Package {creditratingsystem}

Explicit Affordability Ratios and Income Surplus

Description

Calculate stock and service ratios and signed income surplus, then apply user-provided thresholds independently of PD and credit approval.

Usage

debt_to_income(debt_stock, income, income_frequency, currency,
    ratio_period,
               income_basis = "net", context = NULL)
debt_service_to_income(debt_service, income, period, currency,
                       income_basis = "net", context = NULL)
residual_income(net_income, recurring_expenses,
    existing_debt_service,
                new_debt_service, period, currency, context = NULL)
assess_affordability(metrics, threshold_profile, evidence,
    context = NULL)
normalize_income_period(components, target_period, target_basis,
    currency,
  payments_per_year, variable_income_factor, context = NULL)
estimate_recurring_expenses(expenses, target_period, currency,
  payments_per_year, household_profile, benchmarks = NULL,
    context = NULL)
combined_loan_to_value(claims, collateral, allocations, currency,
    context = NULL)
stress_affordability(baseline, scenario, context = NULL)

Arguments

Details

Length-one amounts expand; other vectors require equal lengths. Annualization uses 52 weekly, 26 biweekly, 13 four-weekly, 12 monthly, 4 quarterly and 1 annual payment. Nonpositive income returns a NULL ratio and NONPOSITIVE_INCOME status. No universal thresholds or approval rules are supplied. Threshold comparisons are unrounded, and strict inequalities fail at equality. Missing inputs give INSUFFICIENT_EVIDENCE; known breaches remain visible. Otherwise the assessment is AFFORDABLE or NOT_AFFORDABLE. Household obligations must be reconciled before supplying disjoint totals.

Income normalization instead uses the explicitly supplied payment calendar. Duplicate economic sources count once; conflicting economics fail. Transfers contribute zero. Unknown included income or tax bridges make total income NULL. No taxes are estimated. Expense reconciliation takes the maximum known normalized amount per economic obligation, then applies category floors. It does not infer recurring obligations from raw transactions. Debt service is separate from living costs; unknown debt service stays NULL. Missing required categories and obligations remain explicit.

CLTV counts each allocated collateral once. Lower ranks consume available value first and equal ranks share proportionally. Claim allocation reconciliation uses relative or absolute tolerance 1e-12. Zero collateral value gives NULL ratios. Rank coverage is not an LGD estimate or a legal enforceability assessment.

Stress baseline incomes have economic_id, monthly_amount, net basis and currency; expenses have economic_id, monthly_amount and currency. Contracts have economic_id, currency, principal, remaining_months, annual_rate, repayment (annuity or interest_only), balloon_amount, rate_type (fixed or reset), reset_dates, rate_floor and rate_cap. Rates are nonnegative nominal annual rates divided by twelve. Interest-only requires balloon_amount equal to principal. Stress collateral records use economic_id, value, currency. cltv_scope must declare all_contracts_secured_by_pool or not_applicable. Thresholds explicitly contain minimum_residual, maximum_dsti and maximum_cltv.

Scenario dates are consecutive monthly anniversaries clamped to month-end. Every factor map covers the respective baseline economic IDs exactly, including empty named maps for absent groups. FX is reporting-currency units per source unit, with reporting FX equal to one. Rate shifts apply relative to original coupon only at declared resets, bounded by floor and cap. Projected resets must coincide with payment dates. Annuity payment solves the present-value equation including the terminal balloon and is recomputed at reset. At maturity, actual remaining principal plus interest is due; no refinancing is assumed. Fixed coupons ignore shifts. Residuals accumulate into liquidity, with first strictly negative buffer date reported. Projection continues after exhaustion to expose the funding gap; no default or collections transition is inferred. Annual DTI uses opening debt and current monthly income times twelve; CLTV uses closing debt and the declared common collateral pool. No independent shocks are drawn. Arrays are unnamed lists; objects are named lists, including empty maps.

Value

A CalcResult with ratios and denominator status, signed residuals and component trace, or a separate affordability assessment with rules and headroom.

Examples

debt_to_income(60000, 2500, "monthly", "EUR", "annual")$value
debt_service_to_income(900, 3000, "monthly", "EUR")$value
residual_income(3000, 1000, 900, 200, "monthly", "EUR")$value

Verify Frozen Household Stress Assessments

Description

Checks historically verified economic evidence, household source allocations and frozen contractual stress projections, with separately observed factual payments.

Usage

affordability_stress_backtest(cases, economic_evidence, allocations,
  design, payment_outcomes = NULL, tolerance = 1e-10,
  uncertainty = "none", cluster_resamples = NULL,
  confidence_level = .95, context = NULL)

Arguments

Details

The actual stress engine is rerun with frozen inputs and compared with the original assessment. Unknown or late income is incomplete evidence, not an invented zero. Proportional household shares cannot duplicate an economic source; full recourse is explicit and contractual only. Gross less deductions reconciles to normalized monthly net income. Contractual resets, amortization, FX and policy thresholds remain governed by the original stress input.

Observed scheduled-minus-paid metrics are descriptive facts under complete mature capture; they are not presumed realizations of the scenario debt service. Positive residual income is not zero PD or a lending approval.

Value

A CalcResult with a ValidationResult, verified projection flags, reconciliation findings, separate factual payment metrics, uncertainty and missing/censored period coverage.

Application evidence, selection-aware targets and conditional risk

Description

Freeze dated applicant evidence, retain joint legal roles, construct funded first-default targets and apply explicit conditional models locally.

Usage

build_application_snapshot(applications, feature_records,
    term_events = NULL,
                           required_features = NULL)
assemble_joint_application(application_id, parties, income_items,
                           obligation_items, currency)
build_application_targets(applications, decisions, bookings,
    default_events,
                          exits, horizon_days, evaluation_cutoff,
                          performance_origin = "first_draw_at")
predict_application_risk(snapshot, model, target, calibrator = NULL,
    offset = NULL)
compare_offer_risk(snapshot, offers, model, target, support_bounds,
                   calibrator = NULL, lgd = NULL, offset = NULL)

Arguments

Details

Snapshots select feature revisions with observed_at and known_at no later than information_cutoff, then latest observation and knowledge dates. Ties fail. Source timestamps and requested/offered/accepted/booked terms stay separate. Snapshots have a shared canonical hash; subsequent inference rejects changed content. Objects remain mutable in memory but are copied and integrity checked.

Joint items are deduplicated within income and debt tables by economic_id. Conflicting shared amounts fail. Guarantor resources do not become borrower income. Original liability roles and shares remain; shares need not sum to one. Budget amounts are not multiplied by liability shares. No individual PDs are averaged, and no household identity is inferred. Caller supplies already normalized, recognized monthly amounts in a common currency.

Targets preserve rejected, approved-not-taken-up, accepted-not-booked, booked-not-drawn and funded stages. Known bookings require approval and acceptance; chronology and separate knowledge dates are checked. Information cutoff cannot follow evaluation cutoff. Acceptance attached to a nonapproved decision fails. A target is first default in (origin, origin+horizon], using funded-facility events. Unfunded applications, early exits, incomplete follow-up and pending ascertainment have missing labels. A complete nondefault window yields zero; default at horizon end yields one. Information cutoff remains separate from performance origin. Calendar maturity is reported separately from label_observed. Booking-origin targets explicitly treat booked facilities as the funded risk set even before first draw. One booked facility per application is supported; split multi-facility records explicitly upstream.

Prediction requires model$application_contract with product_ids, population_ids, performance_origin, term_stage and term_features. The last maps term field to fitted feature name and may be an explicitly named empty list. Term-feature names must not also occur among customer features. Product, population and complete target must match; missing/nonfinite model inputs fail. The shared classifier currently supports binary GLM artifacts with identity preprocessing. Optional calibration validates target/version; isotonic additionally declares metadata$input_scale = "probability". Output contains PD, unadjusted PD, model and snapshot identities, target and feature coverage. Risk is separate from creditworthiness budgets and approval decisions.

Offer comparisons require an offered-term conditional model and every scenario term to appear in its feature map. All customer features stay fixed. Marginal support bounds flag extrapolation but cannot establish joint support or causal effects. Optional expected loss is PD times fixed LGD times offered amount; requested_terms$currency is then mandatory and returned. No dynamic EAD, take-up, automatic policy or causal claim is implied.

Value

CalcResult. Snapshots, targets and offer results return lists of row lists; joint application and risk prediction return named records. IDs are strings, dates ISO YYYY-MM-DD. Missing values remain NULL.

Examples

a <- list(list(application_id = "A", customer_id = "C", product_id = "consumer",
               population_id = "new", application_at = "2024-01-01",
               information_cutoff = "2024-01-01", requested_terms = list(amount = 1000)))
f <- list(list(application_id = "A", feature = "income", value = 3000,
               observed_at = "2023-12-01", known_at = "2023-12-02", source_id = "payroll"))
build_application_snapshot(a, f, required_features = "income")$value

Shared Finite Counterfactual Robustness Scenarios

Description

Evaluate candidate feasibility and target success across declared model and measurement scenarios, retaining every failed variant in the denominator.

Usage

assess_counterfactual_robustness(candidates, record, model_variants,
  scenarios, feasibility_contract, target, candidate_ids,
    nominal_variant_id,
  context = NULL)

Arguments

Details

The same additive measurement perturbation applies to baseline and all candidates. Immutable features mean no action, not no measurement uncertainty. Constraints are checked again. Every supplied scenario remains in the normalized denominator, including failed evaluations. A batch callback failure makes that entire scenario unevaluable; no selected rows are silently retried. Feasibility bounds assign unknown feasibility to failure and success respectively. Worst/best target margins are NULL if any margin is unavailable; policy targets have no numeric margin. All-scenario success includes zero-weight scenarios, while weighted fractions follow declared mass. Feature ranges and failure co-occurrence are descriptive, not causal attribution. Finite scenario fractions are not real intervention probabilities or continuous worst-case guarantees.

Value

CalcResult containing nominal and scenario candidate results, conservative success fractions, feasibility bounds, missing-evaluation mass, margins, perturbation summaries and contract hashes.

See Also

compute_credit_ice, find_feasible_credit_counterfactual

Observed Asset-Pool and Tranche Cash Validation

Description

Replay frozen pool forecasts and reconcile independent actual servicing, dated ownership, pool accounts and tranche cash.

Usage

asset_pool_backtest(asset_pool, pool_inputs, forecast_paths,
    forecast_inputs,
  servicing, allocations, pool_accounts, coverage, design,
    cash_contract,
  tranche_contracts = list(), tranche_payments = list(),
  ownership_history = NULL, tolerance = 1e-10, context = NULL)

Arguments

Details

Whole horizons mature before default/prepayment metrics. Incomplete coverage is unknown. Recovery ownership follows the referenced default claim; post-default claim sales require a separate ledger. Original static holdings must match for forecast comparisons. Dynamic or changed holdings receive ledger audits without risk-set-mismatched forecast scores. Tranche cash and principal caps are checked, but contractual waterfall priority, coupons and deficiency ledgers are not inferred. Purchase/sale consideration and completeness are supplied evidence. Forecast scenario weights do not yield empirical confidence intervals. A clean ledger is not release approval.

Value

CalcResult with ValidationResult metrics, explicit applicability, coverage exclusions, ledger violations and cash residuals.

References

BCBS CRE40.93, dated 2019-12-15, transparency of receipts, payments and ledger entries. https://www.bis.org/committees/bcbs/basel-framework/standard/cre/40/inforce/2022-01-01/published/2019-12-15

Dated Behavioral Account Risk and State Dynamics

Description

Reconstruct payment arrears, build dated account snapshots and first-event targets, estimate empirical transitions and calculate descriptive vintage performance. Active zero-balance accounts remain in their risk set.

Usage

compute_days_past_due(obligations, payments, as_of, known_at,
  allocation_policy = "fifo", currency, context = NULL)
assign_delinquency_state(dpd, past_due_amount, exposure, events,
  state_profile, context = NULL)
build_account_month_panel(accounts, statements, account_events,
  observation_dates, reporting_lag_days = 0, obligations = NULL,
  payments = NULL, context = NULL)
build_behavioral_targets(panel, events, horizon_months,
    target_event,
  competing_events, exit_treatment, evaluation_cutoff, coverage,
  context = NULL)
estimate_retail_roll_rates(panel, states, from_date, to_date,
    group_by,
  weighting, missing_successor, context = NULL)
project_retail_roll_rates(initial_distribution, transition_matrices,
  states, default_states, closure_states, exposure_by_period,
  context = NULL)
predict_behavioral_risk(snapshot, model, target, calibrator = NULL,
  offset = NULL, context = NULL)
build_vintage_performance(accounts, panel, events, months_on_book,
  evaluation_cutoff, coverage, default_event, delinquency_states,
  exit_events, context = NULL)

Arguments

Details

Payment reversals are dated and never rewrite an earlier information cutoff. DPD excludes grace/materiality rules. Amounts due today have zero days past due and are excluded from past_due_amount, but retained as unpaid_due_amount. FIFO is a current net-cash allocation contract, not a bank servicing waterfall.

State materiality contains absolute, relative thresholds and and/or logic. Positive arrears and threshold equality qualify. At zero exposure positive arrears satisfy relative materiality. Buckets have state, lower inclusive and upper exclusive integer bounds, gap-free from zero to a final NULL upper bound. event_priority records have event, state, at_risk and absorbing. Explicit events precede the optional dpd_default_rule (minimum_dpd, state, material_required). No bucket is a universal regulatory default definition.

Panel snapshots choose the latest economic statement/revision available by the cutoff. Statements may use a different billing-cycle date. Missing statements produce missing features, not zero balances. Account features include balance, credit_limit, dpd, past_due_amount and utilization. Legal termination is retained; default history survives cure. Customer sums are currency-specific and not PDs.

Targets use (as_of,horizon_end] first-event windows. Previously defaulted and terminated accounts are outside that risk set. Observed events can identify positive labels early. Non-events require explicit mature coverage. Unknown follow-up is NULL when neither coverage nor event is available. Same-day competing target/exit events require upstream resolution. Overlapping account windows and shared customers are dependent observations.

Roll-rate outputs reconcile origin counts/masses to observed destinations plus censoring. Empty observed rows are NULL, never identity rows. Opening-balance weighting measures monetary flows, not account probabilities. Projection keeps first-default states absorbing and separates transition probabilities from conditional exposures. balance_flows are probability flows times destination exposure, not cash transfers. Supplied exposures must reflect drawdowns, amortization and maturity independently of the probability matrix.

Vintage cells use fixed origination-account denominators. Immature or incompletely observed cells remain NULL. Current delinquency stock differs from ever-default; losses are supplied observed loss/recovery cashflows, not ultimate estimates. Age plus cohort equals calendar period, preventing unrestricted causal separation.

Prediction validates snapshot hash, first-default account eligibility, product/state support and maximum_source_age_days from behavioral_contract. The contract requires risk_unit account and event_policy first_default. It invokes supported frozen shared model prediction, then optional calibration. Model-training and calibration knowledge dates cannot exceed the snapshot cutoff. An active zero balance does not force PD to zero. Isotonic calibration must declare metadata input_scale probability. Model actions and future EAD remain separate. Arrays use unnamed lists and objects named lists.

Value

A CalcResult containing reconciled payment records, state assignments, account/customer panels, event labels, empirical transition tables, projected probability/balance paths, behavioral risk predictions or vintage cells.

Build an Explicit Reproducible Calculation Context

Description

Records explicit time, model and data identities with a stable SHA-256 context hash. The system clock is never consulted.

Usage

build_calculation_context(as_of_date = NULL, knowledge_date = NULL,
  horizon = NULL, time_unit = NULL, currency = NULL,
  probability_measure = NULL, conditioning = NULL,
  model_version = NULL, data_manifest = NULL, scenario_id = NULL,
  policy_id = NULL)

Arguments

Details

Without dates the mode is TIME_INDEPENDENT. With both dates the mode is DATED; the dates can be in either order to represent forecast or replay contexts. Hashing sorts object keys and uses a specified binary64 numeric representation, preserving text exactly. Named lists are JSON objects; unnamed lists are arrays. Length-one atomic values are scalars. An unnamed list expresses an explicit empty or length-one array, while structure(list(), names = character()) expresses an empty object. Object keys must be unique printable ASCII text; values may contain UTF-8 text. Non-finite values are rejected. For cross-language interchange, avoid integral JSON literals outside the exact binary64 integer range; represent large non-integer quantities using floating-point notation.

Value

A CalcResult whose value is a context named list with a context_hash. Context-aware functions accept this value or the whole result.

Examples

ctx <- build_calculation_context("2025-12-31", "2026-01-15",
  horizon = 1, time_unit = "years", currency = "EUR",
  probability_measure = "PHYSICAL", conditioning = "survival")
expected_loss(0.02, 0.4, 100000, context = ctx)$unit

Fit and Apply Frozen Probability Calibration

Description

Fit an intercept shift, logistic recalibration, or weighted isotonic step map with explicit target and sample metadata. Apply the frozen artifact with target and model-version checks. All calculations are native R and offline.

Usage

calibrate_intercept(base_pd, target_default_rate, weights = NULL,
  metadata, context = NULL, tolerance = 1e-12, max_iterations = 200)
calibrate_logistic(base_pd, outcomes, weights = NULL, metadata,
  context = NULL, tolerance = 1e-10, max_iterations = 100)
calibrate_isotonic(scores, outcomes, weights = NULL, metadata,
  increasing = TRUE, out_of_bounds = "error", context = NULL)
apply_calibrator(calibrator, values, target, base_model_version,
  context = NULL, groups = NULL)

Arguments

Details

The target contains event, positive horizon, time_unit (days, months or years), probability_measure (PHYSICAL or RISK_NEUTRAL) and conditioning. Calibration metadata contains ISO dates start, end, known_at with start no later than end no later than known-at, a nonempty split_id, and prediction_origin. Empirical fitting requires HOLDOUT or OUT_OF_FOLD; EXTERNAL_TARGET is additionally allowed for intercept shifts. Data origin is USER_SUPPLIED, SYNTHETIC or REFERENCE. These are caller declarations: the functions do not independently establish sample independence or outcome maturity.

Intercept calibration solves the weighted mean equation for \mathrm{logit}^{-1}(\mathrm{logit}(p)+a) while retaining structural zero/one probabilities. Finite-shift endpoint targets and a sample without interior positive-weight probabilities are rejected.

Logistic calibration fits unpenalized binomial maximum likelihood on the base logit. Constant logits, complete separation and quasi separation are reported as controlled failures. Endpoints are never silently clipped. Nonpositive slope is explicitly diagnosed. Fisher covariance assumes independent frequency repetitions; it is not cluster-robust uncertainty.

Isotonic calibration first combines tied scores, then uses weighted pool-adjacent-violators. Its application is left-constant on intervals [x_i,x_{i+1}); there is no linear interpolation. Original input order is preserved. The support policy is stored in the artifact.

Artifacts carry parameters, target, metadata, diagnostics, an input fingerprint and a checked content hash. To exchange across languages, place the artifact in an additional calibration_artifact field of a valid base model and use serialize_fitted_model/load_fitted_model. These retain full numeric precision. Rounded JSON may fail integrity checks intentionally. Hashes detect changed content; they do not authenticate authorship.

Value

A CalcResult. Fit functions return a versioned CalibrationArtifact in $value; application returns a probability vector. Diagnostic and provenance fields remain separate. No result implies validation or approval.

References

https://scikit-learn.org/stable/modules/calibration.html

https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.isotonic_regression.html

https://stat.ethz.ch/R-manual/R-devel/library/stats/html/glm.html

Examples

metadata <- list(
  target = list(event = "first_default", horizon = 12, time_unit = "months",
                probability_measure = "PHYSICAL", conditioning = "PIT"),
  calibration = list(start = "2020-01-01", end = "2021-12-31",
                     known_at = "2022-01-01", split_id = "synthetic",
                     prediction_origin = "EXTERNAL_TARGET"),
  base_model_version = "synthetic_v1", population_definition = "Hand example",
  outcome_policy = "Fully observed first default", weight_type = "frequency",
  data_origin = "SYNTHETIC")
fit <- calibrate_intercept(c(.1, .1), 2/11, metadata = metadata)
apply_calibrator(fit, c(.1, .2, .5), metadata$target, "synthetic_v1")$value

Explicit Calibration, Uncertainty and Model Selection

Description

Native likelihood, pooling, uncertainty and selection methods with frozen artifacts, explicit data roles and retained failed candidates.

Usage

calibrate_beta(base_pd, outcomes, weights = NULL, metadata,
  monotonic = TRUE, boundary_policy = "reject", epsilon = 1e-12,
  context = NULL)
calibrate_grade_pd(counts, defaults, grades = NULL, prior_alpha = 0,
  prior_beta = 0, monotonic = TRUE, target_default_rate = NULL,
  default_grade = NULL, confidence = .95, target_weights = NULL,
  context = NULL)
low_default_bound(defaults, counts, confidence = .95,
  dependence = "independent", correlation = 0, pooling = "none",
  integration_tolerance = 1e-8, context = NULL)
fit_hierarchical_calibration(base_pd, outcomes, groups,
    weights = NULL,
  metadata, variance_bounds = c(1e-8, 25), context = NULL)
calibrate_lgd_level(base_lgd, target_lgd, weights = NULL, metadata,
  transformation = "additive", context = NULL)
estimate_margin_of_conservatism(base_estimate, components,
  confidence = .95, correlation = NULL,
    already_included = character(),
  bounds = NULL, context = NULL)
population_reweighting(sample_features, target_distribution,
  method = "poststratification", cap = NULL,
    outcome_dependent = FALSE,
  context = NULL)
fit_stepwise(X, outcomes, feature_order, candidate_groups,
  mandatory_terms = character(), hierarchy = NULL,
  direction = "bidirectional", criterion = "AIC", estimator = "GLM",
  weights = NULL, max_steps = 100, lr_alpha = .05,
  training = NULL, context = NULL)
fit_regularization_path(X, outcomes, feature_order, penalty_grid,
  split_plan, estimator = "GLM", elastic_net_alpha = 0,
  scoring = "log_loss", weights = NULL, context = NULL)
resample_fit_uncertainty(X, outcomes, feature_order,
    resampling_indices,
  resampling_units, unit_kind, estimator = "GLM", weights = NULL,
  confidence = .95, prediction_X = NULL, selection_spec = NULL,
  context = NULL)
fit_model_selection_protocol(X, outcomes, feature_order, candidates,
  split_plan, metadata, search_budget, scoring = "log_loss",
    context = NULL)

Arguments

Details

Beta calibration fits plogis(c + a*log(p) - b*log(1-p)); (c,a,b)=(0,1,1) is the identity. Active monotonic constraints are reported: interior inverse-information covariance is not boundary-aware inference.

Hierarchical calibration uses supplied logits as offsets, a global intercept, and independent normal group intercepts. Variance is estimated using Laplace marginal likelihood with analytic profile derivatives; conditional modes are solved individually. Existing groups use their mode. Unknown groups integrate over the fitted population using 40-node Gaussian quadrature. Reported group uncertainty conditions on estimated hyperparameters; it excludes hyperparameter and Laplace-approximation error. Variance-boundary estimates are explicit.

Grade PAVA pools empirical or beta-prior means in grade order. Optional target shifts preserve structural zeros and ones and reject unattainable finite shifts. Intervals are exact binomial without a prior, beta posterior with a prior. They describe pre-target pools conditional on pooling, not post-selection or simultaneous coverage. Zero-default cells retain positive upper uncertainty.

Low-default roots solve P(D <= observed defaults | pd)=1-confidence. For zero independent defaults the answer is 1-(1-confidence)^(1/n). The Gaussian factor mode integrates conditional independent binomial probabilities; ordered-tail pooling also reports monotone upper bounds. Coverage is conditional on the declared model, not universal under unknown dependence. Integration diagnostics exclude negligible standard-normal probability outside [-10,10].

Additive LGD leaves raw economic losses unclipped. Multiplicative mode requires nonnegative losses and a positive base mean; a zero target ties all predictions. Logit mode requires interior fractions. Exposure weighting is separate from frequency weighting and must be declared explicitly.

MoC is total adverse bias plus qnorm(confidence)*sqrt(s' C s). Component source IDs prevent duplicate counting; already-incorporated components contribute zero. This declared normal uncertainty policy is not an automatic regulatory default. Poststratification rejects missing positive target strata. Membership odds are q/(1-q)*sample_prior/target_prior; externally supplied propensities do not prove covariate overlap. Clipping occurs before normalization; effective sample size and achieved stratum balance expose its effect.

Stepwise Gaussian likelihood includes residual variance and, for WLS, the weight determinant. AIC/BIC use identical rows; LR compares nested likelihoods and records degrees of freedom. Ties within 1e-10 use lexical action/group order; visited states prevent cycling. Rank deficiency, separation and zero residual variance are failures, retained in candidate evidence. No adaptive selected-model fixed- model inference is claimed.

