Release Notes

v2.4.0

Type: compatible minor release preparation.

Main points:

• Spark transform supports missing_policy="merge" when a merge decision was learned, using native Spark expressions and return_woe=False
• Spark fit with missing merge uses a controlled sampled-to-pandas path and records that the merge decision was learned on the sample
• Spark validation/reporting adds source_profile_, sample-vs-source diagnostics, and missing_sampling_diagnostics
• metadata, bundle outputs, and audit_report.html carry sampling caveats for Spark sampled fit
• audit_report.html is a narrative standalone HTML report with embedded CSS, print-friendly layout, missing policy explanations, bundle inventory, validation alerts, and limitations
• export_bundle(...) includes audit_report.html by default
• docs-site includes a Reveal.js presentation gallery and RiskBands overview deck
• pt-BR/en docs and examples cover missing policy, audit bundle/reporting, and Spark missing merge behavior

Notes:

• missing_policy="standard" remains the default
• Spark-native full fit is not part of this release target
• Spark return_woe=True is still not supported
• no temporal_stable, monotonic_neighbor, custom merge criteria, opaque imputation, or PySpark 4.x support is added
• the audit report supports review, but it does not replace independent formal validation or guarantee regulatory compliance
• joblib/PYSEC-2024-277 is handled as a documented pip-audit exception with no fixed version available; only that advisory is ignored, and the decision is recorded in docs/security/pip_audit_exceptions.md

v2.3.0

Type: compatible minor release.

Main points:

• missing_policy="merge" adds auditable missing-value merge for pandas workflows
• missing_merge_criterion="nearest_event_rate" selects the closest regular bin by fit-time event-rate distance
• missing_merge_criterion="nearest_woe" selects the closest regular bin by fit-time WoE distance
• missing_merge_fallback supports separate_bin and raise
• missing_profile_, missing_decision_log_, missing_merge_candidates_, and missing_merge_map_ preserve the audit trail
• return_woe=True routes merged missing values to the learned destination bin before mapping WoE
• bundle and reporting exports persist merge criterion, fallback, candidates, decision log, and merge map
• standard, separate_bin, forbid, legacy alias compatibility, and RiskBands is Binner are preserved
• docs pt-BR/en and examples describe both merge criteria

Notes:

• missing_policy="standard" remains the default
• PySpark merge is not implemented; PySpark raises an explicit boundary for missing_policy="merge"
• no temporal_stable, monotonic_neighbor, custom merge criteria, opaque imputation, or PySpark 4.x support is added

Documentation update after v2.2.0

Type: docs-only preparation.

Status: documentation update after v2.2.0; no package version bump.

Main points:

• adds a focused missing-policy guide for standard, separate_bin, and forbid
• adds small pandas and PySpark missing-policy demos
• improves docs-site navigation for missing values, bundle fields, and audit logs
• keeps PySpark documented as optional through riskbands[spark]
• does not change core behavior, package version, publish status, tag, or release artifacts

Historical notes:

• at that documentation point, merge policies remained future work
• no opaque intelligent imputation is added
• i18n is not part of this docs-only preparation

v2.2.0

Compatible minor release focused on auditable missing-value policy and compatibility.

Main points:

• missing_policy="standard" is the default and preserves the Sprint A baseline behavior
• missing_policy="separate_bin" is opt-in and creates explicit Missing bins, including categorical missing values
• missing_policy="forbid" raises during fit or transform when selected features contain missing values
• standard is the canonical name for the historical maximize-oriented score strategy
• legacy remains accepted as a compatibility alias for standard
• pandas and PySpark inputs are supported by the missing-policy contract
• bundles persist missing_policy, effective_missing_policy, missing_profile, and missing_decision_log
• old bundles without these fields continue to load as standard
• PySpark remains optional through riskbands[spark] with pyspark>=3.5,<4

Notes:

• merge policies such as merge_nearest_woe and merge_nearest_event_rate are not part of this target
• no opaque intelligent imputation is added
• a full distributed Spark fitting backend is not part of this target

v2.1.0

Compatible minor release focused on the preferred RiskBands name, optional PySpark paths, and validation profiles.