Regularization minimizes weighted average negative log likelihood (Gaussian half squared error) plus lambda*((1-alpha)*sum(beta^2)/2+alpha*sum(abs(beta))). Each fold scales using only its training data; final refit uses development only. Every fold must succeed; exact loss ties prefer stronger penalty. Entity sets are disjoint across fold and stage boundaries, and earlier labels must be known strictly before later observations. Supplied data relationships remain a user responsibility. Standalone selection scoring is intercept plus centered/scaled matrix times coefficients, followed by plogis for GLM. Preserve feature order, training cutoff and the content hash when replaying a SelectionModel.

Resampling repeats complete declared clusters/time blocks and retains every failure. Optional selection repeats within each replicate. Percentile intervals use successful replicates and require at least two; a large failure proportion can invalidate their interpretation. The full protocol performs inner development selection, final development refit, then independent logistic calibration with score tolerance 1e-8. Test labels and predictions never enter optimization. Synthetic engineering evidence is not validation of performance on bank data.

Value

CalcResult. Fitted transformations return a hashed CalibrationArtifact; apply_calibrator validates its target, version, schema, domain and content. Other methods return their estimates and complete diagnostics. Selection methods return a hashed SelectionModel with explicit scoring parameters, training role/cutoff and diagnostic hash. The full protocol additionally returns a common selected_frozen_model for predict_pd, a separate calibration artifact, all candidate evidence, and untouched test indices; it reports no test metrics.

References

The lme4 authors, glmer documentation.https://lme4.github.io/lme4/reference/glmer.html. Friedman, Hastie and Tibshirani, regularization paths and elastic-net penalties. https://www.jstatsoft.org/article/view/v033i01.

Kull, Silva Filho and Flach (2017), Beta calibration. https://proceedings.mlr.press/v54/kull17a.html. Pluto and Tasche, Estimating Probabilities of Default for Low Default Portfolios. https://arxiv.org/abs/cond-mat/0411699. R stats, step. https://stat.ethz.ch/R-manual/R-devel/library/stats/html/step.html.

Examples

low_default_bound(0, 100)$value$bounds
calibrate_grade_pd(c(100, 100, 100), c(1, 4, 2))$value$grades
population_reweighting(c("A", "A", "A", "B"), list(A=.5, B=.5))$value

Proper-prior Bayesian logit and probit posterior simulation

Description

Native multi-chain Metropolis estimation using explicit shared innovations, frozen posterior draws, uncertainty and actual chain diagnostics.

Usage

fit_bayesian_pd(X_train, y_train, feature_order, link, prior_mean,
  prior_sd, sampler_spec, group_spec = NULL, credible_level = .95,
  target, training, model_id, model_version, context = NULL)

Arguments

Details

The posterior combines the Bernoulli likelihood with proper independent normal priors. Symmetric random-walk Metropolis uses caller-supplied independent Gaussian and uniform draws; no proposal adaptation or hidden random seed is used. Retained posterior draws, proposal, initial positions and innovation fingerprint are frozen for audit and prediction. Group intercepts implement partial pooling conditional on their fixed proper prior SD. Parameter order is intercept, slopes, then declared group levels.

Diagnostics include chain acceptance rates, rank-normalized and folded split-Rhat, Geyer initial-positive/monotone paired-sequence ESS for bulk/tails, and raw-quantity mean ESS. ESS is capped by actual draw count. Splitting uses equal first/last halves; an odd middle draw is omitted. Constant chains report unavailable Rhat and zero ESS. DIAGNOSTICS_PASSED requires every coefficient Rhat<=1.01 and bulk/tail ESS>=400. Otherwise a visible failed/insufficient status accompanies the draws; no convergence claim is made. Finite-chain diagnostics cannot prove convergence.

predict_pd averages probabilities over posterior draws, rather than applying the link to mean coefficients. Diagnostics contain equal-tail PD intervals, latent-PD posterior SD, mean-PD Monte Carlo standard error, prediction-specific chain diagnostics and Bernoulli predictive variance. Intervals and MCSE require representative stationary posterior draws, so inspect convergence status before use. These are intervals for conditional PD, not fractional outcome prediction intervals. For grouped fits prediction X is list(X=numeric_matrix,groups=labels); all groups must have been trained.

Prior sensitivity, especially with few defaults, requires refitting explicit substantive alternative priors. No heldout data select a prior implicitly. Model-based posterior uncertainty is conditional on likelihood, priors and independent-row assumptions; it does not establish empirical calibration or independence of repeated borrower forecasts.

Value

FitResult with a data-only ModelArtifact containing all retained posterior draws and chain diagnostics. Use predict_pd with the exact frozen target and features.

References

Vehtari et al. (2021), Rank-normalization, folding, and localization: An improved R-hat for assessing convergence of MCMC. Stan reference manual, Posterior Analysis and Effective Sample Size.

See Also

predict_pd,predict_model_score,serialize_fitted_model

Constrained logistic spline models and portable tree ensembles

Description

Training-only smooth and tree classifiers with explicit constraints, sampling, inner validation and data-only prediction artifacts.

Usage

fit_gam_pd(X_train, y_train, feature_order, linear_features,
    smooth_specs,
  variant, max_iter = 2000L, tolerance = 1e-9, target, training,
  model_id, model_version, context = NULL)
fit_random_forest_pd(X_train, y_train, feature_order, sampling_plan,
  max_depth, min_samples_leaf = 1L, min_weight_leaf = 0,
  criterion = "gini", weights = NULL, target, training, model_id,
  model_version, context = NULL)
fit_gradient_boosting_pd(X_train, y_train, feature_order,
    train_row_ids,
  iterations, max_depth, learning_rate,
    monotonic_constraints = NULL,
  l2_regularization = 1, min_child_weight = 1, min_split_loss = 0,
  max_bin = 256L, inner_validation = NULL,
    early_stopping_rounds = NULL,
  target, training, model_id, model_version, context = NULL)

Arguments

Details

GAM minimizes mean Bernoulli negative log likelihood plus half the smoothing-weighted squared second differences of full spline coefficients. The first coefficient of each smooth is zero to identify its intercept. Monotone spline coefficients are cumulative nonnegative increments, with reversed sign for decreasing terms; constraints operate during L-BFGS-B fitting. Frozen tangent/constant/error boundary behavior preserves the declared policy. A native LP detects feasible separating recession directions in the penalty null space at relative eigenvalue tolerance 1e-10, respecting monotone bounds; smoothing alone does not cure affine separation. Such fits are rejected. Projected KKT residual and EDF on free coefficients are reported. No smoothing-selection or active-constraint uncertainty interval is inferred. No smooths reduces to ordinary logistic MLE.

Forests evaluate all valid weighted CART midpoint splits using the caller's bootstrap/subsample and feature draws. Gain ties use candidate order then increasing threshold at tolerance 1e-14. Weighted leaf proportions are averaged without clipping. OOB predictions exclude all trees containing that original row and report coverage, counts and Brier; OOB is not a temporal or obligor-group holdout. Randomness is explicit in the supplied design rather than hidden in package-global state.

Boosting uses native XGBoost CPU hist Bernoulli loss without stochastic subsampling. Initial margin is the logit of training default fraction. Frozen leaf values already include shrinkage; input casting to float32 and strict-less-than routing match backend prediction. Nonrepresentable values are rejected. All evaluated loss rounds and the selected early-stop prefix are recorded. Feature/label availability and grouping independence remain caller responsibilities; disjoint record IDs alone cannot prove temporal integrity. No empirical calibration or uncertainty interval is claimed by these fits.

Value

FitResult with a frozen data-only ModelArtifact and model-specific numerical, design and validation diagnostics. Use predict_pd or predict_model_score with matching target and features.

References

Eilers and Marx (1996), Flexible smoothing with B-splines and penalties. Breiman (2001), Random Forests. Chen and Guestrin (2016), XGBoost: A Scalable Tree Boosting System. R splines::splineDesign documentation and XGBoost model/monotonic-constraint documentation.

See Also

predict_pd, predict_model_score, serialize_fitted_model

Portable probabilistic classification models and explicit raw scores

Description

Native training-only estimation with frozen data-only prediction artifacts, explicit target and feature ordering, and model-specific diagnostics.

Usage

fit_discriminant_pd(X_train, y_train, feature_order, variant,
  covariance_estimator = "mle", covariance_ridge = 0, priors = NULL,
  target, training, model_id, model_version, context = NULL)
fit_naive_bayes_pd(X_train, y_train, feature_order, feature_types,
    variant,
  alpha = 1, variance_floor = 0, missing_policy = "error",
  unknown_policy = "error", priors = NULL, target, training,
  model_id, model_version, context = NULL)
fit_svm_score(X_train, y_train, feature_order, kernel, C,
    gamma = NULL,
  scale = TRUE, class_weights = NULL, tolerance = 1e-7,
    max_iter = NULL,
  target, training, model_id, model_version, context = NULL)
predict_model_score(fitted_model, X, feature_order, expected_target,
  context = NULL)
fit_firth_logit(X_train, y_train, feature_order,
    fit_intercept = TRUE,
  max_iter = 1000L, tolerance = 1e-9, target, training, model_id,
  model_version, context = NULL)
fit_penalized_glm(X_train, y_train, feature_order, penalty,
    lambda_path,
  selected_lambda, l1_ratio = NULL, fit_intercept = TRUE,
    scale = TRUE,
  weights = NULL, max_iter = 10000L, tolerance = 1e-9, target,
    training,
  model_id, model_version, context = NULL)
fit_ordinal_model(X_train, y_train, feature_order,
    ordered_categories,
  link, default_categories = NULL, max_iter = 1000L,
    tolerance = 1e-9,
  target, training, model_id, model_version, context = NULL)

Arguments

Details

Gaussian discriminants apply Bayes' theorem to fitted class means/covariances. Naive Bayes additionally assumes conditional independence, and its Gaussian variances use maximum-likelihood denominators. These posteriors do not establish empirical calibration.

Elastic-net minimizes weighted mean negative logistic likelihood plus lambda times ((1-alpha)/2 times squared slope norm plus alpha times absolute slope norm). Each path point records objective and KKT residual. Penalties operate on fitted standardized slopes when scale is TRUE. Zero penalty checks rank and separation and reproduces ordinary logistic likelihood. No post-selection uncertainty is supplied.

Firth maximizes Bernoulli log likelihood plus one half log determinant of expected Fisher information. Native brglm2 MPL_Jeffreys with a=1/2 is checked against the exact adjusted score. Full rank is required; single-class samples and separation are permitted. Stored inverse expected information is a local diagnostic, not a penalized profile confidence interval.

Ordinal models obey P(Y<=j|x)=F(threshold_j-x beta). Prediction scores expose all category probabilities in diagnostics. Default categories must be an ordered terminal tail. The parallel-slope diagnostic jointly stacks threshold-wise binary scores and retains cross-threshold covariance in an IID sandwich Wald test. Singular or separated diagnostic fits yield NOT_ESTIMABLE; the diagnostic does not prove proportionality. Covariance parameters are slopes, first threshold and log subsequent threshold spacings.

SVM outputs are signed margins, never PDs without independent calibration. Frozen support vectors/dual coefficients/intercept are portable; higher margins mean default. Preprocessing never uses prediction rows. Model-based likelihood uncertainty assumes independent observations; temporal/group holdout selection and mature labels remain explicit caller responsibilities.

Value

Fits return FitResult with a portable ModelArtifact and diagnostics. predict_model_score returns CalcResult containing a numeric score array, explicit scale, orientation and model lineage; ordinal diagnostics include category probabilities.

References

Firth (1993), Bias reduction of maximum likelihood estimates. Kosmidis and Firth (2021), Jeffreys-prior penalty, finiteness and shrinkage in binomial-response generalized linear models. Tay, Narasimhan and Hastie (2023), Elastic Net Regularization Paths for All Generalized Linear Models. Chang and Lin (2011), LIBSVM: A library for support vector machines. MASS package documentation for lda, qda and polr.

See Also

predict_pd, serialize_fitted_model, load_fitted_model

Native portable MLP, GRU and LSTM credit likelihood models

Description

CPU dense and recurrent neural estimation with actual analytic gradients, explicit Adam orders, frozen weights and causal sequence availability contracts.

Usage

fit_neural_pd(X_train, y_train, feature_order, architecture,
  initial_weights, optimizer_spec, train_row_ids, scale = TRUE,
  inner_validation = NULL, target, training, model_id,
    model_version,
  context = NULL)
fit_sequence_pd(sequences, y_train, feature_order, variant,
    hidden_size,
  initial_weights, optimizer_spec, train_row_ids, label_known_times,
  training_as_of_time, include_time_gaps = TRUE, time_gap_scale = 1,
  scale = TRUE, inner_validation = NULL, target, training, model_id,
  model_version, context = NULL)

Arguments

Details

Forward propagation and full backpropagation are native, with no Python bridge or runtime neural backend. Loss is mean Bernoulli negative log likelihood. Adam bias-corrected moments use a coupled L2 weight-gradient penalty, not AdamW; biases are unpenalized. Optional global gradient clipping occurs after L2. Initialization fingerprints, epoch orders, update counts and full loss paths are recorded. Completion of an epoch budget is not called global optimization convergence.

For input width p and hidden width h, GRU input/hidden weights have shapes (p,3h)/(h,3h), LSTM (p,4h)/(h,4h). Gate-width input and hidden biases are separate. Output weight is (h,1), output bias length one. GRU gate order r,z,n uses reset after the hidden affine: r=sigmoid(input_r+hidden_r), z=sigmoid(input_z+hidden_z), n=tanh(input_n+r*hidden_n), h=(1-z)*n+z*previous_h. LSTM i,f,g,o uses c=f*previous_c+i*g, h=o*tanh(c), sigmoid i/f/o and tanh g. These matrices are transposes of the corresponding PyTorch stored gate matrices. Recurrent and cell states reset to zero for every record and prediction call. Genuine gated BPTT updates all recurrent parameters.

Only mask-one observations whose observation and knowledge times do not exceed forecast time enter sequence state or scaling. Known time cannot precede observation. Padding/future/late-known observations are excluded before processing. At least one observation must remain; retained times are strictly ordered. Time-gap augmentation increases input width by one. Pooling weights each retained sequence observation equally, so longer sequences contribute more to scaling. Label-known times prevent immature negatives, while observed early positive defaults may enter earlier. Correct event evidence and calendar/unit interpretation remain caller responsibilities.

With inner validation, the frozen artifact retains the best validation epoch, including epoch zero when all updates harm validation. Without validation it retains the final epoch. Early stopping never reads an undeclared final test set. Disjoint row IDs do not prove independence of repeated borrower records. No posterior uncertainty or empirical probability calibration is inferred from neural training loss.

Value

FitResult containing only portable architecture, weights, preprocessing, target and training/optimization diagnostics. Use predict_pd or predict_model_score with matching features and target; sequence prediction receives the same sequence-record structure.

References

Kingma and Ba (2015), Adam: A Method for Stochastic Optimization. PyTorch GRU and LSTM documentation for gate equations and parameter conventions.

See Also

predict_pd,predict_model_score,serialize_fitted_model

Dated Collections Episodes, Operational Cure and Recovery Economics

Description

Construct collections histories from explicit dated states and coverage, preserve censored cure and redefault outcomes, separate pre-action information from treatment stages, and evaluate local repayment/recovery scenarios.

Usage

build_collections_episodes(states, actions, coverage, episode_rule,
  evaluation_cutoff, context = NULL)
build_cure_targets(episodes, horizon_days, redefault_horizon_days,
  evaluation_cutoff, context = NULL)
build_collections_action_history(actions, features, outcomes,
    coverage,
  index_stage, required_features, horizon_days,
    maximum_feature_age_days,
  evaluation_cutoff, context = NULL)
simulate_repayment_arrangement(contract, baseline, candidate,
  affordability_paths, discount_factors, period_end_dates, currency,
  context = NULL)
expected_collections_value(actions, discount_factors, ead, currency,
  period_end_dates, context = NULL)

Arguments

Value

Each method returns a CalcResult with explicit limitations.

build_collections_episodes returns episodes, complete cutoff-filtered state_history, unlinked_actions, rule, coverage, evaluation_cutoff and history_hash. Episodes contain identifiers, start/end, OPEN/CURED/TERMINATED status, previous_episode_id, timeline, actions, ever_default, observation_cutoff, cure_candidate_at and episode_hash. Histories and rules are bound by canonical hashes.

build_cure_targets returns one row per episode with horizon_end, cure_label/status, cure_at, redefault_label/status, redefault_at, ever_default and cutoff. Labels are 1 or 0 only when their corresponding event or non-event is established; otherwise NULL with censoring status.

build_collections_action_history returns rows and unindexed_treatments. Rows contain preaction_features, feature_sources, missing/stale features, stage_dates, invalid_stage_order, descriptive_eligible, accepted/performed indicators, prior_or_same_day_outcome_ids, subsequent outcomes, outcome_complete and row_hash. causal_effect_identified is always FALSE.

simulate_repayment_arrangement returns baseline/candidate ledgers, NPV change, principal_arrears_included, contract_changed=FALSE and default_status_changed=FALSE. Each plan contains schedule, discounted net-cashflow NPV, remaining_claim and household residual cash/shortfall dates.

expected_collections_value returns one row per supplied action with expected_net_recovery, its EAD ratio, positive-mass scenario min/max, expanded accepted/declined scenarios, support_status, assumption_reference and causal_effect_identified=FALSE.

Episode and cure boundaries

An episode starts in a declared problem state. A clean transition starts a quiet candidate interval. Episode separation requires complete coverage through max(probation_days, separation_days). A problem event exactly on that boundary interrupts the interval. Terminal states close the episode as TERMINATED. An open data endpoint, a single payment or an unobserved gap never establishes cure. A later problem spell carries the previous episode identifier. Historical default is not erased by operational cure.

The cure target uses probation_days, independently of the longer possible episode separation duration. A supported cure on or before the episode horizon end is a positive observed event, even if later follow-up is censored. A negative label requires coverage from episode start through horizon end and a mature cutoff. A terminal exit within the target window censors a non-cured episode.

Redefault uses a separate risk set starting strictly after cure through the inclusive post-cure horizon end. The first default versus terminal event determines positive redefault versus exit censoring. A negative label requires complete post-cure coverage. The full sealed state history retains terminal events after an episode has already closed, so they cannot disappear from follow-up.

Action information and interpretation

Features must both be observed and known strictly before the actual assignment or offer date. Same-day feature records are excluded because intraday ordering is not established by calendar dates. Eligible records are selected by account/name, ordered by observation date, knowledge date and record identifier. Late logging of the action does not move its information cutoff. Missing and stale values remain visible; no imputation is performed. Duplicate latest observation/knowledge timestamps require explicit resolution.

ASSIGNED, OFFERED, ACCEPTED and PERFORMED dates are checked for chronological order; invalid order makes a row descriptively ineligible. CANCELLED remains a separate stage record. A later accepted/performed status is a post-index outcome of the process, not a pre-index feature. Treatments without the declared index stage are listed separately. Same-day outcomes have unresolved intraday order and are never used as prospective outcomes. Outcome absence is not a zero label when coverage is incomplete. No causal effects, treatment eligibility or action ranking are inferred from these factual groups.

Economic inputs and cash reconciliation

For repayment simulation, interest is an effective rate for each supplied period on its opening principal. It is not an annual rate and receives no hidden day-count conversion. New fees and interest enter one claim ledger. Scheduled cash is capped by the outstanding claim; receipt fractions determine actual cash. The waterfall pays fees, then interest, then principal. Remaining interest is capitalized only when explicitly selected. Unused payment capacity, all waterfall amounts and the closing claim remain visible.

NPV sums received cash less servicing cost after applying period factors. Remaining claims carry no assumed terminal recovery. Household residual cash subtracts scheduled obligations, not the lower cash receipts under assumed nonpayment. Plans do not change a legal contract or a recorded default status.

For expected_collections_value each action row requires action_id, currency, uptake_probability, accepted_scenarios, declined_scenarios, offer_cost, support_status and assumption_reference. Each branch must contain a complete conditional scenario distribution, including when branch probability is zero. A scenario contains unique scenario_id, probability and cashflows. Conditional probabilities lie in [0,1] and sum to one within 1e-12; they are used as supplied. Each cashflow has unique cashflow_id within its scenario, kind RECOVERY or COST, period_end matching a declared date, and finite nonnegative amount.

Joint scenario masses are uptake times accepted conditional mass, or one minus uptake times declined conditional mass. Recovery and servicing costs are discounted per scenario. Offer cost occurs once at time zero; it must not also appear as a scenario cost. Expected recovery weights these exclusive net values. Cashflow IDs prevent repeated records within one scenario; callers must give the same economic cashflow one identifier. Scenario ranges are not confidence intervals and inputs are conditional assumptions, not identified treatment effects.

Examples

plan <- list(plan_id = "base", payments = c(50, 50),
  period_interest_rates = c(0, 0), new_fees = c(0, 0),
  servicing_costs = c(0, 0), receipt_fractions = c(1, 1),
  capitalize_unpaid_interest = FALSE)
contract <- list(currency = "EUR", principal = 100, principal_arrears = 20,
  accrued_interest = 0, fees = 0, as_of = "2024-01-01")
simulate_repayment_arrangement(contract, plan, plan, list(), c(1, 0.95),
  c("2024-02-01", "2024-03-01"), "EUR")$value$npv_change

Validate Dated Collections Cashflow Forecasts

Description

Factual fixed-horizon assessment of frozen repayments, recoveries, costs and recovery timing with explicit historical knowledge and complete cash-capture coverage.

Usage

collections_backtest(forecasts, episodes, cashflows, actions,
  cashflow_coverage, discount_contract, design, uncertainty,
  cluster_resamples = NULL, confidence_level = .95, context = NULL)

Arguments

Details

Amounts are discounted to forecast date by the effective annual rate over actual days divided by 365. Bias is prediction minus observation. Forecast and actual dates use an inclusive forecast-to-horizon window. Only mature horizons with complete cash capture and positive evaluation weight are scored. Late-known cash is excluded; state-history coverage alone cannot establish complete cash capture. Open episodes stay open. Recovery timing is recovery-amount-weighted days, defined only for positive forecast and observed recovery mass. Bootstrap intervals condition on frozen forecasts and become non-estimable when any supplied resample lacks the required outcome. Action-group metrics describe factual assignment cohorts, not causal effects.

Value

A CalcResult containing a ValidationResult, amount and timing metrics, coverage, uncertainty and episode/action/cashflow traces. It does not grant release or empirical approval.

Compile, evaluate and explain local credit decision policies

Description

Deterministic rules keep probability estimates, affordability, eligibility and operational information separate. Evaluation returns a local recommendation with fingerprints and a complete rule trace.

Usage

compile_credit_decision_policy(specification, context = NULL)
apply_credit_decision_policy(policy, risk, affordability,
    eligibility,
  operations, as_of, known_at, context = NULL)
decompose_policy_decision(decision, reason_dictionary,
    context = NULL)

Arguments

Details

Rules have a unique rule_id, nonnegative integer priority, action, reason_code and a list of conjunctive predicates. Lower priorities execute first; equal priorities sort by rule identifier. Predicates name a declared source field and use eq, ne, lt, le, gt or ge. Categories allow equality operators only. Numeric thresholds are finite binary64 values. Contradictory rules and overlapping equal-priority rules with different actions are rejected. Lower-priority overlap and shadowing are reported as diagnostics.

The policy effective interval includes its start and excludes its end. Missing, future, stale or incompatible information takes the explicit missing action before the rule result. Under specification 1.0, approval also requires AFFORDABLE, ELIGIBLE and CLEAR statuses from the three separate nonrisk sources; otherwise it becomes REFER. Gate evaluation never alters a probability. Default action cannot be APPROVE. Approval references record caller assertions and do not establish institutional authorization. No credit, payment, limit or price is executed.

Specification 1.1 requires approval_contract with exactly domain, affordability and required_gates. Domains are RETAIL_CREDIT, WHOLESALE_CREDIT, INVESTMENT and COUNTERPARTY_LIMIT. Affordability mode REQUIRED preserves AFFORDABLE. Only investment/counterparty domains permit NOT_APPLICABLE, with nonempty reason and reference and nonnegative whole-day maximum_age_days. The dated runtime affordability object must then declare NOT_APPLICABLE, the same domain, matching applicability_reference and an exclusive valid_until date. The age limit is inclusive. Missing, stale, future-known, expired or differently bound evidence prevents approval.

Additional gates are distinct named lists with field and value, referencing an explicitly declared categorical field and one allowed category. All gates must pass together. Mandatory ELIGIBLE, CLEAR, current risk and target checks remain independent of rule priority. Failed approval status gates yield REFER. The decision adds approval_domain and approval_gate_trace. Empty gate arrays are permitted but do not waive mandatory gates. Contract domains and legal applicability conclusions are supplied assertions, not inferred approvals.

Specification 1.0 rejects approval_contract, retains its existing hashes and behavior, and cannot silently ignore the new semantics. Artifact envelopes remain schema 1.0; the nested specification carries version 1.1.

Array fields, including category values, rules and predicate lists, must be unnamed lists even for one or zero elements. This preserves canonical JSON array identity across native R and Python. Objects must be named lists. Fingerprints detect content changes; they are not digital signatures.

Value

A CalcResult. Compilation returns a DecisionPolicy artifact; application returns a DecisionArtifact containing action, findings, rule trace, input and policy hashes. Decomposition returns recorded matched reasons and identifies which rule was selected, without recalculating a decision.

Supported First and Purified Second-Order Accumulated Local Effects

Description

Weighted finite-bin local differences with explicit interpolation, empirical centering and unsupported-cell reporting.

Usage

compute_credit_ale(predictor, X, model_contract, selected_features,
  bins, feature_types, weights = NULL, min_bin_count = 1,
    context = NULL)

Arguments

Details

Bins are left-closed/right-open except the final upper endpoint. First-order weighted upper-minus-lower local differences are accumulated, linearly interpolated within bins and centered over the original weighted reference. Second-order four-corner mixed differences are double-accumulated; both first-order accumulated components are removed, followed by weighted empirical centering of the bilinear surface. The purification components are returned. Empty or insufficient cells yield NON_ESTIMABLE_SUPPORT with NULL global curves, while supported local differences remain available. No gaps are filled with zero effects or interpolated across. Numeric categories require a separate methodology and are rejected unless explicitly declared continuous. Counts and correlation do not establish off-manifold boundary support or causal effects. The exact finite-bin interpolation convention is part of the method contract.

Value

CalcResult with centered boundary grid and reference effects when supported, local differences, counts, masses, unsupported cell indices, purification components, correlations and hashes.

References

Apley and Zhu, Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models. https://arxiv.org/abs/1612.08468.

Feasible Ordered Local Credit Prediction Paths

Description

Evaluate an explicit finite ICE path while retaining infeasibility, nonmonotonicity, scalar crossings, optional rating classes and full-policy replays.

Usage

compute_credit_ice(predictor, record, grid, model_contract,
  feasibility_contract, varying_features, thresholds = NULL,
  rating_scale = NULL, policy_evaluator = NULL,
    policy_contract = NULL,
  context = NULL)

Arguments

Details

Bounds, immutable features, integers and one-hot identities are checked before prediction. Linear and derived identities use relative tolerance 1e-10; zero-denominator ratios fail. Infeasible points remain NULL with findings and no predictor or policy invocation. Gaps are never interpolated. Turns are along the supplied ordered path and do not imply a sorted feature axis or a unique continuous threshold. Rating classes use map_pd_to_grade; policy decisions require a separately bound full-policy replay. Content binding is a caller assertion, not authentication. Feasibility is declared structural feasibility, not proof of empirical support or real intervention success. No actual credit action is executed.

Value

CalcResult with grid records, predictions, findings, threshold states, adjacent changes, turns, optional grades and sealed decisions, and contract hashes.

See Also

map_pd_to_grade, apply_credit_decision_policy

Canonical Content Fingerprint for Local Data Contracts

Description

Computes the SHA256 canonical identifier used by local frozen simulation recipes, parameter snapshots and artifact bindings.

Usage

content_fingerprint(value, context = NULL)

Arguments

Details

Object keys must be nonempty printable ASCII and are sorted; string values are UTF-8. Numbers use finite binary64 encoding, with negative zero normalized to zero; equivalent integer and floating values share a digest. Booleans, numbers, strings, arrays, objects and null remain distinct. R named empty lists represent empty objects, while list() is an empty array. The identifier is not a digital signature, authorship proof, or financial schema validation. Exclude an artifact's own digest field before hashing its payload.

Value

A CalcResult whose value is a hexadecimal SHA256 string.

Examples

content_fingerprint(list(principal = 100, currency = "EUR"))$value

Validate Frozen Counterparty Exposure and Closeout Forecasts

Description

Physical per-netting-set exposure quantile and amount validation with separate mature observed-default closeout assessment.

Usage

counterparty_exposure_backtest(forecasts, observations,
    quantile_levels,
  design, uncertainty, cluster_resamples = NULL,
  confidence_level = .95, context = NULL)

Arguments

Details

Quantiles use the weighted empirical inverse CDF; exceptions are strictly greater than the threshold. Atom-dependent model tail mass is reported separately from the nominal tail probability. Market observations require exact horizon alignment and known information. Frozen netting/CSA content is compared, and reported exposure reconciles to the exact per-set value without cross-set offset.

Default-conditioned closeout is separately calculated using the public frozen-collateral-unit routine and declared joint market/default path events. Only mature known closeouts are scored; missing defaults are not zero-loss observations. A mature closeout can remain applicable when a horizon market value is censored. Bootstrap intervals condition on the frozen paths and become non-estimable if a resample lacks required default outcomes.

Value

A CalcResult with a ValidationResult, market exposure and quantile scores, separate conditional closeout errors, coverage, uncertainty and reconciliation findings. Reconciliation conflicts fail the declared contract; performance approval is not inferred.

References

Basel Committee on Banking Supervision, CRE50, publication 2024-07-05, definitions 50.15, 50.19 and 50.22–50.28. https://www.bis.org/committees/bcbs/basel-framework/standard/cre/50/inforce/2019-12-15/published/2024-07-05

Capped issuer, cover-pool and residual-estate recovery

Description

Allocates covered-bond payments under externally supplied legal, segregation, maturity-extension and recovery conventions.

Usage

covered_bond_dual_recourse(cover_pool, cover_pool_paths,
    bond_contract,
  issuer_paths, context = NULL)

Arguments

Details

Cover collections accumulate net of pool costs in a segregated account. Before issuer default, issuer capacity pays due claims; after default, cover cash pays. Default precedes same-day issuer payments. Interest ranks before principal. Declared extension requires issuer default by original maturity and shifts principal while activating only explicitly supplied extra coupons.

Residual estate settlement follows complete cover runoff and every possible maturity date. Recovery is capped by remaining claim and estate capacity, preventing double recovery. Original-date deferral, effective-maturity shortfall, timely-interest failure and ultimate loss remain separate. Legal enforceability and national insolvency rules are external assumptions. No penalty interest or automatic recovery uplift is inferred.

Value

CalcResult with a bound CoveredBondRecoursePaths artifact, source-separated cashflows, account reconciliation and scenario ending shortfalls.

References

https://finance.ec.europa.eu/banking/banking-regulation/covered-bonds_en

Local typed credit data and dated observation preparation

Description

Native offline CSV/JSON import, relational checks, bitemporal joins, first-default windows, workout linkage and group-isolated chronological splits.

Usage

read_local_dataset(path, schema, encoding = "utf-8",
    delimiter = ",",
                   decimal_mark = ".")
validate_credit_dataset(tables, contract_version = "1.0")
join_as_of(left, right, entity_keys = "entity_id", as_of = "as_of",
           known_at = "known_at", observed_at = "observed_at",
           available_at = "available_at", release_id = "release_id",
           suffix = "_right")
build_observation_windows(ratings, default_events, exits,
    horizon_days,
                          evaluation_cutoff,
    entity_key = "entity_id")
build_recovery_episodes(default_events, recovery_cashflows,
    cures = NULL,
                        reopening_days = 90,
    evaluation_cutoff = NULL, closures = NULL)
make_grouped_temporal_splits(panel, train_end, calibration_end,
    test_end,
                            embargo_days = 0,
    entity_key = "entity_id",
                            group_key = "group_id",
    time_key = "observed_at",
                            outcome_end_key = "outcome_end")

Arguments

Details

Imports require exact columns, finite numeric values, explicit numeric units, true/false boolean literals, and JSON-safe integers. Integer text contains signed digits only; IDs larger than 9007199254740991 are strings. Empty files require allow_empty. Duplicate JSON object keys are rejected. Null or empty text is missing only in nullable columns.

Relational tables have keys obligors(entity_id), observations(entity_id, observed_at), facilities(facility_id), defaults(episode_id), recoveries(cashflow_id). Observations require available_at; facilities require entity_id and nonnegative ead; defaults require entity_id and default_at; recoveries require episode_id, cashflow_at, nonnegative amount and kind. Entity foreign keys reference obligors; recoveries reference defaults. Duplicate/missing keys, invalid dates, negative amounts and recovery before default are reported rather than dropped.

The as-of join selects latest observed_at not after as_of, then latest available_at not after known_at. Same-date ties raise DATA_AMBIGUOUS_JOIN. Unmatched rows remain with NULL right fields. This is a fact/revision join, not an interval join.

First-default windows use (origin, origin+horizon] and known events only. Prior default excludes a row. An observed in-window event yields one, complete follow-up without default yields zero, and censored or unfunded rows have missing targets. A future outcome_available_at produces outcome_pending and a missing target unless default is already observed; calendar maturity alone is insufficient. Default wins a same-day exit tie. A rejected application never becomes a good loan label. Complete event reporting through followup_end is a caller assumption; reporting gaps require shortening follow-up. Cure/re-default targets are separate.

Workout reopening uses the preceding default's known cure date. Original EAD is not added across repeats. Unique cashflows remain attached to dated defaults and exactly one workout, with calendar days since first default. Currencies must match. Explicit closure with known closed_known_at alone closes a workout; cure and data-end do not. A cashflow after closure fails pending revised closure or explicit reopening. Amounts are undiscounted, and no LGD or discount rate is inferred.

Temporal splits connect all entity/group links transitively and reserve every component for its latest partition, including future rows beyond test end. Earlier rows are excluded. Training/calibration outcomes plus embargo end by their upper boundary; calibration/test observations are strictly after the lower boundary plus embargo. Test outcomes end by test_end. Empty partitions are reported. This evaluates unseen connected entities, not repeat-customer backtests. All routines are deterministic in-memory preparation, not empirical certification.

Value

A CalcResult with a list of named rows in $value, or a named report. All dates are ISO YYYY-MM-DD strings. Missing outputs remain NULL. Import provenance includes SHA-256 of original bytes. Validation returns valid, violations (one-based row numbers) and row_counts. Recovery returns episodes and cashflows. Splitting returns assignments, counts and group_overlap.

Examples

x <- list(list(entity_id = "A", as_of = "2024-02-01", known_at = "2024-02-01"))
y <- list(list(entity_id = "A", observed_at = "2024-01-01",
               available_at = "2024-01-05", release_id = "v1", score = 2))
join_as_of(x, y)$value

Explicit Dependence Models and Addressed Innovations

Description

Validate dependence matrices, fit Gaussian, t, positive Clayton and Gumbel copulas, transform shared draws, calculate default intersections and write portable innovation files.

Usage

validate_dependence_matrix(matrix, variable_ids, matrix_kind,
  repair_policy = "none", tolerance = 1e-10, context = NULL)
fit_copula(observations, variable_ids, family, input_kind,
    tie_policy,
  correlation_repair = "none", df = NULL, df_bounds = c(1, 100),
  target, training, model_id, model_version, context = NULL)
simulate_copula(dependence_model, independent_uniforms,
    variable_ids,
  auxiliary_uniforms = NULL, context = NULL)
joint_default_probability(p_a, p_b, event_order,
  dependence_model = NULL, joint_probability = NULL, context = NULL)
make_shared_innovations(scenario_ids, time_grid, factor_ids,
    stream_spec,
  seed, output_directory, chunk_rows = 1024, context = NULL)

Arguments

Details

Fitting uses Kendall tau inversion. The t family additionally profiles the joint copula likelihood over degrees of freedom with rank correlation fixed. Clayton and Gumbel fits are exchangeable and support nonnegative average tau. Pairwise composite likelihood diagnostics do not imply joint maximum likelihood or an uncertainty estimate. Shared frailty and scale draws are explicit. Defaults are lower-tail uniform events; latent correlation is not default correlation, and zero t correlation does not imply independence. Matrix repairs are disclosed and are not claimed nearest-correlation solutions. Addressed uniforms use the first32 SHA256 bits at cell midpoints; normal draws use two lanes and Box-Muller cosine. Uniform files match bitwise across languages; normal transforms may differ by floating-point rounding.

Value

CalcResult for matrix validation, simulated uniforms, scalar intersection or innovation manifest; FitResult with a declarative frozen copula model for estimation.

References

Yan (2007), Enjoy the Joy of Copulas, Journal of Statistical Software 21(4), doi:10.18637/jss.v021.i04. Native copula package fitCopula documentation.

Dated relationships and explicit risk adjustment contracts

Description

Granular native methods for typed relationship graphs, conditional support, shared guarantee caps, rating constraints, local override audits and static financial clearing.

Usage

build_entity_graph(entities, edges, relationship_policy, as_of,
    known_at,
  context = NULL)
resolve_parent_links(entity_graph, control_policy, context = NULL)
assign_default_groups(entity_graph, default_link_policy,
    facility_mappings,
  context = NULL)
apply_parent_support(standalone_risk, parent_risk, support_contract,
  joint_default_probability, context = NULL)
allocate_guarantee_coverage(guarantees, claims, allocation_policy,
    as_of,
  known_at, currency, context = NULL)
apply_guarantee_risk(borrower_risk, guarantor_risk,
    coverage_allocation,
  claim_id, loss_contract, joint_default_probability,
    context = NULL)
apply_rating_constraints(base_result, rating_scale, constraints,
  representative_pd_policy, context = NULL)
apply_override(base_result, override_request, policy, as_of,
    known_at,
  context = NULL)
solve_contagion_network(liability_matrix, external_assets,
    external_liabilities,
  entity_ids, currency, model_assumptions, asset_scenarios = NULL,
  tolerance = 1e-8, max_iterations = 10000, context = NULL)

Arguments

Details

Graph intervals are left closed and right open. Parent ambiguity and cycles remain unresolved; ownership never establishes common PD or support by itself. Support integrates conditional rescue in parent-alive and joint-default states, requiring funding sufficient for the whole stated rescue need. Its output is before transfer.

Guarantee allocation consumes each shared cap once across the supplied claims. Payment protection preserves borrower PD and reduces conditional monetary loss by min(covered amount, positive loss) times supplied conditional payout fraction. Previously guarantee-adjusted inputs are rejected even under a new effect ID. Several guarantors require a joint event model. Nominal cap and loss use the claim due-date valuation basis; discount state losses separately.

Rating rules are intersected and inconsistent bounds fail. Nonbinding constraints preserve continuous PD exactly. Grade-only policy exposes any PD/grade inconsistency; representative policy changes PD only on binding. Override history preserves original values and content hashes, rejects duplicate IDs, and records expired requests without applying. A changed evaluation cutoff requires replay from the original; copying an overridden current value is not an unadjusted baseline. Only the requested field changes.

Clearing uses limited liability, absolute priority and one pro-rata debt class. Native lpSolve maximizes payment fractions to select greatest Eisenberg–Noe clearing. A monotone lower iteration and greatest/lower gap diagnose numerical enclosure or multiple clearing points; a small iterate change alone does not prove uniqueness. Outside creditors do not recycle payments. Extreme amount dynamic ranges below 1e-10 of the largest amount fail explicitly. This is a static financial network, not a group PD estimator, dynamic insolvency model or multi-seniority waterfall.

Value

A CalcResult containing the graph, mapping, adjusted risk, immutable local artifact or clearing payments and sensitivity, with diagnostics and limitations.

References

Eisenberg, L. and Noe, T. H. (2001). Systemic Risk in Financial Systems. Management Science 47(2), 236–249. doi:10.1287/mnsc.47.2.236.9835.

Examples

a <- list(model="eisenberg_noe", seniority="single_pari_passu",
          bankruptcy_costs="none", outside_creditor="sink")
solve_contagion_network(matrix(c(0,50,50,0),2), c(20,10), c(50,50),
  c("A","B"), "EUR", a)$value$base$clearing_payments

Portfolio Loss Distribution, Tail Allocation and Concentration

Description

Weighted discrete loss quantiles and exact upper-mass expected shortfall, paired scenario changes, concentration and a homogeneous Gaussian factor benchmark.

Usage

credit_loss_distribution(loss_paths, weights, scenario_ids,
  levels = c(.95, .99), currency, context = NULL)
allocate_expected_shortfall(position_loss_paths, weights,
    scenario_ids,
  position_ids, level, currency, groups = NULL, context = NULL)
incremental_credit_risk(base_loss_paths, change_loss_paths, weights,
  scenario_ids, levels = c(.95, .99), currency, change_id,
    context = NULL)
credit_concentration_measures(positions, weight_field, group_fields,
  weight_unit, context = NULL)
asymptotic_factor_loss(pd, lgd, asset_correlation,
    systematic_factor,
  context = NULL)

Arguments

Details

VaR uses the left inverse CDF. ES selects exactly the upper 1-alpha mass, allocating partial boundary atoms proportionally; level zero is the mean and level one the maximum-loss limit. Signed hedge allocations reconcile to ES. These are supplied-distribution calculations with no inferred confidence intervals. HHI and entropy are descriptive, not regulatory limits. The factor benchmark assumes infinitely granular identical exposures; latent asset correlation is not default correlation. Its unconditional factor law does not establish empirical TTC calibration.

Value

CalcResult with the requested distribution, allocations, paired changes, concentration dimensions or conditional Gaussian benchmark.

References

Rockafellar and Uryasev, Conditional Value-at-Risk for General Loss Distributions. https://sites.math.washington.edu/~rtr/papers/rtr187-CVaR2.pdf

Frozen Held-Out Grouped Permutation Importance

Description

Compare paired prediction losses under explicit shared permutation plans, without refitting or selecting features.

Usage

credit_permutation_importance(predictor, X, y, model_contract,
  feature_groups, permutations, design, loss_contract, row_metadata,
  weights = NULL, strata = NULL, custom_loss = NULL,
  support_evaluator = NULL, support_contract = NULL, context = NULL)

Arguments

Details

Blocks are permuted jointly; outcomes and weights remain with recipients. Fitting-identity overlap, incomplete targets, future-known labels and invalid permutation plans are rejected. Positive-weight infinite log loss is not clipped and raises an explicit non-estimability error. Unsupported rows are reported and retained in denominators. Weighted feature correlation is descriptive; cross-block dependence can be broken. Only declared IID random permutations with at least two plans receive a standard error of the mean loss increase. This excludes cohort, model and selection uncertainty. Fixed plans receive none. No refitting or feature selection occurs; supplied provenance is checked but not authenticated.

Value

CalcResult with baseline and per-plan losses, grouped paired increases, permutation uncertainty, support findings, correlations and content hashes.

First-Default Probabilities from Period Hazards

Description

Converts conditional probabilities into cumulative first-default probabilities or unconditional period first-default masses.

Usage

cumulative_pd_from_conditional(conditional_pd, context = NULL)
marginal_pd_from_conditional(conditional_pd, context = NULL)

Arguments

Details

Survival through period k is the product of one minus each preceding hazard, including the current hazard. Cumulative PD is one minus survival. Period default mass is the current hazard times survival to its start. Calculations use log-survival for numerical accuracy; marginal masses are computed directly rather than by differencing cumulative probabilities. A unit hazard absorbs all remaining survival. Empty inputs return empty vectors. No cure, repeated default, interpolation or annualization is implied.

Value

A CalcResult whose value is a probability vector in chronological order.

Examples

cumulative_pd_from_conditional(c(0.1, 0.2))$value
marginal_pd_from_conditional(c(0.1, 0.2))$value

Historical Decision-Policy Backtest

Description

Audit frozen policy decisions, actual offers and funding stages, matched booked outcomes, declared scenario constraints and paired policy sensitivity.

Usage

decision_policy_backtest(evaluated_grid, grid_inputs,
    historical_cases,
  design, reference_candidate_id, uncertainty = "none",
  cluster_resamples = NULL, confidence_level = .95,
  tolerance = 1e-10, context = NULL)

Arguments

Details

Evaluation profile target_contract_version 1.1 requires a complete target with performance_origin first_draw_at and event default or first_default. The validation design repeats policy_target_contract_version 1.1 and performance_origin, preserving the exact original target. Missing version or 1.0 remains default-only; unknown versions are rejected. Observed decision stages do not imply observed economic outcomes. No lending action or fit is executed. Actual funding requires take-up and an offer. Whole offer hashes bind terms, so changed price/limit scenarios do not inherit factual credit labels. Unfunded outcomes are unknown. Offered limits require frozen offer contract_terms.limit. Mature matched-funded risk is separate from assumed offer economics. Schema 1.1 approval applicability and additional gates are replayed by the real evaluator. Declared scenario violations and historical inconsistencies block reconciliation. Agreement alone does not approve a policy. Full-policy causal value remains unidentified; this validator does not assert an off-policy identification design.

Value

CalcResult with ValidationResult, factual and scenario metrics, paired contrasts and cluster uncertainty, historical audits, constraints and counterfactual applicability.

References

Ehrhardt et al., Reject Inference Methods in Credit Scoring. https://adimajo.github.io/assets/publications/rejectInference.pdf

Local Limit, Quote and Treatment Scenario Economics

Description

Enumerates supplied local offers and action-conditioned cashflows with explicit risk, affordability and eligibility assumptions. No decision is executed and no causal price or treatment effect is estimated.

Usage

compute_limit_frontier(offer_candidates, constraint_profile,
  as_of, context = NULL)
compute_risk_based_price_grid(quote_candidates, pricing_profile,
  as_of, context = NULL)
evaluate_treatment_candidates(treatment_candidates,
    treatment_profile,
  as_of, context = NULL)

Arguments

Details

Targets specify physical default, positive integer horizon, time_unit=days and explicit conditioning. Limit candidates share one comparison target, although tenors may differ. Price targets span the entire quote tenor and end on its last payment. Offer-conditioned risk is supplied, never copied implicitly from a baseline borrower PD.

Limit EL is PD times conditional economic LGD times EAD. Concentration is existing group exposure plus candidate EAD divided by the declared fixed post-decision denominator. Numerical tolerance applies in the units of each comparison. Slacks, violations and binding flags remain separate from affordability and eligibility gates. Unknown affordability caps are infeasible. Feasible offers are enumerated without monotone-risk assumptions. Same-tenor domination means at least as much amount, no more EL and one strict improvement. No hidden cross-tenor utility or interpolation is used.

Pricing starts survival at one. Default and prepayment probabilities are mutually exclusive conditional on opening survival; masses are survival times their respective hazards. Surviving principal follows the supplied full amortization schedule. Fixed simple ACT/365F coupons use opening principal. Default before due misses coupon and exposes opening principal; recovery is principal times one minus LGD at the supplied later date. LGD describes undiscounted net recovery at that date, and may be signed. An already discounted LGD would double discount. Prepayment after coupon returns all opening principal plus its fee. Continuing accounts pay coupon and scheduled principal. Servicing cost applies to all opening survivors. Full surviving repayment is required; no residual value is assumed. Same-date funding discount factors must agree within a quote.

Booked NPV subtracts initial amount and origination cost, adds upfront fee and discounted expected cashflows. Funding enters only through discount factors. Reported EL is already reflected in recovery cashflows and must not be subtracted again. Per-offer NPV is takeup times booked NPV minus offer cost, paid once regardless of uptake. No demand elasticity or unique break-even rate is inferred.

Treatment expected recovery/cost cashflows are already event-probability-weighted within acceptance. Accepted value subtracts accepted upfront cost; total expected value mixes uptake and explicit rejected value and subtracts offer cost once. The baseline is an independent comparator. Unsupported predictions give null values and cannot be ranked. Supported but unaffordable treatments retain their value and infeasibility reason. Allowed supported feasible treatments rank by expected value, ties by ID. Historical selection and missing counterfactuals are not repaired by this conditional scenario comparison.

Value

CalcResult with SHA-256-bound LimitFrontier, PriceGrid or TreatmentEvaluation, candidate fingerprints, monetary components, separate feasibility reasons and declared assumptions.

Held-Out Policy Evaluation, Constrained Selection and Robustness

Description

Replays frozen local policies, separates supplied prospective offer economics from observed matched-funded evidence, and checks finite shared perturbations without inferring causal policy effects.

Usage

evaluate_decision_grid(policy_candidates, cases, candidate_inputs,
  observed_outcomes, evaluation_profile, evaluation_cutoff,
    context = NULL)
select_constrained_policy(evaluated_grid, context = NULL)
assess_policy_robustness(policy, baseline_cases,
    perturbation_scenarios,
  robustness_profile, as_of, context = NULL)

Arguments

Details

Evaluation profile target_contract_version 1.1 requires a complete target with performance_origin first_draw_at and event default or first_default. The validation design repeats policy_target_contract_version 1.1 and performance_origin, preserving the exact original target. Missing version or 1.0 remains default-only; unknown versions are rejected. Observed decision stages do not imply observed economic outcomes.

Evaluation roles are VALIDATION, OOF_VALIDATION or TEST; prediction design is HELD_OUT or OUT_OF_FOLD. OOF_VALIDATION requires OUT_OF_FOLD. Per-case model and calibration exclusion and pre-registration are caller assertions, not an independent audit. Register candidates, objective and constraints before declared outcome access. Policies use the same physical default target with integer day horizon. Historical policy knowledge cutoff equals the case decision date.