Main points:

• RiskBands is the preferred public estimator name; Binner remains compatible
• min_n_bins records a soft quality status without forcing artificial cuts
• sample_size controls PySpark fit sampling
• pandas/PySpark inputs are detected automatically in fit and transform
• pandas outputs remain pandas; PySpark outputs remain PySpark
• fit(validate=True) and transform(validate=True) create validation profiles with separate fit and transform reports
• v2.1.0 bundles persist schema/version metadata, profiles, separate validation reports, sampling/backend metadata, and data schema when available
• PySpark remains optional through riskbands[spark] with pyspark>=3.5,<4

Notes:

• PySpark fit uses controlled sampling plus the current pandas engine
• PySpark transform and validation profiles use native Spark expressions and aggregated profiles
• A full distributed Spark fitting backend is not part of this release

v2.0.3

Patch release focused on release hardening and deterministic operational behaviour.

Main points:

• stronger categorical handling for rare categories, missing values, and unknown categories
• safer export_bundle(...) outputs with sanitized names and a traceable manifest
• explicit force_numeric support
• stronger quality gates with ruff, coverage-enabled pytest, pip check, bandit, and pip-audit
• supply-chain constraints to avoid the vulnerable ortools 9.11.4210 -> protobuf 5.26.1 resolver path
• README and release governance updates for pandas, Spark/Databricks usage, overrides, auditable export, assets, and local prompts

v2.0.2

Release focada em auditabilidade real, inspeção mais amigável e experiência pública mais robusta.

Principais pontos:

• nova camada pública de export com export_binnings_json(...)
• novo bundle auditável com export_bundle(...)
• metadata_ mais forte, incluindo pesos do score e contexto efetivo do fit
• novas tabelas públicas score_table() e audit_table()
• aliases mais descobríveis para inspeção de bins
• nova camada pública de plots para bad rate, heatmap, share temporal e score components
• correção do alinhamento temporal da estratégia supervisionada, melhorando diagnostics e visualizações
• benchmark assets da documentação regenerados com charts mais largos e menos traces vazios
• docs-site reforçado para onboarding, auditoria e interpretação visual

v2.0.1

Patch release para fechar a publicação pública com consistência:

• corrige a resolução de riskbands.__version__ no pacote instalado fora do source tree
• adiciona teste de regressão para a leitura de versão via metadata distribuída
• preserva integralmente a renomeação para stable, a documentação nova e o fluxo de release da série v2

v2.0.0

Release de consolidação pública:

• renomeação definitiva do valor público de score_strategy de generalization_v1 para stable
• remoção do nome antigo da API pública, exemplos, smoke tests, labels e documentação principal
• docs-site reorganizado para onboarding, primeiros passos e navegação mais clara para novos usuários
• páginas dedicadas para score_strategy, normalization_strategy, woe_shrinkage_strength, Optuna e interpretação de outputs
• notebooks e exemplos alinhados ao fluxo amigável no estilo sklearn e pandas
• preparação explícita do fluxo de release para validação, GitHub Pages e publicação em PyPI via Trusted Publishing

v1.2.0

Evolução importante da ergonomia da API pública:

• Binner mais alinhado a convenções de sklearn e pandas
• suporte amigável a fit(df, y="target", column="feature")
• transform(...) e fit_transform(...) com comportamento mais previsível para DataFrame e Series
• aliases públicos como max_n_bins e monotonic_trend
• novos métodos de inspeção: binning_table(), summary(), report(), score_details(), diagnostics() e plot_stability()
• atributos pós-fit mais fáceis de descobrir
• notebook novo com Plotly e dados sintéticos para onboarding da biblioteca

v1.1.0

Evolução importante da camada de scoring:

• caminho legado preservado explicitamente como legacy
• novo objective temporal introduzido e hoje exposto publicamente como stable
• pesos configuráveis, normalização absolute e shrink de WoE
• integração consistente com Binner, BinComparator, relatórios auditáveis e Optuna
• novo exemplo mínimo comparando legacy versus stable

v1.0.0

Mudanças estruturais importantes já refletidas no repositório:

• rename destrutivo para riskbands
• Binner estabelecido como classe principal pública
• namespace legado nasabinning removido
• direção de documentação orientada a benchmark estabelecida nos exemplos do repositório