Approved cases need versioned offer_id, version, known_at, currency, target, risk_conditioning=CONDITIONAL_ON_THIS_ACCEPTED_OFFER, assumption_reference, support_status, and nullable takeup_probability, conditional_booked_expected_loss, conditional_booked_npv, ead and offer_cost. SUPPORTED assumptions require finite values; positive-weight unsupported approved offers make monetary metrics null. Per-applicant NPV is takeup times book NPV minus offer cost; book NPV already includes losses and funding. EL is a separate metric, not subtracted again. Nonapproval costs are explicit. All metrics use the same total applicant weight.

Unfunded rows cannot supply fabricated defaults, losses or profits. Funded outcome horizons begin at actual funding and equal the target horizon. Funding knowledge cannot precede funding. Complete labels need outcome knowledge no earlier than horizon end; incomplete labels remain null. Observed results require recommended approval, exact booked-offer hash match, funding known by cutoff and, for mature statistics, complete outcomes known by cutoff. Changed assumptions or versions are deliberately not treated as equivalent offers. Missing funding, unfunded, different offers and incomplete outcomes have distinct statuses. Rejected cases are never counted as nondefaults. Loss/profit bounds cover only missing outcomes inside the known matched-funded cohort, not unknown funding or causal policy value. Full candidate causal value remains null.

Selection accepts only VALIDATION/OOF_VALIDATION. TEST grids may be evaluated but cannot select policy. The objective is expected_npv_per_application and tie break candidate_id. Each constraint has distinct constraint_id, metric, operator (le/ge) and threshold. Metrics include approval_rate, refer_rate, decline_rate, blocked_rate, expected_loss_per_application, expected_npv_per_application and expected_funded_ead_per_application. Slacks and binding flags are recomputed. Unidentified candidates are infeasible. Objective/constraint-direction Pareto domination is checked before utility/ID ties. Empty feasible sets return INFEASIBLE, not an arbitrary candidate. Full candidate and policy bindings are verified; the selected policy is returned unchanged. Final assessment requires an untouched test period.

Robustness replays one fixed policy on a finite shared scenario set. Scenario weights are nonnegative and sum to one. Each case utility/feasibility map explicitly covers APPROVE, REFER, DECLINE and BLOCKED, with nullable finite monetary utility and nullable Boolean feasibility. Approved cases require offer_id, version, amount and annual_rate terms. These supplied terms are not computed by the rule policy. Action flips, changed input sources, both-approved amount/rate changes, feasibility and regret remain separate. A changed source is associated with a flip, not causal attribution.

Regret compares only explicitly feasible actions with known utilities. Unknown positive-weight case/scenario comparisons propagate null worst-case regret. Known infeasible action rates are accompanied by unknown-dimension findings. Worst-case mean regret ranges only over supplied positive-weight scenarios. Bootstrap validity and dependence are caller responsibilities; no automatic confidence interval, counterfactual identification or continuous uncertainty bound is asserted. No financial action is executed.

Value

CalcResult with bound DecisionGrid, PolicySelection or PolicyRobustness, input and artifact hashes, complete rule traces, separate assumption/outcome metrics, feasibility reports and explicit unidentified quantities.

References

https://scikit-learn.org/stable/modules/classification_threshold.html separates probability fitting and threshold selection and warns against tuning on the same training observations; reviewed 2026-10-13. Local contracts and economic algorithms here are library-specific.

Cashflow Double Test and Independent Default Triggers

Description

Evaluate simultaneous debt-service and value-coverage breaches, independent unlikeliness-to-pay and material-arrears triggers, and absorbing first default.

Usage

detect_cashflow_default(path_records, default_policy, currency,
    context = NULL)

Arguments

Details

The economic double test requires simultaneous DSCR and LTV breaches on the same facility and date. A declared policy may use it as a UTP indicator only while a debt-service shortfall remains after explicitly available liquidity. Buffers must be legally and operationally available by the payment date, net of prior claims; hypothetical collateral proceeds or uncommitted refinancing do not qualify. They never cancel other independently established UTP signals. Regulatory-reference default is UTP OR material arrears exceeding the applicable day threshold. Adequate value never cancels either trigger. Zero debt service has no DSCR breach; positive debt and zero assets have an LTV breach. Supplied evidence must cover the assessment subject, including relevant obligations outside the simulated facility. Material-arrears counters follow the referenced materiality rule and must advance by calendar time or reset within an interval. Obligor assessment requires a common facility grid. Default remains absorbing; cure and redefault require separate episode logic. Detection dates are grid endpoints, with the preceding nondefault endpoint reported. Initial defaults are left censored and excluded from initially performing PD risk denominators.

Value

A CalcResult containing policy and hash, facility_tests, assessments and first_default_events, including initial_default, last_nondefault_date, observed_through and default_time_status.

References

Basel Framework, CRE36.68–36.75: https://www.bis.org/committees/bcbs/basel-framework/standard/cre?allChapters=true&chapter=36.

Dated Drawn Exposure Schedules and Realized Conversion Factors

Description

Accrue balances through explicit dated events with known default timing, and retain raw realized CCF values with undefined-denominator statuses.

Usage

exposure_schedule(contract, events, annual_interest_rates,
    period_end_dates,
  day_count, default_timing, allow_overlimit, currency,
    context = NULL)
realized_ccf(reference_drawn, reference_limit, ead_at_default,
    reference_date,
  default_date, reference_known_at, default_known_at,
    evaluation_cutoff,
  currency, minimum_undrawn = 0, context = NULL)

Arguments

Details

Interest accrues on drawn principal between successive events using actual calendar days and the declared divisor. It is not capitalized. Principal and interest repayments are explicit separate events and cannot exceed the corresponding balance. The event ledger exposes every applied balance change. A bullet payment can give positive before-event exposure and zero after-event principal; interest persists until paid explicitly.

CCF is (EAD at default minus reference drawn) divided by (reference limit minus reference drawn). Negative and above-one CCF values are retained. Nonpositive undrawn commitment yields NULL CCF with NO_POSITIVE_UNDRAWN; a small positive commitment at/below the threshold yields BELOW_MINIMUM_UNDRAWN. No artificial floor, ceiling, limit-change adjustment, cancellation-right model, future borrower behavior or legal netting rule is inferred.

Value

A CalcResult. exposure_schedule returns a schedule and event ledger bound by schedule_hash, with facility/valuation conventions. Each period contains opening and closing principal, accrued interest, limit, nonnegative undrawn commitment, overlimit amount, exposures before/after endpoint events and selected conditional_ead. realized_ccf returns raw ccf, denominator_status, reference amounts, undrawn commitment, exposure increase and reconstructed EAD when defined.

Examples

realized_ccf(40, 100, 70, "2023-12-01", "2024-01-01",
  "2023-12-01", "2024-01-02", "2024-01-31", "EUR")$value

Signed Netting, Collateral Unit Ledgers and Frozen Closeout Exposure

Description

Apply explicit legal membership, margin, market and timing contracts to user-valued aligned counterparty scenarios.

Usage

aggregate_netting_set_exposure(trade_value_paths,
    netting_agreements, fx_paths,
  as_of, reporting_currency, valuation_convention, context = NULL)
simulate_collateral_margin(netting_paths, margin_policies,
    call_calendar,
  collateral_market_paths, context = NULL)
closeout_exposure_at_default(netting_paths, collateral_paths,
    default_events,
  intermediate_cashflows, freeze_timing, cashflow_convention,
    context = NULL)

Arguments

Details

Signed values aggregate only within sets before positive-part flooring. No cross-set offset, interpolation or automatic legal opinion is supplied. Prior settlements precede calls, followed by immediate settlements. Target units use declared valuation-date net value, price, FX and haircut. Pending units are included in the committed balance to prevent duplicate calls. MTA is measured as gross reporting market value; future pending settlements remain visible.

Positive collateral is recognized after received haircut. Nonremote posted collateral is a full nominal return claim, while posted haircut only sizes required posting. Enforceable posted-claim netting gives max(V-received+posted,0); otherwise exposure is max(V-received,0)+posted. Reuse does not imply reinvestment benefits. Initial units are the balance before first-date events; pre-snapshot unsettled obligations require an explicitly extended scenario.

At default, frozen units are revalued at closeout market, FX and haircut; later margin is ignored. Outstanding intervening claims are included exactly once. Same-day intraday default cashflows are outside this date-only contract. Missing path extension fails. The caller establishes the margin observation anchor; regulatory MPOR floors and capital approval are not inferred.

Value

CalcResult with signed net values and positive exposures, a bound collateral scenario artifact with unit balances, pending transfers and call ledger, or closeout components and exposures. Margin-period days use the declared observation anchor; default-to-closeout days are a separate output.

References

BCBS CRE50.15–19 and 50.23, published 5 July 2024, https://www.bis.org/committees/bcbs/basel-framework/standard/cre/50/inforce/2019-12-15/published/2024-07-05.

Frozen CCF Conditional Means and Direct EAD Distributions

Description

Fit historically aligned conversion-factor mean regressions or nonnegative direct-exposure distributions, and replay models or contractual schedules.

Usage

fit_ead_model(exposure_episodes, X_train, feature_order,
    target_mode, model_spec,
  target, training, model_id, model_version, currency,
    weights = NULL,
  minimum_undrawn = 0, fit_intercept = TRUE, context = NULL)
predict_ead(fitted_model_or_schedule, facility_covariates,
    feature_order,
  expected_target, reference_dates, default_dates, currency,
  reference_drawn = NULL, reference_limit = NULL,
    quantile_levels = NULL,
  context = NULL)

Arguments

Details

CCF is (EAD-drawn)/(limit-drawn), without clipping. Frequency least squares and HC0 covariance estimate an identity conditional mean. Reconstructed negative EAD fails rather than being floored. Mean standard errors hold reference amounts fixed; a CCF mean alone does not identify exposure quantiles.

Direct EAD models zero probability separately from positive lognormal amounts. The positive log regression uses maximum-likelihood residual variance, requires nonzero residual dispersion and full rank, and yields mean exp(mu+squared_sigma/2). Multiplication by positive probability gives total mean. Quantiles include the zero atom. none requires all positive targets; logit requires both classes; all_zero fits an exact point mass without positive-stage parameters. Positive-stage support is diagnosed separately. Predictive quantiles assume a correctly specified lognormal distribution and omit parameter-estimation uncertainty.

Predictors and knowledge dates are supplied explicitly; truth of source evidence cannot be inferred. Models use native declarative artifacts with input fingerprints, retained provenance and cross-language replay. Schedule prediction interpolates nothing and respects its before/after-payment default convention. No PD or regulatory conversion factor is supplied.

Value

FitResult with chronological diagnostics, excluded rows, regressions, optional positive-EAD gate, covariance, feature support and provenance. CalcResult predictions contain conditional_ead, dates and currency; CCF adds conditional_ccf and coefficient-based mean standard errors; direct EAD adds positive_probability and optional quantiles.

Estimate observed long-run annual default rates

Description

Aggregates explicitly selected complete mature annual cohorts available at the knowledge date.

Usage

estimate_long_run_pd(cohorts, selected_cohort_ids, knowledge_date,
    weighting,
  uncertainty, dependence_assumption, representativeness, target,
  confidence_level = .95, context = NULL)

Arguments

Details

Periods are start-inclusive, end-exclusive calendar years; February 29 anniversaries use February 28. Annual mean equally weights cohort rates. Pooled weighting divides total defaults by total denominator. Year-t intervals require independent nonoverlapping years and at least two cohorts; they are approximate and intersected with [0,1]. Exact binomial intervals require independent homogeneous obligors and nonoverlapping periods. Representativeness and independence remain unverified user assumptions. Repeated obligors and dependent defaults require other uncertainty methods. An observed long-run rate is not automatically TTC.

Value

CalcResult with value list pd, lower, upper and cohort, target, weighting and assumption diagnostics. Bounds are NULL when uncertainty is none.

References

Clopper and Pearson (1934), as documented in stats::binom.test.

Sampling Aware Default and Joint Loss Estimates

Description

Estimate risk for one initially performing subject, keeping exact IID binomial uncertainty separate from weighted or dependent scenario summaries.

Usage

estimate_simulation_risk(path_records, sampling_contract,
  confidence_level = 0.95, context = NULL)

Arguments

Details

IID estimates require unit weights and return a Clopper–Pearson interval. The zero-default upper bound is positive. A zero plug-in standard error is not certainty. Loss Monte Carlo standard error uses the IID sample variance and does not establish tail robustness. Weighted scenarios and dependent or QMC points receive no invented IID confidence interval. Kish effective weight count is a concentration diagnostic, not a substitute binomial trial count. Expected loss is the mean joint default loss. Conditional mean EAD and conditional mean LGD generally do not factor into it; EAD-weighted LGD provides the monetary factorization when defined. A zero-EAD default leaves individual LGD undefined. Economic losses and LGDs are not capped. Censored positive-weight performing trials are rejected instead of silently dropped.

Value

A CalcResult with PD, its applicable interval, expected loss, conditional loss/EAD/LGD, EAD-weighted LGD, sample standard errors where justified, effective weight count, excluded initial defaults, and explicit empty-risk-set status.

References

Clopper and Pearson (1934), Biometrika 26(4), 404–413, doi:10.1093/biomet/26.4.404. See also https://stat.ethz.ch/R-manual/R-devel/library/stats/html/binom.test.html.

Definition-specific tranche risk and paired structure stress

Description

Separates payment-event probability, ultimate capital loss and discounted contractual cash shortfall; compares explicit structures on shared scenarios.

Usage

estimate_tranche_credit_metrics(tranche_paths, target,
    discount_factors,
  confidence_level = .95, context = NULL)
stress_tranche_structure(baseline_paths, alternatives, target,
  discount_factors, confidence_level = .95, context = NULL)

Arguments

Details

Discounted shortfall is the positive part of principal promised at legal maturity plus coupons due on actual opening balances minus dated receipts. Signed components are also returned. This includes payment timing and differs from undiscounted principal loss. Loss outside the selected event remains explicit. Open runoff leaves ultimate loss and discounted shortfall unidentified.

Exact Clopper-Pearson intervals and sample-mean discounted loss standard errors apply only to declared IID equally weighted scenarios. Fixed weighted or dependent/QMC scenarios have no IID uncertainty. Conditional principal severity is tied to the selected event; zero event probability yields null severity.

Stress outputs amount and rate deltas with changed notionals disclosed. Different joint default mappings may share identical innovations. Individual scenario-risk intervals are not paired-difference confidence intervals. No causal or optimization claim is made.

Value

CalcResult containing bound TrancheCreditMetrics or TrancheStructureStress with source and innovation bindings.

References

https://stat.ethz.ch/R-manual/R-devel/library/stats/html/binom.test.html

Expected Default Loss

Description

Computes undiscounted expected default loss as PD times LGD times EAD.

Usage

expected_loss(pd, lgd, ead, context = NULL)

Arguments

Details

All inputs are scalar or one-dimensional numeric vectors. Length-one inputs expand; all other lengths must agree. Empty inputs and implicit vector recycling are rejected. The horizon is the horizon of the supplied PD. LGD and EAD must be deterministic or jointly specified conditional on default. Multiplying separately estimated marginal means is generally invalid when residual dependence remains. No currency is inferred and LGD is not capped. Binary exponent scaling preserves representable products when a naive intermediate multiplication would underflow; non-finite final results are rejected.

Value

A CalcResult whose value is the expected loss, in the EAD monetary unit.

Examples

expected_loss(0.02, 0.4, 100000)$value
expected_loss(c(0, 0.1, 1), 0.5, 200)$value

Documented Expert Probabilities, Agreement and Exact Finite Networks

Description

Fits elicited proper Beta priors, updates compatible observations, pools distributions, applies joint likelihood evidence, diagnoses judgments, and performs bounded exact Bayesian-network variable elimination.

Usage

fit_elicited_beta_prior(elicitation, elicitation_contract,
  max_cdf_error = .01, context = NULL)
update_expert_beta_binomial(prior, defaults, exposures,
  observation_contract, predictive_exposures = 0,
  credible_level = .95, context = NULL)
pool_expert_probability_distributions(components, weights,
  support, rule, exclusion_reasons = NULL, context = NULL)
update_odds_from_evidence(prior_pd, evidence_groups,
  dependence_contract, previous_evidence_ids = NULL, context = NULL)
check_expert_preference_consistency(comparisons, items,
  incomplete_policy, tolerance = 1e-8, context = NULL)
assess_expert_rating_agreement(first_ratings, second_ratings,
  ordered_grades, rater_pair, weighting, missing_policy,
  bootstrap_indices = NULL, confidence_level = .95, context = NULL)
compile_discrete_bayesian_network(nodes, network_contract,
  elimination_order = NULL, max_factor_cells = 1000000,
    context = NULL)
infer_discrete_credit_network(compiled_network, evidence,
  query_node = NULL, context = NULL)

Arguments

Details

Targets contain event default, positive horizon, time unit (days/months/years), PHYSICAL or RISK_NEUTRAL measure, and conditioning. Elicitation and observations retain distinct evidence. Moment priors use alpha=mean*concentration and beta=(1-mean)*concentration. Quantile fitting minimizes weighted squared CDF residuals in log shapes bounded by [-18,18], using native nlminb from starts (1,1), (2,18), (18,2), with residual, boundary and local-Jacobian identification checks. Prior concentration is not an empirical count.

Conjugate updates add defaults to alpha and nondefaults to beta. Posterior mean, median, credible interval and prior weight are separate. Future count probabilities integrate binomial probabilities over the posterior Beta. Existing evidence IDs cannot be reused; current prior availability cannot be later than observations. Complete conditionally iid sampling is a supplied assumption, not an inferred property.

Linear pools average normalized densities. Logarithmic pools exponentiate weighted log densities and normalize on declared finite support. Disjoint positive-weight support is an error. Histogram densities are constant in bins; cumulative output is at right endpoints. Opinion correlation is not independent evidence.

Likelihood updates add joint log likelihood ratios. Distinct groups require conditional independence under both outcomes. Zero/one priors remain dogmatic under finite positive likelihood ratios.

Preference priorities solve log least squares under a zero-sum log-priority constraint in each connected component. Pair and triangle residuals and DFS cycle witnesses diagnose rather than repair inconsistency. Uncertainty is judgment disagreement only; no sampling interval is asserted.

Weighted kappa compares observed agreement with independent empirical rater marginals. Expected agreement within 1e-15 of one is degenerate. Optional paired percentile bootstrap assumes independent cases unless caller supplies a valid cluster resampling design. Any degenerate replicate suppresses the interval instead of being silently discarded. Agreement does not establish predictive validity.

Networks require every parent-state combination once, nonnegative CPT rows normalized within 1e-12, an acyclic graph and dated parameter provenance. Exact log-domain factor multiplication and elimination marginalize absent evidence. Querying an observed node still checks evidence likelihood. Impossible evidence errors; positive likelihood underflow is separately flagged with finite log likelihood. Every intermediate obeys the frozen size limit. These are observational probability networks, not automatically causal models.

Value

A CalcResult whose value holds the sealed distribution/network, posterior/count distribution, pool, evidence contributions, preference diagnostics, agreement matrix/kappa, or exact queried probabilities and likelihood. Hashes identify frozen content, not authenticated authorship.

References

SHELF: https://shelf.sites.sheffield.ac.uk/. R Beta distribution: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/Beta.html. De Carvalho et al., logarithmic pooling: https://arxiv.org/abs/1502.04206v2. Crawford and Williams (1985), A note on the analysis of subjective judgment matrices, Journal of Mathematical Psychology 29, 387–405. Revelle, weighted kappa: https://personality-project.org/r/psych/help/kappa.html. Scutari, exact inference: https://www.bnlearn.com/about/teaching/eng-apm-inference.pdf.

Examples

target <- list(event="default", horizon=1, time_unit="years",
  probability_measure="PHYSICAL", conditioning="performing")
contract <- list(population_id="synthetic", reference="synthetic exercise",
  known_at="2024-01-01", expert_ids=list("one"),
  evidence_ids=list("elicitation-one"), target=target)
prior <- fit_elicited_beta_prior(
  list(method="moments", mean=.1, concentration=20), contract)
prior$value$alpha

Frozen Qualitative Rubrics and Dated Structured Evidence

Description

Compiles explicit categorical or numeric qualitative criteria and evaluates known, valid structured facts with evidence and conflict traces. No narrative inference or automatic conversion of points to PD occurs.

Usage

compile_qualitative_rubric(criteria, rubric_contract, score_policy,
  interactions = NULL, rating_limits = NULL, context = NULL)
evaluate_qualitative_evidence(compiled_rubric, observations,
    evidence,
  as_of, known_at, conflicts = NULL, context = NULL)

Arguments

Details

Criteria weights are normalized once at compilation and never renormalized because evidence disappears. Categorical levels must exactly partition declared categories. Numeric levels must be finite contiguous disjoint intervals, left closed/right open except the last upper endpoint, with points respecting any declared increasing or decreasing risk order. Higher points always indicate greater risk.

Each criterion declares required evidence types and a positive minimum number of independent source groups. Independence is a supplied assertion, not inferred from document counts. Missing policy is refer or worst. Non-applicable policy is separately refer, worst or fixed; fixed points must be within the criterion range and require valid supporting evidence. Unknown facts never become good or average observations.

Only verified, known and valid evidence supports facts. Validity intervals are half open. Observation time cannot exceed its knowledge time. Unknown/future/expired records are explicitly excluded. Multiple active identical facts can combine evidence; differing facts always refer until an explicit known valid resolution. Open conflicts, incompatible resolutions and unavailable resolution selections remain unresolved.

Weighted criterion points are summed, nonnegative adverse interaction deltas added and declared floor/cap applied. Unknown contributions produce an interval and no point estimate. Unknown interactions either refer or conservatively apply the adverse delta. Missing/assumed and non-applicable records do not receive invented levels. Active rating limits intersect; inconsistent or unresolved limits refer. Points are not mapped automatically to PD or grade.

Evidence coverage uses original normalized criterion weights. A conservative numerical score may coexist with REFER when evidence coverage or judgment remains unresolved; that score does not authorize a credit action. Content hashes record exact rubric and sorted input-record content, not author authentication.

Value

A CalcResult containing the sealed rubric, or score and interval, readiness/referral status, evidence coverage, criterion and interaction contributions, rating limits, conflicts, record exclusions, input hashes and declared cutoffs/target.

References

SHELF structured elicitation background: https://shelf.sites.sheffield.ac.uk/. These functions implement explicit local contracts and do not claim an elicitation workshop occurred.

Frozen Linear and Grouped Shapley Credit Attribution

Description

Exact linear-predictor contributions and exact or permutation grouped Shapley attribution with explicit reference distributions.

Usage

explain_linear_credit_prediction(linear_model, transformed_inputs,
    baseline,
  feature_groups, offset = NULL, baseline_offset = 0,
    context = NULL)
estimate_credit_shapley_values(predictor, record, background,
    model_contract,
  feature_groups, masking_contract, method = "exact",
    permutations = NULL,
  max_coalitions = 4096, context = NULL)

Arguments

Details

Linear contributions are coefficient times displacement from the reference and add on the linear predictor scale. Offsets are separate. Nonlinear response and response differences are shown separately; logit contributions are not additive PD changes.

Interventional masking breaks dependence between observed and masked features. Conditional distributions are supplied by the user and require independent validation. Exact Shapley weights are factorial subset weights. Permutation contributions telescope and are averaged; only declared IID random permutations with at least two draws receive a sampling standard error. Fixed permutation plans have none. This uncertainty excludes model and reference estimation uncertainty. Local accuracy and complete-coalition prediction consistency are checked. Coalitions are cached within the explicit budget.

For the product function x*y at (2,3) against (0,0), exact contributions are (3,3). With perfectly correlated equal-probability (0,0)/(1,1), explaining f=x conditionally at (1,1) yields (.25,.25) about baseline .5. Attributions explain a supplied reference game, not causal effects or credit approval.

Value

CalcResult containing contributions, reference and prediction, reconstruction/local-accuracy diagnostics and content hashes.

Frozen training transformations and auditable financial features

Description

Estimates preprocessing and supervised binning only from declared training rows, applies frozen artifacts, and constructs units-aware ratios and exact-calendar vintage-aware lags.

Usage

fit_preprocessor(train_features, feature_schema, target, training,
  model_id, model_version, context = NULL)
transform_features(features, fitted_preprocessor, context = NULL)
build_financial_ratios(financial_statements, ratio_spec,
    context = NULL)
build_distributed_lags(series, observations, lag_spec, frequency,
  gap_policy = "null", context = NULL)
fit_monotone_bins(x_train, y_train, feature_name, direction,
    min_bin_count,
  max_prebins, target, training, model_id, model_version,
    context = NULL)
compute_woe_iv(bin_counts, smoothing, woe_sign, context = NULL)

Arguments

Details

Standard scaling uses population standard deviation after imputation; robust scaling uses median and type-7 interquartile range. Zero scale becomes one and is reported. Missing and unknown categories have distinct declared handling; categorical mode ties are lexical. One-hot output names and lineage are frozen. Constant columns are retained and require explicit downstream model treatment.

Monotone bins preserve ties, build approximate count prebins, merge small bins with the nearest-risk neighbor (left on ties), then pool adjacent monotonicity violations. This is not claimed globally optimal. Missing has a separate bin exempt from support and monotonicity. Right-closed boundaries use observed maxima; the final NULL upper boundary has no upper limit. Transformation returns bin IDs, not WOE. WOE uses class-conditional smoothed shares; IV is sign-invariant and equals the sum of contributions.

Ratios require aligned statement dates and matching input units, with explicit missing/invalid-denominator statuses instead of infinities. Lags select actual calendar dates and the latest available vintage at the query information cutoff. Month/year lags clamp month ends. Future revisions cannot alter historical-cutoff replay.

Value

Fits return FitResult with ModelArtifact. Transform returns CalcResult containing design_matrix, feature_order and lineage with coverage diagnostics. Ratios and lags return row records with values, statuses and eligibility. WOE returns bin shares, weights of evidence, IV contributions and total IV.

References

Pool-adjacent-violators principle: https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.isotonic_regression.html. All support, unit and vintage policies are explicitly declared library contracts.

Finite-Domain Feasible Credit Counterfactual Search

Description

Enumerate an explicit candidate domain under structural constraints and declared costs, preserving search limits and failed candidates.

Usage

find_feasible_credit_counterfactual(predictor, record, candidates,
  model_contract, feasibility_contract, cost_contract, target,
    candidate_ids,
  max_candidates = 10000, policy_evaluator = NULL,
    policy_contract = NULL,
  context = NULL)

Arguments

Details

For changes divided by feature scales, L1 sums weighted absolute changes, L2 takes the square root of weighted squared changes, and change-count sums weights for exactly changed features. Candidates are visited by cost then ID. Infeasible candidates are reported without prediction. Every evaluated failure remains visible. A complete finite search reports FEASIBLE_SOLUTION or NO_FEASIBLE_SOLUTION. SEARCH_LIMIT remains the status whenever candidates remain unevaluated, even if a feasible witness exists. Optimality concerns only the supplied finite domain and declared costs, never continuous or causal global optimality. Policy targets replay full gates rather than equating low risk to approval. No real credit action is executed.

Value

CalcResult with search status, best witness, all evaluated/successful candidates, unevaluated IDs, domain scope, baseline and contract hashes.

See Also

compute_credit_ice

Estimate and Apply Binary Probability Models

Description

Fits an unpenalized Bernoulli likelihood natively in R and applies frozen coefficients using an explicit target and feature contract.

Usage

fit_binary_glm(X_train, y_train, link, feature_order, fit_intercept,
  target, training, model_id, model_version, weights = NULL,
  weight_type = "frequency", offset = NULL, max_iter = 100L,
  tolerance = 1e-10, context = NULL)
predict_pd(fitted_model, X, feature_order, expected_target,
  offset = NULL, context = NULL)

Arguments

Details

The estimator uses stats::glm.fit. No preprocessing, feature selection or calibration is inferred. Integer frequency weights mean repeated identical observations, not sampling, class-balancing or EAD weights. Covariance is inverse expected Fisher information with independent Bernoulli observations and unit dispersion. It is not cluster-robust. AIC and BIC use the Bernoulli likelihood, with BIC sample size equal to the sum of frequency counts.

The design must have full rank at relative singular-value tolerance 1e-10. A bounded linear program checks complete and quasi separation: maximize mean signed margin with every signed margin nonnegative and each direction coefficient between minus one and one, after column normalization. Mean margin above 1e-8 is rejected; feasibility tolerance is 1e-7. Nonconvergence, unusable information, excessive scaled likelihood score or fitted numerical probability boundaries raise controlled errors. No penalized fallback occurs. Offsets are supplied anew for prediction. Unsupported models or preprocessing are rejected. Generic prediction checks method/parameter-kind agreement, complete target and feature order; a supplied context knowledge date must not precede model availability. Common diagnostics bind canonical model/input hashes, training fingerprint/date, target and prediction count. Typed categorical, sequence and grouped inputs are preserved. No empirical fit or implicit calibration is performed, and PIT outputs are never relabelled TTC.

Value

Fitting returns a FitResult with model and value$model, including coefficients, expected-Fisher covariance, target, training, provenance and diagnostics. Prediction returns a probability-valued CalcResult. Fit diagnostics include likelihood, deviance, AIC/BIC, frequency count, rank, condition number, score and its threshold, iteration count and warnings.

References

R Core Team, https://stat.ethz.ch/R-manual/R-devel/library/stats/html/glm.html; lpSolve, https://cran.r-project.org/package=lpSolve.

Examples

target <- list(event = "default", horizon = 1, time_unit = "years",
  probability_measure = "PHYSICAL", conditioning = "performing")
training <- list(data_fingerprint = "analytic-counts", split_id = "train",
  known_at = "2025-12-31", data_origin = "SYNTHETIC")
fit <- fit_binary_glm(matrix(numeric(), 2, 0), c(0, 1), "logit",
  character(), TRUE, target, training, "intercept-example", "1",
  weights = c(7, 3))
predict_pd(fit, matrix(numeric(), 1, 0), character(), target)$value

Compatible PD Pooling, Held-Out Stacking and Frozen Model Graphs

Description

Combines aligned probability estimates, fits an actual held-out penalized logistic meta-model, executes a typed frozen graph and explains its result relative to explicit baseline inputs.

Usage

combine_pd_signals(pd_signals, weights, combination_scale,
  combination_contract, context = NULL)
fit_stacked_pd(out_of_fold_base_predictions, outcomes,
    base_model_ids,
  split_plan, meta_model_spec, target, training, model_id,
    model_version,
  weights = NULL, context = NULL)
evaluate_model_graph(model_graph, input_data, parameter_snapshots,
    context)
decompose_hybrid_prediction(hybrid_result, baseline_inputs,
  decomposition_scale = "native", context = NULL)

Arguments

Details

Pools normalize nonnegative weights. Log-odds pooling preserves finite-weight zero/one limits, rejects simultaneous positive-weight zero and one, and never clips. Mixtures use supplied row-specific conditional regime probabilities; they are not event unions. Mixture within- and between-regime Bernoulli variance sum to p(1-p). Dispersion is not uncertainty.

Stacking validates forward-time, entity and group separation. Every base lineage declares model ID/version, target, held-out origin, reference, training fingerprint, known-at and fit/preprocessing/selection row IDs. These must lie inside fold training rows. Base fits precede validation forecasts; known prediction dates cannot exceed forecasts. Used labels must be known and full horizons mature even for early known defaults. Negative labels require declared complete-through coverage. Date horizons use integer days or whole months, clamping month-end dates. Outer development/calibration/test entities and groups are disjoint. Only held-out development rows train the meta-model. Declared lineage is checked, not authenticated. Penalties are caller selected; no test-based selection occurs.

The meta-model uses actual native penalized binomial likelihood. Positive-weight constant bases are removed and frozen; all-constant models fit a real intercept and prediction flags deviations from the ignored constant training values. Log-odds inputs must be interior. No post-selection uncertainty is claimed. Use predict_pd with base-model IDs as feature order on the returned fitted artifact.

Graph contracts contain unit, target, ordered row IDs, as_of and known_at. Units are design, probability, loss_fraction and money. Designs add feature_order; money adds currency. Vectors exactly match rows without recycling. Supported operations are input (input_id), predict_pd (design port, snapshot_id), apply_calibrator (pd port, snapshot_id, base_model_version, optional groups), combine_pd_signals (ordered named signal ports and pooling parameters), log_odds_shift (pd port, aligned finite shifts, effect ID/reference), and expected_loss (pd/lgd/ead ports and loss assumption/reference). Frozen parameters must be known by the cutoff. New fitting, external state access, unknown operations/parameters and graph cycles are rejected. Financial feedback networks require their domain solver before insertion as typed inputs.

Expected loss requires performing PD, default-conditional LGD/EAD and matching event/horizon/measure/rows/dates, nonnegative monetary EAD and explicit deterministic_conditional_amounts or conditional_independence_EAD_LGD assumption. Raw economic LGD may exceed one or be negative. This product is not the expectation of generally dependent random exposure and severity. Repeated declared effects in ancestry or across loss legs are rejected. Identical computation content reuses a local cache. All snapshots, node results, effects and cache keys are retained with a content seal.

Decomposition verifies the seal and replays nodes with relative tolerance 1e-12 plus absolute 1e-14 for native numerical differences, with exact contracts and trace relationships. Each contribution changes one whole input from the common baseline, holding others at baseline. Final equals baseline plus those changes plus an explicit interaction remainder. These are neither causal effects nor Shapley values. Local native deltas for same-contract single-parent steps expose calibration/adjustment effects; they must not be summed across branches. Log-odds decomposition requires interior final probabilities in every scenario. Content hashes do not authenticate source evidence.

Value

A CalcResult with pooled values and diagnostics; a FitResult for stacking; a sealed graph CalcResult with replayable lineage; or baseline, input contributions, interaction remainder, local steps and exact reconciliation diagnostics.

References

Held-out stacked-generalization contract: https://sklearn.org/1.7/modules/generated/sklearn.ensemble.StackingClassifier.html. Pooling and decomposition implement the explicitly stated algebraic definitions.

Gaussian Importance Sampling for First Default and Loss

Description

Evaluates a local deterministic default evaluator under a complete Gaussian mean-shift proposal with ordinary density-ratio estimates.

Usage

importance_sample_defaults(standard_normal_innovations,
    proposal_shift,
  evaluator, importance_contract, context = NULL)

Arguments

Details

For x=z+mu the likelihood ratio is exp(-mu'z-mu'mu/2). Freeze the proposal independently of evaluated IID draws. The evaluator defines default, not the sampling transformation. Ordinary finite-sample PD estimates can exceed one and are not clipped. Self-normalized estimates are generally biased. Zero proposal defaults provide no exact target-PD upper bound. QMC/dependent designs have no IID standard error. Finite second moments are required but not established for arbitrary caller loss functions. Density-ratio numerical underflow and overflow are rejected.

Value

CalcResult containing ordinary PD and loss estimates, likelihood ratios and contributions, separate self-normalized PD, weight concentration and applicable IID sample standard errors.

References

Art B. Owen, Monte Carlo theory, methods and examples, Chapter 9. https://artowen.su.domains/mc/Ch-var-is.pdf

Dated Recovery Valuation and Raw Economic Workout LGD

Description

Value known realized recovery/cost cashflows using explicit discount and FX contracts, preserving open-workout incompleteness and raw economic LGD.

Usage

discount_recoveries(cashflows, discount_curve, fx_contract,
    default_date,
  evaluation_cutoff, reporting_currency, context = NULL)
workout_lgd(ead_at_default, discounted_net_recoveries,
    episode_status,
  default_date, evaluation_cutoff, currency, context = NULL)

Arguments

Details

Each contribution is amount times FX times discount factor; costs reduce net recovery. Exact date/currency keys are required, and duplicates are rejected. Future or not-yet-known cashflows are excluded visibly, not replaced with forecasts. The supplied discount and FX contracts define the economic valuation convention; no regulatory rate is supplied.

Loss is EAD minus net recovery. Raw economic LGD can be below zero or above one and is never clipped. An open workout does not establish a final loss merely because data ended. These methods do not estimate remaining recovery. Products use scaling to avoid intermediate overflow; nonfinite final products, sums or ratios fail.

Value

A CalcResult. discount_recoveries returns discounted gross recovery, costs, net recovery, signed individual contributions and excluded cashflows with reasons. Diagnostics retain valuation dates, discount/FX convention and input fingerprint. workout_lgd returns economic_loss, economic_lgd_to_date, economic_lgd, completeness, denominator status and the input exposure/net recovery. Final economic_lgd is NULL for OPEN workouts; both ratios are NULL at zero EAD while monetary loss remains.

Examples

workout_lgd(100, 80, "CLOSED", "2024-01-01", "2024-06-30", "EUR")$value
workout_lgd(100, -20, "OPEN", "2024-01-01", "2024-06-30", "EUR")$value

Declared Adverse-Scenario Conditional LGD Comparison

Description

Compare matched portfolio covariates under a frozen conditional LGD model and dated scenario policy, optionally with supplied paired model draws.

Usage

estimate_downturn_lgd(fitted_lgd_model, baseline_covariates,
    adverse_covariates,
  aggregation_weights, feature_order, expected_target,
    downturn_definition,
  baseline_precision_design = NULL, adverse_precision_design = NULL,
  precision_feature_order = NULL, model_draws = NULL,
  uncertainty_contract = NULL, interval_level = 0.95,
    context = NULL)

Arguments

Details

No scenario selection, regulatory surcharge or positivity floor is inferred. A model can imply a negative adverse-minus-baseline difference. Identical scenarios produce zero difference. Marginal training ranges do not establish joint support or causality. Uncertainty is NOT_ESTIMATED unless paired model draws are supplied. Each draw evaluates both scenarios; empirical linear quantiles (type7) summarize baseline, adverse and paired difference, retaining the declared interpretation. No statistical coverage guarantee is created from user-supplied illustrative draws.

Value

CalcResult with baseline_lgd, downturn_lgd, signed difference, unit-level predictions, normalized weights and uncertainty. Diagnostics retain policy/model hashes, explicit policy and baseline/adverse marginal range flags.

Frozen Conditional LGD Mean and Distribution Models

Description

Estimate fractional logit, variable-precision beta and structural zero-and-one inflated beta models, preserving explicit conditional targets.

Usage

fit_fractional_lgd(X_train, lgd_targets, feature_order, link,
    covariance,
  target, training, model_id, model_version, weights = NULL,
  weight_type = "frequency", fit_intercept = TRUE, max_iter = 200,
  tolerance = 1e-10, context = NULL)
fit_beta_lgd(X_train, lgd_targets, feature_order, mean_link,
    precision_link,
  target, training, model_id, model_version,
    precision_design = NULL,
  precision_feature_order = NULL, weights = NULL,
    weight_type = "frequency",
  fit_intercept = TRUE, max_iter = 500, tolerance = 1e-10,
    context = NULL)
fit_zero_one_inflated_lgd(X_train, lgd_targets, feature_order,
    component_spec,
  target, training, model_id, model_version,
    precision_design = NULL,
  precision_feature_order = NULL, weights = NULL,
    weight_type = "frequency",
  fit_intercept = TRUE, max_iter = 500, tolerance = 1e-10,
    context = NULL)
predict_lgd(fitted_model, X, feature_order, expected_target,
    default_condition,
  precision_design = NULL, precision_feature_order = NULL,
  quantile_levels = NULL, context = NULL)

Arguments

Details

Fractional logit maximizes the Bernoulli quasi likelihood and uses the sandwich covariance. Frequency weights enter the meat linearly; HC1 applies N/(N-k). Endpoints remain unchanged, and constant interior targets can yield zero empirical covariance. Full-rank and finite-MLE separation checks are enforced.

Beta uses shape parameters mu*phi and (1-mu)*phi, with logit mean and log precision. Analytic score and observed information include mean/precision cross derivatives. Covariance assumes independent, correctly specified beta observations. The effective normalized score threshold is max(sqrt(tolerance),1e-7), recorded in diagnostics. Native R reports gradient evaluations rather than iteration counts. Perfectly explained interior responses lack a finite precision estimate and are rejected.

The inflated model estimates boundary probability b, conditional-one probability o and an interior beta model. Endpoint probabilities are b*(1-o) and b*o; the mean is b*o+(1-b)*mu. All components require positive-frequency support and identifiable fits. Quantiles include the endpoint atoms. Component marginal support violations are reported; these do not establish joint support.

Prediction retains fitted provenance and rejects conflicting context target or model-version fields; currency is checked when declared in the target. Models are native declarative artifacts replayable without fitting code callbacks. Native optimizers can take different iteration counts. No downturn margin, regulatory approval or open-workout censoring correction is implied.

Value

FitResult with coefficients, covariance, support ranges, input fingerprints and fit diagnostics. predict_lgd returns CalcResult with mean_lgd, distribution, optional precision, endpoint probabilities and quantiles, plus model fingerprint and marginal training-range violations.

References

Papke and Wooldridge (1996), Journal of Applied Econometrics 11, 619–632. Cribari-Neto and Zeileis (2010), Journal of Statistical Software 34(2), doi:10.18637/jss.v034.i02. Ospina and Ferrari, Inflated Beta Distributions, https://arxiv.org/abs/0705.0700.

Residual Terminal Payouts and Open Workout Valuation

Description

Fit a cure and recovery hurdle with censored resolution timing, then value remaining terminal payouts without duplicating known cashflows.

Usage

fit_recovery_hurdle(episodes, cashflows, X_train, feature_order,
  stage_spec, target, training, model_id, model_version, currency,
  fit_intercept = TRUE, context = NULL)

estimate_open_workout_lgd(episodes, cashflows,
    fitted_recovery_model,
  X, valuation_contracts, future_cost_cashflows, feature_order,
  expected_target, stage_spec, as_of, currency, extrapolation,
  context = NULL)

Arguments

Details

The stage requires payout_definition RESIDUAL_TERMINAL_GROSS_RECOVERY_OR_CURE_VALUE; cashflow_pattern SINGLE_TERMINAL_PAYOUT_AFTER_REFERENCE; timing_outcome_independence CONDITIONAL_ON_REFERENCE_COVARIATES; cashflow_completeness COMPLETE_THROUGH_OBSERVATION_END; payout_horizon ULTIMATE; positive_distribution lognormal; timing_distribution weibull, lognormal or loglogistic; a nonnegative cure_residual_value_ratio; and a nonempty landmark_definition. The complete stage object must match prediction. Completeness is a user assertion about supplied histories, not an imputation of unreported payments.

The core target horizon is a resolution reporting reference. It does not truncate ultimate terminal payout. predict_lgd returns payout_horizon ULTIMATE and an undiscounted residual loss ratio, not whole-workout discounted LGD or loss at the finite reporting horizon. If cure probability is p, positive recovery probability among noncured cases is q, fixed cure ratio is c and positive log-ratio parameters are mu and variance v, expected terminal value ratio is p*c + (1-p)*q*exp(mu+v/2). Loss is one minus this ratio. Its quantiles respect both the zero-recovery and cure-value atoms and can be negative.

Default and all reference features must be known at the reference landmark. Only resolutions observed by observation end and known at training cutoff count as events. Other episodes are right-censored. Training ratios require positive residual exposure. At least one observed cure, one noncured zero payout and two positive noncured payouts are required, with identified nonseparated designs and positive lognormal dispersion. Observations have unit frequency.

Positive post-reference cash recoveries must occur at terminal resolution. Cure asset value must equal residual exposure times the declared cure ratio and cannot duplicate a terminal cash recovery. Costs are not gross payout targets. The closed-case outcome factorization assumes conditional independence of terminal outcome/amount and resolution timing, and independent right censoring given reference covariates. Model availability and feature definitions must match the prediction stage; later features are not silently substituted.

Open-workout valuation includes known past net flows once. Conditional remaining survival is S(age+t)/S(age). Resolution mass in each grid interval receives its right-endpoint default-anchored discount factor. Supplied future costs are deducted once without inferred costs beyond the schedule. Unknown tail discount value leaves final loss and final LGD null; within-grid amounts remain available. A supplied conditional average tail discount factor requires an economic assumption reference. Closed cases use actual cashflows and cure value without forecast. Zero EAD leaves the ratio null and preserves monetary loss.

Value

fit_recovery_hurdle returns a FitResult with frozen gate models, positive amount regression, censored AFT model, target, stage, provenance and identification diagnostics. estimate_open_workout_lgd returns a CalcResult with per-episode known cashflows, cure value, future terminal value within the grid, supplied future costs, unresolved probability and value, economic losses and nullable final LGD.

References

NIST/SEMATECH e-Handbook, Lognormal Distribution: https://www.itl.nist.gov/div898/handbook/eda/section3/eda3669.htm.

lifelines WeibullAFTFitter parameterization: https://lifelines.readthedocs.io/en/latest/fitters/regression/WeibullAFTFitter.html. The specific hurdle factorization and terminal stage are explicit modeling assumptions.

See Also

predict_lgd, discount_recoveries, fit_aft_model, predict_event_curves

Joint Path, Exclusive Event and Portfolio Loss Expectations

Description

Integrate losses on explicitly aligned probability scenarios. Preserve joint default, LGD, exposure, discount and currency dependence, with finite partition expectations and discrete portfolio tail allocations.

Usage

expected_loss_from_paths(default_indicator, lgd_paths, ead_paths,
  scenario_weights, discount_factors, scenario_ids, time_grid,
  currency, probability_measure, uncertainty = "none",
    context = NULL)
lifetime_expected_loss(marginal_pd, conditional_loss,
    discount_factors,
  time_grid, currency, probability_measure, context = NULL)
event_conditioned_loss(states, event_names, currency,
  probability_measure, context = NULL)
counterparty_expected_loss(default_indicator, net_exposure_paths,
  recovery_paths, scenario_weights, discount_factors,
    counterparty_id,
  netting_set_id, scenario_ids, time_grid, currency,
    probability_measure,
  uncertainty = "none", context = NULL)
aggregate_portfolio_loss(facility_loss_paths, portfolio_mapping,
  scenario_weights, fx_rates, scenario_ids, reporting_currency,
  probability_measure, quantile_levels, context = NULL)

Arguments

Details

For scenario s and period t, joint path loss is the sum of D(s,t) * LGD(s,t) * EAD(s,t) * DF(s,t). The probability-weighted mean integrates all supplied dependencies before aggregation. At most one first default prevents counting the same default again in later periods. The conditional independence comparator multiplies conditional mean LGD and conditional mean discounted EAD within each default period, weighted by its default mass. The difference from joint loss is the aggregate within-default covariance contribution; it is not necessarily positive. The counterparty comparator instead separates default from the complete recovery/exposure/discount product under the supplied scenarios.

The lifetime function multiplies marginal first-default mass by conditional joint loss and fixed period discount. Stochastic loss/discount dependence needs joint path integration. The event function integrates a full mutually exclusive partition exactly once, which allows default and transfer or support events to co-occur without counting their overlap twice. Event meanings and state losses remain explicit caller inputs.

Portfolio paths are aligned, converted with scenario FX, allocated, and summed before tail calculation. Weighted VaR uses the left inverse empirical CDF. Expected Shortfall averages exactly the upper (1 - level) probability mass: all mass strictly above VaR plus only the necessary fraction at VaR. Boundary mass is allocated proportionally across tied scenarios, so facility tail contributions sum to portfolio ES. Zero-weight scenarios do not affect the CDF. The result is not the sum of standalone facility VaRs or a simple conditional mean given loss greater than or equal to VaR.

Products use power-of-two scaling to avoid intermediate overflow when the final product is representable; nonfinite final products or sums fail with controlled NUMERICAL_FAILURE. Native floating-point underflow is still possible. No stochastic paths are generated and no dependence model is inferred by these methods.

Value

A CalcResult with monetary unit and explicit assumptions. expected_loss_from_paths returns expected_loss, path_losses, period_contributions, marginal_default_probability, default_weighted_lgd, default_weighted_discounted_ead, conditional_independence_loss, conditional_lgd_exposure_covariance_contribution and standard_error. Undefined conditional means in zero-default-mass periods are NULL. Diagnostics retain time grid, IDs, measure, residual and canonical input-path fingerprint.

lifetime_expected_loss returns lifetime_expected_loss, period_contributions and total_default_probability. event_conditioned_loss returns expected_loss, state contributions and probability_sum. counterparty_expected_loss extends joint-path results with independent_default_exposure_loss and default_exposure_dependence_contribution; diagnostics retain the netting-set identity and number of negative net values floored.

aggregate_portfolio_loss returns portfolios and unallocated_shares. Each sorted portfolio contains aligned path_losses, expected_loss, named facility_expected_loss and tail_metrics. Each tail includes level, value_at_risk, expected_shortfall and named facility_es_contributions.

Interpretation and limitations

The reported iid standard error measures Monte Carlo integration uncertainty only. It excludes parameter, model and time-discretization uncertainty and is not supplied for unequal scenario probabilities. Finite scenarios can instead represent an explicitly supplied discrete probability law.

All inputs must share their declared scenario meanings, target event, horizon, valuation date and measure. IDs only verify alignment, not economic compatibility. A risk-neutral discounted credit loss is not automatically a complete CVA or capital calculation. Legal netting, collateral closeout, own-default effects and regulatory acceptance are not inferred. Signed raw economic LGD is retained in general loss methods; the counterparty method explicitly uses bounded recovery.

References

Acerbi and Tasche (2002), On the coherence of Expected Shortfall, https://arxiv.org/abs/cond-mat/0104295, Definition 2.6 and equation (3.4), with sign and confidence-level convention translated to upper-tail monetary losses.

Examples

r <- expected_loss_from_paths(matrix(c(1, 0), 2, 1),
  matrix(c(0.6, 0), 2, 1), matrix(c(100, 100), 2, 1),
  c(0.5, 0.5), 1, c("default", "survive"), 1,
  currency = "EUR", probability_measure = "PHYSICAL")
r$value$expected_loss
lifetime_expected_loss(c(0.1, 0.05), c(40, 50), c(1, 0.95), c(1, 2),
  currency = "EUR", probability_measure = "PHYSICAL")$value

Release-Aware Macro Dynamics, Cycle Laws and Satellite Transmission

Description

Actual native VAR, Johansen cointegration, Gaussian hidden-regime and dynamic-factor estimation, explicit shared scenario innovations, conditional Gaussian and reverse stress, and dated satellite regressions. Forecast averages do not silently become through-the-cycle estimates.

Usage

transform_macro_history(history, variable_transforms, frequency,
    fit_window, 
    information_cutoff, context = NULL) 
fit_var(transformed_history, lags, deterministic = "constant", 
    covariance_estimator = "df_adjusted", training, target,
    model_id, 
    model_version, context = NULL) 
fit_vecm(level_history, rank, lags,
    deterministic = "constant_inside", 
    integration_order = "I1", training, target, model_id,
    model_version, 
    context = NULL) 
fit_regime_switching(transformed_history, regime_count, lags = 0,
    initialization, 
    max_iterations = 200, tolerance = 1e-08, training, target, 
    model_id, model_version, context = NULL) 
fit_dynamic_factor(transformed_history, factor_count,
    initialization, 
    ar_limit = 0.995, variance_bounds = c(1e-08, 10000),
    max_iterations = 300, 
    tolerance = 1e-08, training, target, model_id, model_version, 
    context = NULL) 
block_bootstrap_scenarios(residual_history, shared_draws,
    factor_ids, scenario_ids, 
    time_grid, frequency, block_length, scheme = "moving",
    weights = NULL, 
    context = NULL) 
simulate_macro_scenarios(fitted_model, innovations,
    initial_state = NULL, regime_uniforms = NULL, 
    output_space = "model", sampling_contract = NULL,
    context = NULL) 
condition_macro_paths(fitted_model, innovations,
    conditioning_points, initial_state = NULL, 
    sampler = "gaussian_equality", sampling_contract = NULL, 
    context = NULL) 
fit_satellite_model(target_series, macro_history, feature_spec,
    estimator, 
    link, split_plan, training, target, target_unit, model_id, 
    model_version, weights = NULL, covariance = "classical", 
    hac_lags = 0, residual_process = "white_noise", context = NULL) 
apply_satellite_model(fitted_satellite_model, scenario_cube,
    idiosyncratic_innovations, 
    initial_target_state, context = NULL) 
reverse_stress_scenario(fitted_model, loss_model, horizon,
    target_loss, shock_metric, 
    initial_state = NULL, shock_bounds = NULL, initial_shock = NULL,
    
    tolerance = 1e-07, max_iterations = 500, context = NULL) 

Arguments

Details

Transformation fitting uses training-only release vintages and retains an original-level inverse anchor. Later revisions cannot alter frozen training values. VAR reports estimated companion roots without repair. VECM preserves explicit rank, integration and deterministic placement, including rank-zero and one-lag cases. Trace statistics do not automatically infer rank.

The hidden-state EM likelihood uses regime-specific regressions on observed lagged levels. Every supplied start is evaluated and failures retained. Regimes are ordered by intercept, filtered probabilities are distinct from training-window smoothed probabilities, and local convergence is explicit. A one-regime fit reproduces the corresponding Gaussian regression.

Dynamic factors use genuine Gaussian Kalman likelihood with latent and idiosyncratic AR(1) states, identified loading anchors and explicit parameter bounds. Missing cells are integrated out without future imputation. Native R uses KFAS. Frozen artifacts retain observation/transition/covariance matrices and the terminal filtered distribution. Nonconvergence remains visible in fit diagnostics.

Stationary Gaussian TTC initialization solves equilibrium moments only for stable constant dynamics. Unit-root VECMs and deterministic trends fail this law. Declared cycle mixtures verify phase-weight mass but do not establish the caller's economic equilibrium assumptions. Regime stationary probabilities alone do not establish an autoregressive joint invariant law. Factor forecasts include filtered state uncertainty through explicit initial draws. All output cubes preserve the full sampling contract, named axes, measure and information cutoff.

Gaussian equality conditioning updates the entire correlated path and includes latent or stationary initial uncertainty. Support checks use unit-normalized covariance. Singular density is labeled in standardized-support coordinates; full-rank density uses physical coordinates. Density is not probability of an exact continuous event. Conditional stress retains its parent PIT/TTC scope. Reverse optimization reports a local feasible shock or explicit failure; it makes no general global optimality or infeasibility claim.

Satellite fits exclude future releases, immature training labels and held-out target values. WLS covariance conventions distinguish precision weights from binary GLM frequency weights. AR residual pairs do not cross entity boundaries or calendar gaps. Application exposes baseline, macro and idiosyncratic link contributions; zero macro deviations preserve a supplied baseline in delta mode. Shared macro paths and their cycle contract are preserved.

Value

FitResult containing a common portable ModelArtifact for estimators; otherwise CalcResult with a MacroHistory, innovation cube, ScenarioCube or reverse-stress optimization result and diagnostics. Artifact replay uses explicit innovations, not merely matching random seeds.

References

https://www.statsmodels.org/stable/vector_ar.html

https://search.r-project.org/CRAN/refmans/urca/html/ca.jo.html

https://search.r-project.org/CRAN/refmans/KFAS/html/SSModel.html

Local Market Credit Pricing, Calibration and Historical Physical PD

Description

Eight native methods with explicit risk-neutral cashflows, quoted-population historical physical estimation, and identified and assumed parameter diagnostics.

Usage

build_credit_quote_snapshot(quotes, instruments, discount_curve,
    snapshot_contract, context = NULL)

cds_expected_cashflow_legs(hazard_curve, discount_curve,
    cds_contract, context = NULL)

bootstrap_cds_hazard_curve(quotes, discount_curve,
    calibration_contract, context = NULL)

risky_bond_expected_cashflows(hazard_curve, discount_curve,
    bond_contract, context = NULL)

calibrate_bond_hazard_curve(quotes, discount_curve,
    calibration_contract, context = NULL)

diagnose_market_credit_identifiability(calibration,
    alternative_assumptions, diagnostic_contract, 
    context = NULL)

fit_market_to_physical_pd_bridge(observations, feature_order,
    bridge_contract, folds, model_id, 
    model_version, context = NULL)

compare_cds_bond_credit_basis(cds_calibration, bond_calibration,
    comparison_contract, context = NULL)

Arguments

Details

Common contracts

Python contracts are dictionaries with array fields as lists. Native
R uses named
lists for objects and unnamed lists for JSON arrays, including
empty/single arrays.
All date fields are ISO calendar dates (YYYY-MM-DD). The information
clock is
calendar-day resolution; use a separately constructed
cutoff-eligible daily input
when intraday ordering matters. Public functions accept optional
context; its
measure, currency, valuation date and knowledge cutoff must be
compatible where
applicable. Numeric inputs must be finite.

Both hazard_curve and discount_curve contain:

| Field | Meaning |
| --- | --- |
| as_of, known_at | Valuation origin and input vintage; knowledge
cannot follow valuation. |
| currency | Explicit matching currency identifier. |
| time_basis | Exactly ACT/365F: actual calendar days divided by
365. |
| knots | Strictly increasing positive interval ends in years from
origin. |
| rates | One rate per interval; final rate continues after the
final knot. |
| extrapolation | Exactly FLAT_LAST_RATE, explicitly authorizing
that continuation. |
| probability_measure | Required RISK_NEUTRAL for hazard curves. |

Hazards are nonnegative. Discount forward rates may be negative.
Discount factors
are D(t)=exp(-integral r) and survival is S(t)=exp(-integral
lambda).
Very large exponents fail explicitly instead of overflowing or
silently producing
zero survival. Curve fitting records the latest quote/discount input
knowledge
in the fitted hazard vintage; an earlier known discount curve can
price a later
eligible quote at the same valuation origin.

Instrument contracts share as_of, currency,
probability_measure=RISK_NEUTRAL
and default_settlement=AT_DEFAULT. Calibration additionally requires
instrument_type (CDS or BOND), issuer_id and seniority on every
instrument.
Each calibration fits a single issuer and seniority.

Local quote snapshot

build_credit_quote_snapshot(quotes, instruments, discount_curve,
snapshot_contract, context=None)

instruments is a nonempty array with unique instrument_id and
complete
instrument contracts. Each quote has unique quote_id, instrument_id,
quote_type, unit, value, observed_at, known_at, and optional paired
bid, ask. Knowledge cannot precede observation. Supported
conversions are:

| quote_type | Accepted units | Normalized unit |
| --- | --- | --- |
| CDS_SPREAD | BASIS_POINTS, FRACTION | Annual coupon fraction |
| CDS_UPFRONT | PERCENT_NOTIONAL, FRACTION | Fraction of notional |
| BOND_CLEAN_PRICE | PER_100_PAR | Price per 100 outstanding par |

snapshot_contract requires as_of, known_at, currency, integer
maximum_age_days >= 0, stale_policy (EXCLUDE, KEEP_FLAGGED) and
bid_ask_policy (REQUIRED, OPTIONAL). It returns all normalized rows
with
eligible and findings. Future information, currency mismatch,
crossed/outside
bid-ask quotes and conflicting values for the same
instrument/type/vintage are
ineligible. Stale rows are either excluded or explicitly retained
with a flag.
Different seniorities remain separate and appear in
cross_seniority_groups.
The snapshot is canonically hashed. It never downloads market data.

The calibration quotes argument accepts eligible normalized snapshot
rows
(snapshot.value['quotes'] filtered by eligible) or equally explicit
local
rows. Each needs quote_id, eligible=True, observed_at, known_at,
instrument, quote_type, normalized_unit, value. Quote knowledge must
not
exceed valuation. Eligibility assertions supplied without a snapshot
remain the
caller's data-preparation contract; calibration cannot attest to
unseen quotes.

CDS cashflows and bootstrap

cds_expected_cashflow_legs(hazard_curve, discount_curve,
cds_contract, context=None)

cds_contract additionally requires recovery in [0,1], nonnegative
annual
coupon, protection_start, maturity, settlement_date, signed
accrued_premium, signed upfront, and premium_mode. Positive upfront
and
settlement-accrued amounts are paid by the protection buyer;
negative amounts are
rebates received by the buyer. All results are per unit notional.

* CONTINUOUS: premium integrates D(t) S(t) over the protection
interval.
  Settlement accrued premium must be zero. This provides the exact
flat oracle
  par_spread=(1-recovery)*hazard.
* SCHEDULED: requires day_count (ACT/360 or ACT/365F), boolean
  accrual_on_default, and contiguous schedule rows with
accrual_start,
  accrual_end, payment_date. Accrual end cannot exceed maturity; the
last end
  equals maturity. The first period can start before valuation.
Coupon payment
  can lag accrual end: survival is tested at accrual end and the
surviving coupon
  is then owed at payment, with no additional post-maturity default
risk.
  Default accrual integrates actual linear accrued fraction,
including any
  pre-valuation portion, over the covered default interval.

Protection PV is (1-R) integral D S lambda. The premium and
accrued-default
annuities are separate.
fair_upfront=(protection-coupon*risky_annuity)/D(settlement)
-accrued_premium; the residual discounts the difference from
supplied upfront.
par_spread excludes settlement accrued cash, so it is the ratio of
protection
PV to total risky annuity, not an implicit rescaling of a supplied
accrued quote.
The payment/accrual conventions are deliberately explicit.

bootstrap_cds_hazard_curve(quotes, discount_curve,
calibration_contract, context=None)

The contract requires fixed recovery, positive maximum_hazard and
positive
price_tolerance. Maturities must be positive and distinct. Spread
quotes set
coupon to the normalized spread and upfront to zero, and require
zero settlement
accrued premium. Upfront quotes retain the instrument's fixed coupon
and accrued
settlement amount. A bounded scalar root is solved for each next
hazard segment.

The returned sealed MarketCreditCalibration includes feasible, the
fitted or
partial hazard_curve, quote residuals, survival, solver bracket
residuals and
boundary statuses. A missing nonnegative bracket or unmet residual
tolerance is
reported; negative hazards are never clipped. Partial curves are not
accepted by
basis/identifiability methods. Local knot sensitivities identify
their fixed-prior-
knots scope; curve_sensitivities additionally inverts the full
quote/hazard
Jacobian for recursive quote and recovery effects. Full
sensitivities are withheld
at a parameter boundary or deficient rank. A perfectly fitted zero
spread with
100% recovery does not identify a hazard.

Bonds and hazard calibration

risky_bond_expected_cashflows(hazard_curve, discount_curve,
bond_contract, context=None)

The contract requires recovery, fixed liquidity_spread,
recovery_convention (PAR, TREASURY, MARKET_VALUE),
cashflow_assumption (CONTRACTUAL_NONOPTIONAL,
EXOGENOUS_OPTION_SCENARIO),
positive outstanding_principal, nonnegative accrued_interest,
settlement_date=as_of and nonempty schedule. Each schedule row has
strictly
increasing future payment_date, nonnegative coupon_amount,
principal_amount.
Principal payments must sum to outstanding principal. Optional
cashflows require
an additional option_scenario_id; an externally supplied scenario is
not an
option price. Bond prices are at same-day settlement, with no hidden
funding or
forward-settlement adjustment.

* PAR: recovery at default is R times remaining outstanding
principal. The
  principal decreases only at scheduled amortization dates.
* TREASURY: recovery is R times the then risk-free value of all
remaining
  promised coupon/principal payments. This differs from recovery of
par.
* MARKET_VALUE: recovery is R times pre-default value. The
deterministic
  valuation discounts promised cashflows at r + liquidity +
(1-R)*lambda.

Survival-weighted promised coupon and principal PVs remain separate
from recovery
PV under all conventions. For market-value recovery, recovery PV is
the exact
residual between the reduced-form value and survival-weighted
promised payments.
Liquidity is an exogenous discount wedge applied to both promised
and default
cashflows. Treasury's replacement claim uses the risk-free curve
without that
wedge. Output separately reports the risk-free, credit-only and
credit-plus-
liquidity prices and their additive price effects. Clean price is
dirty price
minus supplied accrued interest; clean_price_per_100 divides by
outstanding par.
No credit/liquidity attribution is identified merely by this
arithmetic.

calibrate_bond_hazard_curve(quotes, discount_curve,
calibration_contract, context=None)

Each bond quote additionally requires strictly positive
price_standard_error
in price-per-100 units. The contract requires hazard_knots,
initial_hazards,
positive maximum_hazard, fixed recovery, fixed liquidity_spread,
positive
integer max_evaluations, and
weight_interpretation=KNOWN_GAUSSIAN_PRICE_STANDARD_ERRORS. Native
bounded
nonlinear optimization minimizes the sum of squared price residuals
divided by
those known errors. Python uses bounded least squares; R uses native
L-BFGS-B
with a numerical pricing Jacobian. Solver iteration counts need not
match.

Outputs retain raw/weighted residuals, fitted prices, objective,
convergence,
Jacobian, column-scaled rank/condition, active boundaries and
covariance when
available. Intervals are untruncated local 95% Wald approximations
conditional
on fixed recovery/liquidity, model correctness and supplied Gaussian
price errors.
They are withheld at boundaries or deficient rank; they are neither
global
profile intervals nor evidence that recovery or liquidity was
learned from data.
Overparameterized fits remain flagged instead of claiming unique
identification.

Identifiability and basis

diagnose_market_credit_identifiability(calibration,
alternative_assumptions, diagnostic_contract, context=None)

calibration is a sealed feasible CDS/bond calibration value.
alternative_assumptions is a nonempty array with unique scenario_id,
recovery
and, for bonds, liquidity_spread. Every alternative actually refits
hazards.
diagnostic_contract requires nonnegative equivalence_tolerance, one
positive
quote_standard_errors value per quote, two-element recovery_bounds,
and for
bonds two-element liquidity_bounds. Bounds must contain the
reference and all
alternatives. The objective is squared standardized quote residuals,
optionally
plus independent recovery information.

Optional independent_recovery_observation requires value, positive
standard_error, nonempty evidence_id, and eligible known_at. Its
independent
Gaussian measurement assumption is explicit. The joint Jacobian
includes hazard,
recovery and (bond) liquidity columns plus the independent
observation row.
Profiles report actual alternative hazards/objectives, failed
feasibility and
near-equivalent sets. This is a finite declared scenario profile,
not a search
proving the global identified set. The normal equivalence tolerance
is a supplied
objective tolerance, not an automatically inferred confidence
threshold.

compare_cds_bond_credit_basis(cds_calibration, bond_calibration,
comparison_contract, context=None)

The comparison contract requires fixed recovery,
bond_liquidity_spread,
funding_treatment=NO_ADDITIONAL_FUNDING_WEDGE, integer
maximum_quote_age_days, and a full comparison_cds_contract. The
method checks
issuer, seniority, currency, curve valuation/vintage, exact discount
conventions,
recovery, liquidity, quote freshness and maturity coverage.
Mismatches return
comparable=False, explicit findings and null basis, rather than an
unmarked
number. Matching curves are repriced with the same CDS schedule. The
sign is
CDS par spread minus bond-implied equivalent CDS par spread. Output
includes both
spreads, basis in fractions/basis points and survival difference. A
matched
synthetic world gives zero basis. This is a convention-conditional
market
comparison, not a pure default signal or trading recommendation.

Historical physical-PD bridge

fit_market_to_physical_pd_bridge(observations, feature_order,
bridge_contract, folds, *, model_id, model_version, context=None)

Each observation contains unique observation_id, entity_id, as_of,
label_known_at, complete_through, market_curve_as_of,
market_curve_known_at,
market_curve_hash, features_known_at, features mapping exactly the
declared
feature names, binary outcome, default_date (null for no observed
default),
initial_state=PERFORMING, and
outcome_origin=OBSERVED_ISSUER_DEFAULT.
Market features/curve vintages must be available by the prediction
date; target
windows must be fully mature and available by fit cutoff. A default
must occur
within the declared target horizon. Feature hashes document supplied
lineage;
the method does not reconstruct unseen historical quote books.

The bridge contract requires:

* target: event=first_default, positive horizon, time_unit in
days/months/
  years, probability_measure=PHYSICAL,
conditioning=initially_performing.
  Date windows use integer days or an integer number of calendar
months, clamping
  calendar month ends.
* feature_probability_measure=RISK_NEUTRAL,
market_feature_contracts: one
  object per feature with probability_measure=RISK_NEUTRAL, positive
horizon,
  time_unit and nonempty definition. Feature horizons may differ
from physical
  target horizons because the learned conditional model is explicit.
* outcome_origin=OBSERVED_ISSUER_DEFAULT, nonempty
target_population,
  population_selection=MARKET_QUOTED_ISSUERS,
  transformation=IDENTITY_PREDECLARED_FEATURES, known_at, split_id,
  link (logit, probit, cloglog), and optional data_origin.

folds contains nonempty unique zero-based train and validation index
arrays.
Training and validation entities are disjoint, validation rows occur
in at most
one fold, and every training label is known strictly before the
earliest
validation prediction date. Each fold fits an actual Bernoulli GLM;
no full-fit
predictions are substituted. Warmup rows have null out-of-time
predictions.
The final model fits all eligible observed rows for subsequent
prediction.
Out-of-time Brier/log loss and entity-cluster Brier uncertainty use
only real
observed outcomes. Native GLM coefficient covariance is also
retained. Separation,
rank deficiency and nonconvergence fail explicitly.

The frozen model uses common ModelArtifact serialization and generic
predict_pd dispatch, requiring matching features and the PHYSICAL
target. An
estimated conditional historical association is not a universal
Radon–Nikodym
measure change, a causal premium decomposition, an
unquoted-population estimate,
or a justified TTC estimate. Selection into the quoted population
remains an
explicit limitation. There is no automatic identity Q-to-P
conversion.

References and checks

The recovery distinctions and reduced-form valuation follow the definitions in Darrell Duffie, Defaultable Term Structure Models with Fractional Recovery of Par (1998). ISDA CDS Standard Model documentation illustrates the broader contractual conventions necessary for standard-model reproduction; this implementation makes no such equivalence claim.

Default tests include actual fits, independent closed-form prices, independent accrual quadrature, amortization conservation, inverse quote calibration, recursive sensitivity refits, nonidentification with and without independent recovery data, zero-basis matching, temporal/entity leakage and frozen model interchange. CREDITRATINGSYSTEM_EXTENDED_TESTS=true adds the recovery/liquidity/recovery- convention calibration grid. Default CI avoids that exhaustive grid.

Value

A CalcResult with pricing, quote snapshot, calibration, identifiability or basis values; the historical bridge returns FitResult with a common ModelArtifact. Infeasible CDS calibrations retain explicit partial results. The details describe all result fields and limitations.

See Also

predict_pd, serialize_fitted_model, load_fitted_model

Dated Local Model Roles and Monitoring Escalation

Description

Preserve model-role event history and apply explicit monitoring policies without external deployment or recalibration.

Usage

assign_model_role(history, registry, assignment, context = NULL)
rollback_model_assignment(history, registry, target_event_id,
  rollback_event, context = NULL)
classify_monitoring_findings(metric_panel, policy, context = NULL)

Arguments

Details

Roles are champion or challenger. A profile/purpose has one champion slot named champion; named challenger slots allow parallel evaluations. Events append in nondecreasing knowledge order. Repeating identical content is idempotent; conflicting IDs or equal role/slot effective boundaries fail. Registered model content and exact feature order must match. References are user assertions.

select_effective_model accepts the history and reconstructs champion intervals using only events known at its cutoff. Intervals include the start and exclude the next known event's start. Events remain unchanged; a backdated later-known assignment cannot affect an earlier knowledge replay. History alone has no final expiry. Rollback is a new event and does not cancel independently scheduled future changes.

Monitoring operators are gt, ge, lt or le. Persistence is a positive integer; severity is WARNING or CRITICAL. Every expected series is required, so an entire absent series is insufficient information. Supported input statuses are OK, MISSING_OUTCOME, INSUFFICIENT_INFORMATION, TECHNICAL_ERROR and NOT_APPLICABLE. Missing information breaks the breach streak. A premature breach is WATCH; a known nonbreach is WITHIN_POLICY. The trace retains every period; overall status describes only the last declared period. Precedence is technical error, critical, warning, missing outcome, insufficient information, not applicable, watch, within policy. Only persisted warning/critical breaches receive the supplied action code.

Value

A CalcResult containing a content-hashed append-only assignment history or a monitoring status with full rule trace.

Mature-outcome monitoring, distribution drift and paired diagnostics

Description

Local quantitative monitoring under explicit coverage, maturity, cluster resampling and null-path contracts. No model promotion or recalibration is performed.

Usage

build_monitoring_cohorts(predictions, outcome_events, coverage,
  target_contract, evaluation_cutoff, context = NULL)
monitor_rating_cohort(cohort, group_by, reference_window,
    minimum_count,
  uncertainty, cluster_resamples = NULL, confidence_level = .95,
  context = NULL)
detect_credit_feature_drift(reference, current, feature_order,
    method,
  permutations, multiple_testing, dependence_assumption,
  reference_weights = NULL, current_weights = NULL,
    bandwidth = NULL,
  frozen_bins = NULL, smoothing = .5, permutation_units = NULL,
  context = NULL)
detect_pd_calibration_drift(residuals, method, parameters,
    null_paths,
  null_spec, alpha, validation_null_paths = NULL,
  confidence_level = .95, context = NULL)
compare_challenger_cohorts(champion, challenger, cluster_resamples,
  selection_protocol, confidence_level = .95, context = NULL)
diagnose_fairness_slices(cohort, group_labels, reference_group,
  use_basis, minimum_count, risk_threshold, cluster_resamples,
  confidence_level = .95, context = NULL)

Arguments

Details

The default window is open at forecast time and closed at forecast plus horizon_days. Zero labels require mature, fully covered windows. Early known defaults remain positive but ineligible before horizon maturity. Prior defaults/exits exclude forecasts; later exits censor unobserved outcomes. Cure never removes the first default within the event horizon. LGD and incomplete workout labels are not treated as binary default observations.

Migration grades derive from dated, known rating events; prediction outcome_grade is ignored. Origin grade carries forward only under complete mature coverage. Coverage used for migration must also certify complete rating-event capture. First default uses the declared default grade. Hashes detect bundle edits but do not prove source declarations.

Cohort metrics use weighted AUC/Gini, Brier, exact log loss and calibration errors. One-class AUC is null. Impossible log loss retains infinite status and null numerical loss. Exact Clopper–Pearson rate intervals require unit weights, unique clusters and independent homogeneous Bernoulli outcomes. Bootstrap quantiles use type 7; unidentified replicate counts and degenerate bands remain visible.

KS is a weighted empirical CDF supremum. Wasserstein integrates the absolute CDF difference. MMD is the squared weighted empirical RBF embedding distance including diagonals, not the unbiased U-statistic. Missing values are complete-case exclusions for numeric tests and distinct PSI bins. PSI uses smoothed fixed-bin shares and separate unknown-category bins; it has no universal significance threshold.

Numeric permutation p-values equal (1 + null exceedance count)/(B+1), using a relative 1e-14 comparison tolerance. The reported Monte Carlo standard error is approximate and requires independent uniform valid permutations. A 95 percent exact binomial interval for the underlying permutation-tail probability remains nondegenerate at sample boundaries. Hand-selected paths check arithmetic only; exhaustive enumeration inference is not claimed. Cluster paths must move complete equal-size clusters, preserving within-cluster order; weights travel with observations. Marginal multiplicity correction is explicitly chosen. Input drift does not establish concept drift.

CUSUM initializes two sums at zero and updates max(0, up+residual-allowance) and max(0, down-residual-allowance); statistic is their maximum. EWMA initializes at zero, updates smoothing*residual+(1-smoothing)*previous, and uses the absolute value. Threshold is ascending null-path-maximum order statistic ceiling((B+1)*(1-alpha)). Insufficient simulation count is rejected. Crossing is strictly greater than threshold. Exchangeability of calibration maxima and a future whole null path yields an unconditional finite-horizon rank bound; it does not guarantee performance under a misspecified or estimated null. Calibration exceedance frequency and optional independent validation are distinct. Horizon extension requires recalibration.

Challenger differences are challenger minus champion on jointly eligible matching cases. Missing counterparts and eligibility disagreements are separate outputs. Group selection rates use all known decisions, while outcome errors, TPR/FPR and calibration use mature observed cases. Fairness gaps use shared cluster paths; intervals are pointwise. Selective outcomes, small groups and boundary samples limit interpretation. No automatic deployment or fairness certification is implied.

Value

A CalcResult with the cohort bundle, monitoring panel, feature findings, sequential path and calibrated threshold, paired metric differences, or group slices, plus diagnostics and explicit limitations.

References

Gretton et al. (2012), A Kernel Two-Sample Test, Section 2.2, equation (5): https://www.jmlr.org/papers/volume13/gretton12a/gretton12a.pdf. NIST CUSUM and EWMA recurrences: https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc323.htm and https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc324.htm. Kleinberg, Mullainathan and Raghavan (2017), Inherent Trade-Offs in the Fair Determination of Risk Scores: https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.ITCS.2017.43.

Default curve inversion, annual hazards and rating migration

Description

Separates conditional, cumulative and marginal default probabilities, and discrete migration from continuous-time intensity models.

Usage

conditional_pd_from_cumulative(cumulative_pd, context = NULL)
hazard_to_pd(hazard, interval_lengths, context = NULL)
estimate_transition_matrix(ratings, observation_windows, states,
    default_state,
  estimator, knowledge_date, observation_scheme, exits = NULL,
  cohort_period_days = NULL, days_per_year = 365.25,
  empty_row_policy = "error", context = NULL)
project_transition_matrix(matrix, horizon, matrix_kind, states,
    default_state,
  step_length = NULL, time_unit = "years", context = NULL)

Arguments

Details

Conditional periods after complete default have NULL values and an undefined mask. Annual hazards integrate over positive year fractions using stable exponential arithmetic.

Cohort estimates divide complete-case endpoint counts by origin counts. Censored records are excluded unless default was already observed; informative censoring can bias this estimator. Duration estimates divide all observed off-diagonal event counts by originating exposure time and set diagonal intensities to negative row totals. This is a homogeneous Markov maximum-likelihood estimator only when all actual state-change times are observed. Periodic panel observations require a different likelihood. State at window start is the last known rating at or before start; later events occur on (start,end]. Future-known records are ignored; duplicate rating dates require upstream resolution. Default is absorbing and observed cures are rejected. Unknown rows error unless no migration is explicitly assumed and reported.

Discrete projection uses an integer number of periods and repeated matrix multiplication. Generator projection uses native Matrix matrix exponential. Inputs must have valid stochastic or generator rows and exactly absorbing default. Projection assumes time-homogeneous Markov dynamics; no generator is inferred from a transition matrix.

Value

CalcResult. Curve inversion returns conditional_pd and undefined_after_full_default. Hazard conversion returns conditional_pd, cumulative_pd, survival and marginal_pd. Estimation returns matrix, matrix_kind, counts and exposure_years, with exclusions and unidentified rows in diagnostics. Projection returns matrix, pd_by_origin and survival_by_origin.

References

Jackson (2011), Multi-State Models for Panel Data: The msm Package for R, doi:10.18637/jss.v038.i08. See also Matrix::expm.

Explicit Gaussian factor PIT and TTC transformation

Description

Conditions or unconditions a latent Gaussian default threshold under an explicit factor model. Scalar inputs expand to equal-length vectors.

Usage

pit_ttc_transform(pd, systematic_factor, asset_correlation,
    direction,
  factor_distribution, horizon, time_unit, probability_measure,
    context = NULL)

Arguments

Details

PIT equals Phi((Phi-inverse(TTC)-sqrt(rho)*z)/sqrt(1-rho)). TTC integrates over the factor distribution; a zero factor does not generally yield TTC. Correlation and factor suitability are user assumptions. Boundary saturation can lose inverse information and is reported.

Value

CalcResult containing probabilities and factor, horizon and conditioning diagnostics.

References

BIS (2005), An Explanatory Note on the Basel II IRB Risk Weight Functions, https://www.bis.org/publications/explanatory-note-basel-ii-irb-risk-weight-functions.

Shared Default Groups, Portfolio Migration and Paired Dependence Stress

Description

Native portfolio preparation, factor or supplied-copula default paths, once-only terminal guarantee protection, physical rating migrations and paired stress comparisons. Defaults are shared across all positions in a declared default-equivalence group.

Usage

prepare_credit_portfolio(positions, entity_graph,
    default_link_policy,
  portfolio_contract, guarantee_allocations = NULL,
  netting_contracts = NULL, context = NULL)
simulate_credit_portfolio(prepared_portfolio, group_pds,
    dependence_contract,
  shared_innovations, conditional_loss_paths, levels = c(.95, .99),
  context = NULL)
simulate_portfolio_migrations(initial_grades, transition_matrices,
  shared_uniforms, migration_contract, valuation_contract = NULL,
  context = NULL)
stress_portfolio_factor_structure(prepared_portfolio, group_pds,
  dependence_contract, shared_innovations, conditional_loss_paths,
  stress_cases, levels = c(.95, .99), context = NULL)

Arguments

Details

The prepared artifact binds facility lots, default groups and reservations with a content hash. Complete legal evidence, facility totals and default-link policy are supplied by the caller; they are not legally authenticated. Collateral recovery belongs in the conditional LGD path and is not reapplied.

For Gaussian factors, common normals are transformed by the declared PSD correlation matrix; systematic loadings plus independent residual normals produce unit-variance latent assets. Asset correlation is distinct from default-indicator correlation. Supplied dependent uniforms can instead come from the granular copula methods. Each group produces one default event that applies to all its facilities. EAD and LGD paths are bound to the same scenario/driver contract, preserving supplied dependence with recovery and exposure.

Gross loss is default indicator times EAD times raw LGD. Guarantees pay only positive remaining loss when the borrower defaults and the guarantor survives the terminal horizon. First-loss protection pays up to its reserved monetary cap. Pro-rata protection first scales positive gross loss by the covered-EAD fraction capped at one, then applies monetary and remaining-loss caps. These are explicit terminal-horizon payoff assumptions, not a regulatory capital substitution formula or a model of interim guarantee settlement.

Migration paths retain the initial grade, every period grade and first-default time. First-default time remains fixed after an explicitly allowed cure. The output separately reports weighted empirical terminal frequencies and products of the declared transition matrices. Arbitrary temporal dependence of supplied uniforms does not establish the Markov law. Optional marked loss equals supplied initial value minus supplied terminal mark minus cash distributions; signed valuation gains remain present.

Dependence stress preserves random coordinates and labels whether conditional recovery/exposure paths are fixed or replaced. Monetary comparisons are paired finite differences. Identical marginal PD parameters need not imply identical finite-sample default frequencies after a factor-structure change. No hidden PD uplift, covariance repair, collateral recovery, additional netting or risk-neutral conversion is performed.

Value

CalcResult containing a sealed prepared portfolio, joint default/monetary loss paths with weighted risk distribution, grade and first-default paths with optional marked values, or baseline/stress results with paired changes.

See Also

assign_default_groups, allocate_guarantee_coverage, credit_loss_distribution, incremental_credit_risk

Ordered Cure and Subsequent Re-Default Projections

Description

Reuse a frozen Aalen–Johansen model to distinguish reversible state occupancy, sustained cure and first default after cure.

Usage

predict_cure_and_redefault(current_episode, model,
    action_conditions,
  horizon_grid, state_recoveries, target, currency,
  extrapolation = "error", context = NULL)

Arguments

Details

Cure states must already encode the specified probation rule in training data. The augmented projection remembers cure and then marks the first subsequent default permanently. Historical default is never erased. Initial default occupancy does not count as re-default. Model training must be known by the episode cutoff. Action conditioning is observational and does not identify a causal treatment effect. Unsupported flat projections do not change model parameters.

Value

A CalcResult with state occupancy, sustained-cure probability, ordered cure-then-redefault probability, expected undiscounted recovery and support diagnostics.

Consolidate Stored Model and Policy Reasons

Description

Deterministic same-scale contribution grouping with separately reproduced sealed policy evidence.

Usage

rank_credit_reason_codes(contributions, reference_basis,
    reason_dictionary,
  policy_decision = NULL, max_reasons = NULL, context = NULL)

Arguments

Details

Signed transformed-feature contributions are summed within a reason before testing adverse direction. Non-adverse groups are suppressed with backing evidence retained. Model reasons rank by descending adverse net magnitude, then priority and code. Policy reasons rank separately by priority and code, retaining matched rule traces and selected status. Their units are never added to model effects. Unmapped contributions and absolute contribution coverage are explicit. Content hashes detect alteration but do not authenticate evidence. Reasons are descriptive, not causal effects or automatically legally sufficient adverse-action notices.

Value

CalcResult with ranked model reasons, separate policy reasons, suppressed/omitted groups, unmapped contributions, coverage and provenance hashes.

Dated local ratings, immutable model versions and historical replay

Description

Build typed rating subjects, route profiles, assess evidence, select explicit historical model versions, preserve risk quantities and trace local lifecycle events. Batch inference and replay use frozen declarative recipes.

Usage

build_rating_subjects(tables, links, as_of, known_at)
route_rating_profile(snapshot, rules, as_of, known_at)
assess_rating_applicability(snapshot, contract, as_of, known_at)
register_rating_model_version(registry, bundle, registration)
select_effective_model(registry, assignments, profile_id, purpose,
    as_of, known_at)
compose_risk_assessment(assessment_id, subject, components, binding,
    as_of,
                        known_at, calculated_at, effects = NULL)
check_assessment_freshness(assessment, policy, events, as_of,
    known_at)
transition_assessment_state(assessment, event, transitions)
rate_subject_batch(requests, registry, batch_id, known_at)
apply_assessment_event(state, event, dependencies, known_at)
replay_assessment_history(requests, registry, expected, known_at,
                          replay_id = "replay")

Arguments

Details

All schemas use version 1.0. Dates have calendar-day resolution. Effective intervals are [effective_from,effective_to); NULL end means unbounded. Knowledge cutoff is inclusive. Use named lists for objects, unnamed lists for arrays including singletons, and structure(list(), names=character()) for empty objects. Row tables also accept data.frame. Hashes detect changed content; they do not authenticate an author.

Subject types remain distinct. An overlapping effective version creates an explicit conflict and is excluded from the resolved graph. Relationships are legal_obligor, joint_liability, economic_dependency, default_correlation, data_identity, contains, pd_reference. Unresolved link endpoints are reported. No PD, grade or loss is copied between subject levels.

Routing requires matching snapshot economic date and snapshot knowledge no later than the query. Attribute observation cannot follow its knowledge date. Future observations and facts learned after snapshot cutoff are unavailable. Conditions combine with AND. Operators are eq, ne, lt, le, gt, ge, in, not_in, present, absent. Missing values fail all comparisons except absent; ordered comparisons are finite numeric. Membership values are arrays. Empty conditions explicitly match all. Smallest matching priority wins. Equal best priority yields AMBIGUOUS, no match NO_ROUTE, unique match ROUTED with complete trace. No expression evaluation occurs.

Applicability distinguishes missing, unavailable, stale, conflicting and outside training support facts. These yield REFER. A known disallowed applicability field yields NOT_APPLICABLE; otherwise APPLICABLE. Support bounds are inclusive marginal ranges, ages are nonnegative integer days, and required feature names are distinct. These statuses are neither default labels nor lending decisions.

Registry versions are immutable content-bound entries, sorted by model ID/version. Model training knowledge must precede registration. Identical bundle and registration are idempotent; a differing bundle or registration under the same model ID/version fails. Selection requires one effective known assignment and the exact pinned bundle hash registered by historical cutoff. It reports SELECTED, MISSING_MODEL, AMBIGUOUS_MODEL, NOT_YET_EFFECTIVE or EXPIRED_MODEL. Later registered models do not replace an explicit historical version. Dependency/configuration maps participate in the bundle hash and are supplied by the caller.

Composition retains PD, GRADE, LGD, EAD, EL and SCORE as separate quantities and variants. Every definition includes event, positive horizon, time_unit (days/months/years), probability_measure (PHYSICAL/RISK_NEUTRAL), conditioning. PD lies in [0,1]; EAD is nonnegative. LGD and EL are finite and may reflect net gains. Grades name their scale and graded quantity (PD, EL or SCORE). Definition differences within subject/variant, duplicate quantities, unlinked other subjects and mixed currencies are explicit composition conflicts; no automatic inference or aggregation occurs. Effects have kind support/transfer/override/cashflow and unique integer order. Input/output component references must exist, output level must match, and producer order must precede consumer order. Repeated output writes and cycles fail. Caller must identify the same economic effect consistently to avoid hidden double counting.

Composed assessments start DRAFT with history. Freshness checks age, information age, effective expiry and known new material events, including late corrections of older facts. Equality to an age limit remains current. Status is CURRENT, STALE or SUPERSEDED. Historical queries cannot use assessment lifecycle state from a later knowledge date.

Lifecycle state names are DRAFT, SUBMITTED, REVIEWED, FINAL, SUPERSEDED, with explicit caller-supplied transitions. Exactly one transition must match, all required gates must be TRUE, reasons nonempty, and dates/history ordered. Duplicate events fail. Conflicted assessments cannot finalize. FINAL only permits a declared move to SUPERSEDED, which is terminal. Append-only history includes previous assessment hash, old/new state and full event; earlier input objects remain unchanged. FINAL is local processing status and conveys no institutional approval.

Batch requests pin model identity and bundle hash. A recipe has method, version 1.0 and parameters. The application_risk method accepts input=list(snapshot=...) and parameters target plus optional calibrator/offset. Its snapshot cutoff must not exceed request knowledge. The binary_glm method accepts input X (row matrix or row lists) and feature_order; parameters expected_target plus optional offset. This finite dispatcher executes supported library inference only; no user code, callbacks or executable serialization. Model registration must be available by each individual request cutoff, and request knowledge must be available by the batch cutoff. Each unique string row_id retains success or a controlled error and missing value; sorted row order is input-permutation invariant. Hashes bind input, recipe, model, numeric result, sorted request collection and registry.

Event projection matches explicit dependency type/ID pairs and returns sorted unique affected assessment IDs without recalculating or mutating risk/accounts. Same ID/revision/content is DUPLICATE with unchanged state; differing content under that key fails. Corrections have distinct revisions referencing the latest processed revision and retain original history. Future events are DEFERRED; state containing future processed events is rejected for a historical cutoff.

Replay executes supplied pinned historical requests and compares input/recipe/model/ result hashes. Missing bundles or references and changed hashes yield NOT_REPRODUCIBLE with findings; complete agreement yields REPRODUCED. Numerical hashes are exact, so numerical runtime differences remain visible. No current champion or external source is loaded. Reproducibility is not empirical validation.

Value

CalcResult with named-list value. Objects and row arrays are data only. Registry, assessment, batch and event state have canonical SHA-256 content hashes. Outputs are detached R values; consumers reject mutated hashes.

Content-bound assessment pipelines

The unchanged batch signature additionally accepts recipe method assessment_pipeline, version 2.0. Its identical data-only recipe must be stored in the registered bundle's configuration$assessment_recipe. Dependencies are ID-keyed artifacts, each with kind, content, canonical content sha256 and known_at; kind innovations identifies supplied random inputs. Empirical fitting consumes frozen method outputs; numerical market pricing and explicitly driven simulations execute their native granular methods.

Input contains snapshot, ID-keyed datasets (value/as_of/known_at), assessment_id, assessment_known_at, calculated_at, lifecycle_events and dependency events. Snapshot/data/model/artifact information must be available at assessment_known_at. Dataset economic dates cannot follow the subject snapshot. Calculation lies between assessment knowledge and the historical request cutoff; lifecycle events cannot exceed that request cutoff. Later review history never changes the original input information set.

Recipe parameters explicitly contain routing_rules, applicability_contract, steps, components, effects, transitions, event_dependencies and nullable decision. Exactly one applicable route is required before execution. Steps contain step_id, method_id, arguments, argument_types and output_contract. References have the exact object form list('$ref'=list(source="step", id="pd", path=list("value",0))). Sources input/artifact/model/step resolve dataset envelopes, frozen content, primary model under ID primary, or predecessor outputs. Both languages use zero-based array indices. Ready DAG steps run in sorted ID order; unknown predecessors and cycles fail. No user code or user-selected imports are evaluated.

Argument types explicitly select numeric_matrix, numeric_vector, character_vector or logical_vector conversion. Other arguments retain their typed list structure. Output contracts require shape (scalar/vector/matrix/object/records), unit and nullable target. Optional fields contain path, unit, target and nullable target_path; the latter binds an actual nested target, including scenario sampling law. Component references must match producing quantity/currency and full target. Physical and risk-neutral measures and PIT/TTC sampling laws are never inferred or interchanged. A PRE_FITTED empirical step requires exactly one artifact argument referencing a frozen output with the same method_id. Registry, lifecycle and recursive batch operations use dedicated stages, not arbitrary DAG calls.

The Shapley operation accepts a data-only predictor with kind FROZEN_PD, model, feature_order and expected_target. A trusted internal generic-PD call is constructed; model identity and probability scale must match its explanation contract.

Composition is followed by declared lifecycle transitions and revisioned dependency event projection. Repeated identical lifecycle events are idempotent; different content under the same event identity fails. A dependency correction retains prior history. Projection does not silently change risk; recomputation uses a new dated request. Optional decision arguments invoke the frozen policy separately, with PD and target equal to a composed PD component.

Version-2 rows include assessment_hash, event_state_hash, artifact_manifest_hash, innovation_hash (manifest entries tagged innovations) and nullable decision_hash, in addition to legacy input/recipe/model/result hashes. Expected replay rows must contain these fields. Full calculation, composition and ordered history are rerun; corrupt/missing artifacts produce NOT_REPRODUCIBLE findings. Successful values retain routing/applicability, ordered step argument/value hashes, final assessment, artifact manifest, event state/trace and separate decision. Numerical hashes are exact; cross-runtime last-bit differences remain visible. Hashes establish content integrity, not signature authenticity, empirical validity or institutional approval.

Examples

empty <- structure(list(), names = character())
s <- list(subject_id = "O", subject_type = "obligor", as_of = "2024-01-01",
          known_at = "2024-01-01", attributes = empty)
r <- list(list(rule_id = "R", profile_id = "P", profile_version = "1",
               priority = 1, conditions = list(), effective_from = "2024-01-01",
               effective_to = NULL, known_at = "2023-12-01"))
route_rating_profile(s, r, "2024-01-01", "2024-01-01")$value$status

Reconcile Dated Card Authorizations and Settlements

Description

Replaces settled reservations by booked money, preserving event identity and separate knowledge time. No total account exposure is inferred.

Usage

reconcile_card_authorizations(authorizations, events, policy, as_of,
  known_at, currency, context = NULL)

Arguments

Details

A partial capture of 80 from a reservation of 200 leaves booked exposure 80 and pending authorization 120. Complete capture releases the unused reservation. Reversal reduces booked exposure without reopening the hold. Expiry is exclusive; late settlement needs explicit policy permission. Cancellation closes a hold without refunding a purchase. Overcapture and reversal above the original remaining settlement are rejected.

Amounts are reconciled in integer minor units, with an explicit maximum of 2⁠^⁠40 units for each amount and account aggregate. Binary representation error tolerance in units is max(1e-9, min(1e-3, 8*2⁠^⁠(-52)*abs(amount/minor_unit))). This is not a business rounding rule. The scenario conversion factor is supplied, not an estimated CCF, and can produce fractional minor-unit scenario exposures. Only supplied reservations and their booked charges are covered; account payments and installment conversions require a separate cash ledger.

Value

A CalcResult with authorization and account results, ordered event ledger, excluded records, explicit policy and input/output fingerprints.

Historical Retail Payment and Revolving Features

Description

Replay point-in-time payment ledgers and separate balance movement, limit changes and minimum/full repayment behavior.

Usage

build_payment_history_features(obligations, payments, coverage,
    account_id,
  as_of, information_cutoff, lookback_days, allocation_policy,
  currency, grace_days, context = NULL)
build_revolving_features(balance_history, statements, coverage,
    account_id,
  as_of, information_cutoff, lookback_days, currency,
    context = NULL)

Arguments

Details

Payment histories replay each day using the earlier of the feature cutoff and that historical date. Later-known reversals do not rewrite earlier days. Supply the complete contractual ledger, not only window-dated records. Grace changes only the assessment date, never the actual due date or DPD. Each obligation due inside the window is assessed at due date plus grace. Unelapsed grace and obligations not contemporaneously known are reported as unassessable. Payment ratios aggregate allocated amounts over assessed amounts; a zero denominator is NULL. CURRENT and ARREARS describe unpaid-ledger states and are not regulatory default classifications. Partial/unknown history leaves complete features NULL while retaining known-history summaries.

Revolving snapshots select latest-known revisions and require one resolved record per observed day. Missing daily balances are not carried forward. Complete overlimit-day totals require all calendar dates and complete coverage. A zero limit gives NULL utilization. Explicit inactive accounts differ from undrawn active accounts. Closed accounts require zero residual balance and limit. Utilization change equals balance change divided by the new limit plus previous utilization times the relative limit cut using the new limit denominator. This separates net balance movement from limit effects, not gross draw transactions. Trend is endpoint utilization change per elapsed calendar day.

Statement paid_by_due must mean allocated payments by contractual due date, not eventual receipts. Observation must extend through due date. Full, minimum without full, below minimum and no-amount-due states remain separate. Statement counts refer to observed supplied evidence rather than asserting unseen history exists.

Value

CalcResult containing daily history, assessments and descriptive summaries with explicit missing coverage. Diagnostics retain cutoff policy, coverage and excluded revisions.

See Also

compute_days_past_due, build_bureau_features

Booked Cashflow and Retail Liquidity Features

Description

Compute locally classified dated cashflows, own-transfer reconciliation, uncertainty, volatility and observed liquidity buffers.

Usage

build_transaction_features(transactions, daily_balances,
    category_mapping,
  coverage, customer_id, own_account_ids, as_of, information_cutoff,
  lookback_days, currency, context = NULL)

Arguments

Details

Latest knowledge-visible releases are selected before the posting-date window. Value dates remain visible. Reversals must reference earlier visible same-account originals of opposite sign and cannot cumulatively exceed them. They inherit the original classification while retaining their own cash date. Supply pre-window originals needed for lineage. Cross-window reversals remain actual cash movements.

Categories are SALARY, OTHER_INCOME, BUSINESS_REVENUE, CONSUMPTION, DEBT_SERVICE, BORROWING, REFUND, OWN_TRANSFER, OWNER_DRAW, TAX and UNKNOWN. Unmapped codes remain UNKNOWN. Only salary and other income enter classified income; consumption, debt service and tax enter classified expense. These are classified subtotals, not claims about the economic nature of unknown amounts. Borrowing, business sales, refunds and owner draws are not household income.

Own-transfer legs must identify two explicitly owned accounts and opposite same-currency amounts. Single visible legs remain unmatched transfers, not income. Own transfers and owner draws are excluded from external cashflow statistics. Daily external net cashflow includes zero posting days; its standard deviation uses the population denominator and numerical scaling. Incomplete history leaves complete statistics NULL while retaining known-record statistics.

Balance snapshots require one resolved observation per account-day. Negative accounts count separately as overdraft account-days. Household buffer observations require all declared accounts that day. Missing balance days are never carried forward or interpreted as zero overdrafts. Full overdraft totals require complete source coverage and every declared account-day.

Value

CalcResult with category credit/debit/net totals, known and coverage-qualified cashflows, daily volatility, category uncertainty, observed liquidity, classified transaction lineage and unmatched own transfers.

See Also

build_payment_history_features, reconcile_income_evidence

Dated Bureau, Thin-File and Income Evidence

Description

Describe local credit files, evidence availability and reconciled income without inferring zero debt from missing sources or summing the same salary across sources.

Usage

build_bureau_features(snapshots, tradelines, inquiries,
    provider_mapping,
  customer_id, provider_id, as_of, information_cutoff,
    lookback_days,
  max_age_days, currency, context = NULL)
build_thin_file_features(bureau_features, evidence,
    eligibility_profile,
  customer_id, as_of, information_cutoff, context = NULL)
reconcile_income_evidence(declarations, documentary_evidence,
  transaction_evidence, inclusion_policy, customer_id, as_of,
  information_cutoff, currency, period, basis, context = NULL)

Arguments

Details

Latest knowledge-visible revisions are selected before economic filters. Bureau tradelines bind both snapshot_id and snapshot_release_seq; late corrections cannot rewrite earlier evidence. Inquiry counts use the interval (as_of-lookback_days, as_of] and supplied local history. Closed balances remain separate from open balances. Missing values and partial coverage prevent complete open totals; zero limits give NULL utilization. Absent and no-hit files never establish zero household debt. Maximum reported DPD describes the supplied tradeline history rather than an invented lookback maximum. No provider score is reproduced.

Thin-file strata describe evidence, not default classes. Each feature requires all named evidence kinds with acceptable age and coverage. Supplied thresholds and partial-evidence policy are explicit. Common direct outcome labels cannot be designated input features. Use lists for JSON-style arrays, including single-element arrays.

Income types are SALARY, BONUS, PENSION, RENT_NET, NET_BUSINESS_INCOME and OTHER_NET_INCOME. Each stable component is selected once from its newest eligible source evidence according to the declared priority. Salary and bonus need distinct identities. Business revenue, borrowing, owner transfers and refunds are not inferred as disposable income. Gross/net, period and currency mismatches are excluded without guessed conversion. Document support is supplied evidence, not independent authenticity verification. Source values and discrepancies from the declaration remain visible; a zero declaration has no relative denominator. Incomplete coverage or missing expected components leave total income NULL while retaining the known subtotal.

Value

CalcResult with bureau file state, coverage, age and features; evidence strata and supported/unsupported features; or reconciled components, selected subtotals, nullable total and discrepancies. Diagnostics retain source revisions, policy or mapping fingerprints and excluded evidence.

See Also

validate_retail_dataset, aggregate_household_obligations

Point-in-Time Retail Records and Household Obligations

Description

Validate explicit local retail schemas, freeze typed dated relationships, and separate unique household obligations, budget allocations and person-level legal liability.

Usage

validate_retail_dataset(tables, data_contract, as_of,
  information_cutoff, context = NULL)

build_customer_household_graph(entities, relationships, as_of,
  information_cutoff, context = NULL)

aggregate_household_obligations(obligations, household_graph,
  coverage, currency, period, guarantee_budget_policy,
    context = NULL)

Arguments

Details

Each table specification has table, role, primary_key, fields, foreign_keys, date_order, event_field, known_field, feature_fields and allow_extra_fields. Roles are IDENTITY, FEATURE, OUTCOME or DECISION. Fields declare name, type (string/date/number/integer/boolean) and boolean nullable. Foreign keys declare field, table and target_field. Date-order pairs declare earlier and later fields. Keys and features are explicit arrays; use lists for JSON-style arrays, including single-element arrays. Event and known fields must have date types. Source coverage rows declare source_id, status and known_at. Future coverage assessments appear UNKNOWN. Eligible children must reference cutoff-eligible parents. Repeated applications remain separate when application IDs differ. Violations use zero-based row indices in both languages.

Only FEATURE tables can designate model inputs. User-prohibited fields and common direct outcome names cannot be designated features. This contract does not infer the meaning of arbitrary derived variables or establish their lending suitability. Self-referencing payment foreign keys can preserve reversal identity; economic allocation remains separate.

Relationships run from customers to households (MEMBER_OF), applications (COAPPLICANT_ON), facilities (BORROWER_ON or GUARANTOR_ON), or businesses (OWNER_OF). The latest cutoff-known release ordered by known_at and release_seq is selected before filtering its half-open effective interval. Identity is never inferred from address or name. Duplicate active customer/target/role edges and multiple active economic households per customer are rejected.

Debt fractions belong only to borrower/guarantor roles and lie in [0,1] or are NULL. Internal borrower budget shares cannot sum above one per facility. Legal fractions can each equal one for joint borrowers. Household economic membership is not an automatic legal default unit, guarantee or owner-PD transfer.

Aggregation selects the latest known obligation snapshot per facility and counts each linked facility once per household. Legal liability and internal budget allocation remain distinct. Closed facilities require explicit zero residual amounts. Missing snapshots, amounts or statuses and partial/unknown coverage leave totals NULL even if known subtotals equal zero. Missing budget shares separately prevent complete budget totals. Snapshot age, release and source remain visible. Content bindings prevent silent modification of a supplied graph.

Value

A CalcResult. Validation returns valid, applicability, violations, row counts, eligible row indices and source coverage. Graph construction returns a dated, content-bound typed graph, memberships and liability paths. Aggregation returns household known subtotals and nullable totals, budget allocations, individual legal liabilities and source coverage.

See Also

compute_days_past_due, join_as_of

Examples

entities <- list(
  list(entity_id = "person", entity_type = "CUSTOMER", known_at = "2024-01-01"),
  list(entity_id = "house", entity_type = "HOUSEHOLD", known_at = "2024-01-01"))
relationships <- list(list(relationship_id = "membership", release_seq = 1,
  source_id = "person", target_id = "house", role = "MEMBER_OF",
  effective_from = "2024-01-01", effective_until = NULL,
  known_at = "2024-01-01", liability_fraction = NULL, budget_share = NULL,
  evidence_reference = "verified household membership"))
graph <- build_customer_household_graph(entities, relationships,
  "2024-01-31", "2024-01-31")
graph$value$memberships

Dated BNPL Purchase Plans and Allocated Obligations

Description

Aggregate known purchase plans, payment/refund allocations, separate disputes and explicit provider coverage.

Usage

build_bnpl_obligation_features(plans, installments, events,
  provider_coverage, customer_id, as_of, information_cutoff,
  lookback_days, near_term_days, currency, context = NULL)

Arguments

Details

Latest knowledge-visible revisions are used. Financial events PAYMENT, MERCHANT_REFUND, PAYMENT_REVERSAL and REFUND_REVERSAL require explicit same-plan installment allocation. No allocation rule is guessed. Reversals identify earlier original events and cannot cumulatively exceed them. Net paid plus refunded amounts cannot exceed the contractual installment; excess customer cash refunds are separate cashflows. Financial events cannot follow actual plan closure. Closed plans require zero residual known unpaid installments.

DISPUTE_OPEN and DISPUTE_CLOSE have zero financial amount. Closing identifies one earlier opening on the same plan. Disputes do not themselves cancel debt or imply default. Explicit refinancing must reference an earlier visible plan; repeated purchases alone are not refinancing.

Concurrency uses actual half-open purchase/closure intervals visible at the feature cutoff, not assumed contractual maturity. It is a retrospective known history rather than a contemporaneous-knowledge replay. Active contracts and positive-outstanding plans are separate counts. Near-term cashflows use (as_of, as_of+near_term_days]; overdue and due-today amounts are separate. Incomplete provider coverage, incomplete schedules or a plan without installments leave total outstanding NULL. Known subtotals remain available.

Value

CalcResult with plan and provider concurrency, repeat use, explicit refinancing, outstanding installments, near-term due cashflows, separate disputes and source-release diagnostics.

See Also

build_transaction_features, aggregate_household_obligations

Validation of Frozen Retail Risk Pools

Description

Replay dated pool assignments and separately assess mature PD, closed-workout LGD, observed default EAD, concentration, residual heterogeneity and migration.

Usage

retail_pool_diagnostics(fitted_model, pool_assignments,
    assignment_inputs,
  predictions, cohort_outcomes, workouts, links, estimation_profile,
  design, previous_assignments = NULL, uncertainty = "none",
  cluster_resamples = NULL, confidence_level = .95,
    tolerance = 1e-10,
  context = NULL)

Arguments

Details

No pool refit is performed. One forecast per declared default unit is required. Whole horizons mature before PD metrics; missing/open workouts do not invent final LGD. Constant pool PD has no identifiable within-pool rank. Relationships never propagate default labels. Observed diagnostic summaries do not replace frozen forecasts. Independent-unit binomial and episode intervals are omitted; supplied whole-cluster draws provide conditional uncertainty. A residual range or sparse sample does not establish pool homogeneity.

Value

CalcResult containing ValidationResult with separate parameter metrics and applicability, residual diagnostics, concentration, migrations, exclusions and cluster intervals.

References

BCBS CRE36, published 2022-12-08, paragraphs 36.16–17 and 36.24. https://www.bis.org/committees/bcbs/basel-framework/standard/cre/36/inforce/2023-01-01/published/2022-12-08

Frozen Retail Pools and Local Risk Parameters

Description

Fit deterministic constrained adjacent score pools, replay frozen assignments and estimate separately weighted local PD, LGD, EAD and CCF with explicit maturity and default conditioning.

Usage

fit_retail_risk_pools(observations, pooling_spec, target, training,
  model_id, model_version, context = NULL)
assign_retail_risk_pool(observations, fitted_model, as_of,
  information_cutoff, expected_target, uncovered_policy,
    context = NULL)
estimate_pool_risk_parameters(pool_assignments, cohort_outcomes,
  workouts, estimation_profile, evaluation_cutoff, calibration_id,
  calibration_version, context = NULL)

Arguments

Details

Strata explicitly include product, borrower_type, facility_type and delinquency_state. Each segment declares segment_id, exact values for all strata, finite score_min/score_max, interior increasing cutpoints and separate pd/lgd/ead monotonicity directions none/increasing/decreasing. Score and strata must not be direct outcome fields. Spread thresholds separately declare nullable pd/lgd/ead maximum ranges of original nonempty-bin means; NULL disables a restriction. Zero support minima explicitly permit absent conditional-parameter evidence. Use lists for JSON-style arrays, including single-element and empty arrays.

Training features must be known by index. The whole horizon and label receipt must precede training cutoff. Default dates are in (index, horizon end]. Nondefault rows cannot carry LGD/EAD. Only CLOSED workout LGD enters as a numeric LGD label; OPEN final LGD is NULL. Conditional outcome knowledge dates lie between actual default and training cutoff. Missing numeric outcomes have NULL knowledge dates.

Initial bins are left closed/right open with the final endpoint included. Greedy adjacent merging prioritizes undersupported blocks, choosing admissible neighbors by the sum of squared normalized available parameter-mean differences, then the leftmost index on ties. With no support merge available, it considers adjacent monotonicity violations. Every merge must satisfy enabled original-bin mean-spread limits. The complete merge path and separate PD/LGD/EAD diagnostics are retained. This is a local heuristic, not a proof of global optimality or infeasibility. Spread constraints describe between-bin means, not statistical proof of homogeneity. Any remaining violation leaves the affected whole segment INSUFFICIENT_SUPPORT and unassignable. No global monotone LTV law is assumed.

Assignment never refits the artifact. Missing features, unseen strata, out-of-domain scores and unsupported segments remain uncovered. IDs bind model/version, ruleset, segment and merged bin interval. Artifact integrity, target identity, model availability and feature chronology are checked. A dated delinquency change can legitimately change the pool.

Parameter estimation uses mature whole-horizon unit outcomes with exposure weights frozen at index. Censored, unknown and unassigned outcomes are separate exclusions, not nondefaults. A default unit appears once. Workouts must match cohort default and reference date. discount_recoveries values explicit cashflows at the supplied workout valuation knowledge date; workout_lgd and realized_ccf preserve raw economic ratios, open-workout incompleteness and undefined zero denominators. One workout per observation is allowed.

Table status describes mature-cohort support only; separate parameter statuses identify defined and absent/zero-weight conditional samples. Weighted estimates retain numerators, denominators, positive-weight counts, observed ranges and independent-episode weighted-mean standard errors; fewer than two contributing episodes leaves the error NULL. Open workouts, missing default workouts, zero-EAD closed workouts and undefined CCF remain separate. Sparse pools retain descriptive estimates with insufficient-support status.

calibrate_grade_pd without priors or monotonic recalibration supplies unweighted exact binomial intervals; low_default_bound supplies a one-sided independent binomial bound. These refer to unit counts, not exposure-weighted PD. Inference is conditional on chosen pools, without pool-selection uncertainty. Local cohort PD is not automatically TTC or long-run PD. Missingness, censoring, dependence and representativeness require separate assessment; these routines do not establish regulatory approval.

Value

FitResult with bound pooling ModelArtifact; CalcResult with bound RETAIL_POOL_ASSIGNMENTS; or CalcResult with a bound CalibrationArtifact of type RETAIL_POOL_RISK_PARAMETERS and pool risk table. The latter is not a score transform for apply_calibrator.

See Also

serialize_fitted_model, calibrate_grade_pd, low_default_bound, discount_recoveries, workout_lgd, realized_ccf

Explicit Small-Business and Household Cash Interactions

Description

Separate business operating cash, matched owner transfers, household debt and typed legal liability evidence.

Usage

build_retail_sme_features(business_cashflows, owner_drawings,
  household_obligations, liability_links, coverage, business_id,
  owner_customer_ids, business_debt_facility_ids,
  household_debt_service_window, as_of, information_cutoff,
  lookback_days, currency, context = NULL)

Arguments

Details

Business categories are SALES, OPERATING_COST, TAX, CAPEX, DEBT_SERVICE, FINANCING, OWNER_DRAW and OWNER_CONTRIBUTION. Signed operating cash comprises sales, operating costs and tax; adding capex and then debt service yields cash before and after business debt. Financing and owner transfers are excluded. Sales are not household income. Owner draws are business debits, contributions credits. Other categories can include explicitly classified signed reimbursements.

Owner drawing legs must match a supplied business posting and opposite household amount once. Local household evidence is asserted, not independently obtained. Unmatched draws remain visible without becoming operating income. Owner households are joined by explicit IDs and counted once. Missing household totals or shared business/household debt facility IDs suppress combined debt subtraction. Cash after household debt excludes living costs and is not affordability or distributable income. No currency or debt-service period conversion is guessed.

Legal types are LIMITED_OWNERSHIP, PERSONAL_GUARANTEE and UNLIMITED_LIABILITY. Fractions lie in [0,1] or are NULL; caps are nonnegative or NULL. Half-open dated links remain explicit evidence; ownership never automatically becomes a guarantee or copied owner PD. Customer concentration uses squared net-sales shares only when total sales are positive and all customer amounts nonnegative. Calendar-month sales and peak-month share are observations, not a seasonal forecast. Negative monthly net amounts make the peak-share statistic undefined.

Value

CalcResult with business cash sequence, owner transfers, monthly and customer sales, matched household debt, overlap exclusions and typed legal evidence.

See Also

aggregate_household_obligations, build_transaction_features

Contractual tranche cash accounts and payment priorities

Description

Allocates actual pool cash to senior costs, interest, liquidity, reserves and tranche principal with explicit triggers and terminal identification.

Usage

run_tranche_waterfall(asset_pool_paths, structure_profile,
    context = NULL)

Arguments

Details

Pool interest and principal/recovery enter separate accounts. Pool workout and servicing costs and nonoverlapping additional deal fees rank first. Tranche coupon and prior arrears follow seniority. Reserve and repayable liquidity provide explicit support. Liquidity repayment precedes reserve refill and tranche principal; it can use principal even where direct principal-to-interest diversion is prohibited. Remaining interest is diverted to principal or equity as configured. Outstanding interest arrears block all principal distributions.

OC uses performing principal, pending net modeled recovery and actual cash over protected outstanding debt. IC uses interest after fees over protected coupon and arrears. A breach forces sequential principal and may authorize principal-to-interest diversion. Ratios with zero denominators are null and do not trigger.

REQUIRE_COMPLETE requires full pool runoff, releases final reserve, disallows new terminal liquidity and records ultimate residual principal write-down. KEEP_OPEN returns null ultimate loss. Timely-interest and legal-maturity failures persist after subsequent cure. No interest on arrears or post-maturity penalty claim is invented. Initial reserve is separately funded support.

Value

CalcResult containing a hashed TrancheWaterfallPaths artifact with dated tranche cashflows, account conservation residuals, persistent payment-failure indicators and ending principal loss identification.

Estimated logistic scorecards, odds scales and contribution explanations

Description

Fits a logit scorecard using the native GLM engine with separation, rank and convergence checks, then replays contributions and explicitly interpretable score transformations.

Usage

fit_scorecard(design_matrix, outcomes, feature_order, fit_spec,
    target, training,
  model_id, model_version, weights = NULL, max_iter = 100,
  tolerance = 1e-10, context = NULL)
score_scorecard(fitted_model, transformed_features, feature_order,
  expected_target, context = NULL)
scale_score_points(log_odds, base_score, base_odds,
    points_to_double_odds,
  odds_convention, score_direction, context = NULL)
apply_qualitative_adjustments(qualitative_inputs, coefficient_spec,
  neutral_levels, base_value, adjustment_scale, context = NULL)
scorecard_reason_codes(score_contributions, reference_profile,
    reason_policy,
  context = NULL)

Arguments

Details

Preprocessing and supervised binning must already be frozen within the training fold. Feature-kind declarations do not independently establish absence of leakage. The model contains an intercept and neutral-profile contribution reference. Complete or quasi separation is a controlled error; no proprietary coefficients are supplied.

Points are affine in default log-odds, with slope minus PDO/log(2) for higher-is-safer and plus PDO/log(2) for higher-is-riskier. Base odds may use either direction, which is explicitly converted. Returned inverse parameters reconstruct PD without rounding.

Qualitative effects subtract neutral-level coefficients. Log-odds effects return adjusted PD; integer rating-notch differences do not invent a PD or unknown rating-scale boundaries. Reasons compare model contributions with a declared reference, rank adverse differences, break ties lexically and retain full reconciliation. Explanations do not imply causality.

Value

FitResult for fitting; otherwise CalcResult. Scoring returns PD, log_odds, intercept, feature and reference contributions and differences. Scaling returns points and inverse parameters. Adjustments return contributions, total, adjusted value and optional PD. Reasons return selected reasons, all differences, total difference and unreported difference.

References

Native binomial logit estimation follows stats::glm.fit; independent tests recover grouped binomial maximum-likelihood estimates and verify score inversion and contribution arithmetic.

Selection Mechanisms, Reject Inference and Honest Evidence Scopes

Description

Native dated selection panels, logit propensities, sequential inverse probability weights, hypothetical augmentation, binary bivariate-probit selection likelihood, augmented population means and partial-identification diagnostics. Missing own-contract outcomes never become measured validation labels.

Usage

build_selection_dataset(applications, decisions, bookings, outcomes,
  as_of, target_population, target, context = NULL)
fit_selection_propensity(X, selected, feature_order, mechanism,
    folds,
  row_metadata, target, training, model_id, model_version,
    context = NULL)
selection_ipw_weights(observed, propensities, target_weights = NULL,
  stabilization = NULL, trim = NULL, context = NULL)
fuzzy_reject_augmentation(outcomes, observed, baseline_pd,
    application_ids,
  augmentation_weight = 1, context = NULL)
parcel_rejected_applications(bands, accepted_defaults,
    accepted_counts,
  rejected_counts, log_odds_shift = 0, uniforms = NULL,
    context = NULL)
fit_joint_selection_model(X_outcome, X_selection, selected,
    outcomes,
  outcome_features, selection_features, target, training, model_id,
  model_version, rho_grid = NULL, selected_rho = NULL,
    exclusion = NULL,
  context = NULL)
estimate_population_risk_aipw(outcomes, observed,
    outcome_predictions,
  propensities, fold_lineage, target_weights = NULL,
    clusters = NULL,
  context = NULL)
reject_inference_bounds(outcomes, observed, target_weights = NULL,
  strata = NULL, risk_bounds = NULL, context = NULL)
reject_inference_sensitivity(baseline_pd, observed, outcomes,
  log_odds_shifts, target_weights = NULL, threshold = NULL,
    context = NULL)
prepare_reject_evaluation(candidates, evaluation_rows, as_of,
    target,
  outcome_comparability = NULL, context = NULL)

Arguments

Details

The panel preserves offered-but-not-taken-up, taken-up-but-not-funded, immature funded and censored funded strata. The caller constructs maturity/exit-policy targets before joining; partial default-free histories must not be supplied as mature nondefaults.

Propensity fitting calls the native binary GLM primitive, with independent forward/entity folds. Warmup OOF predictions remain NULL, never in-sample replacements. The final frozen model uses the declared training sample. Deterministic separation fails explicitly.

IPW is observed target mass divided by the product of sequential conditional propensities. Support restriction is reported with retained/excluded mass and Kish effective sample size. Conditional MAR and positivity are assumptions, not conclusions of fitting.

The joint binary likelihood is Phi(-b) for unselected rows, Phi2(a,b;rho) for selected defaults and Phi2(-a,b;-rho) for selected nondefaults. Native deterministic 64-node Plackett Gauss-Legendre quadrature and analytic scores drive bounded L-BFGS optimization, initialized by genuine separate probit fits. No continuous-outcome Heckman correction is substituted. Correlation/coefficient boundaries, tiny likelihood cells at or below 1e-14, separation and failed convergence are not accepted fits. Estimated correlation requires explicit exclusion assumptions; fixed-correlation covariance conditions on the supplied correlation. Observed-information covariance assumes IID observations. Hessian conditioning and weak identification are reported. Generic predict_pd returns unconditional Phi(X beta) for the exact potential-own-contract target and union feature order: outcome columns then selection-only columns.

AIPW averages m+S/e*(Y-m). Cluster influence totals give an empirical sandwich standard error and untruncated normal 95 percent interval. Adequate nuisance convergence and independent clusters are required. Out-of-domain estimates are reported rather than clipped. Double robustness concerns a population mean under MAR, not individual reject PD or general MNAR identification.

Worst-case binary bounds use observed default mass and unobserved mass. Optional restrictions remain assumptions. Log-odds sensitivity reports observed versus assumed components; correlation sensitivity is available through the joint model's fixed-rho profiles. Pseudo outcomes never become evaluation truth. Evaluation-only SplitPlans exclude training-entity or temporal contamination and expose distinct actual, external-proxy and synthetic-engineering scopes. No lending actions or pseudo-label performance rankings occur.

Value

CalcResult with method-specific panel, weights, scenarios, mean/influence uncertainty, partial-identification bounds, sensitivity surface or evaluation-only SplitPlan. Fit functions return FitResult with a common portable ModelArtifact and likelihood/OOF diagnostics. Hypothetical labels are marked validation-ineligible; all-synthetic evaluation carries synthetic provenance.

References

Ehrhardt et al., Reject Inference Methods in Credit Scoring. https://adimajo.github.io/assets/publications/rejectInference.pdf

Binary endogenous-probit sample selection documentation. https://search.r-project.org/CRAN/refmans/endogenous/html/sampselect.html

Robins, Rotnitzky and Zhao (1994), Estimation of Regression Coefficients When Some Regressors Are Not Always Observed. doi:10.1080/01621459.1994.10476818

Validation of Frozen Selection Candidates

Description

Compare common application populations without fitting models; separate funded factual, synthetic and transported-proxy evidence and audit assumption-conditioned selection arithmetic.

Usage

selection_bias_sensitivity(candidates, evaluation_rows,
    selection_results,
  design, reference_model_id, uncertainty = "none",
    cluster_resamples = NULL,
  confidence_level = .95, outcome_comparability = NULL,
  tolerance = 1e-10, context = NULL)

Arguments

Details

Whole horizons must mature even for known early defaults. Negative labels require complete coverage. Pseudo labels never become validation truth. Funding differs from acceptance. Each candidate covers the same population; time/entity leakage is excluded. IPW ESS and trimmed mass remain explicit; zero support prevents full-population identification. AIPW retains mandatory frozen out-of-fold nuisance lineage. No selection model is refitted. Model hashes and source declarations are not externally authenticated. Arithmetic agreement does not approve a policy. Brier and AUC contrasts are descriptive, without an implicit multiple-testing acceptance rule. Bootstrap uncertainty conditions on frozen models.

Value

CalcResult with ValidationResult, scoped metrics, paired differences, exclusions, actual cluster intervals, assumption-labelled transport results and reconciliation violations.

References

Ehrhardt et al., Reject Inference Methods in Credit Scoring, sections 1–2. https://adimajo.github.io/assets/publications/rejectInference.pdf

Exchange Versioned Local Model Artifacts

Description

Writes and reads non-executable language-neutral model metadata and numeric parameters with canonical SHA-256 integrity verification.

Usage

serialize_fitted_model(fitted_model, destination,
    schema_version = "1.0")
load_fitted_model(manifest_path, allowed_schema_versions = "1.0")

Arguments