By AFML Chapter

Chapter-to-Module Mapping

Chapter 2: Financial Data Structures

Information-driven bars replace fixed-time sampling with activity-based aggregation, producing returns that are closer to IID normal. CUSUM and z-score filters extract structurally meaningful events from the bar stream.

• data_structures — bar construction (dollar, volume, tick, imbalance, run bars). Entry point for raw trade data.
• filters — CUSUM and z-score event filters
• etf_trick — synthetic ETF and futures roll utilities

Chapter 3: Labeling

Triple-barrier labeling converts filtered events into ML labels with controlled profit-taking, stop-loss, and holding-period barriers. Meta-labeling separates direction from sizing.

• labeling — triple-barrier and meta-labeling. Entry point for label generation.
• bet_sizing — probability-to-position-size conversion

Chapter 4: Sample Weights and Uniqueness

Overlapping labels violate IID assumptions. Sequential bootstrap and uniqueness weighting correct for this by measuring and accounting for label overlap.

• sampling — indicator matrix and sequential bootstrap. Entry point for overlap-aware sampling.
• sample_weights — uniqueness and time-decay weighting
• sb_bagging — sequentially bootstrapped bagging ensembles

Chapter 5: Fractional Differentiation

Fractional differencing finds the minimum transformation order that achieves stationarity while preserving predictive long-memory in price series.

• fracdiff — FFD and standard fractional differencing. Entry point.

Chapter 6: Ensemble Methods

Bias-variance decomposition and bagging diagnostics determine whether bagging or boosting improves ensemble quality under financial label structure.

• ensemble_methods — bias/variance diagnostics, aggregation, bagging-vs-boosting recommendation

Chapter 7: Cross Validation in Finance

Standard k-fold CV leaks information through overlapping labels. Purged k-fold with embargo removes this leakage source.

• cross_validation — purged k-fold CV with embargo. Entry point.
• backtesting_engine — walk-forward and CPCV validation

Chapter 8: Feature Importance

Multiple importance methods (MDI, MDA, SFI) are needed to detect substitution effects and unstable features before deploying models.

• feature_importance — MDI, MDA, single-feature importance (Rust)
• fingerprint — model fingerprinting for partial and pairwise effects

Chapter 9: Hyper-parameter Tuning

Tuning must use purged CV to avoid leakage-inflated scores. Randomized search is preferred for large parameter spaces.

• hyperparameter_tuning — grid/randomized search with purged CV scoring

Chapters 10-12: Position Sizing and Robust Backtesting

Backtesting is a scenario sanity check, not a performance estimator. CPCV provides path distributions instead of point estimates.

• bet_sizing — dynamic and reserve sizing for execution
• backtesting_engine — walk-forward, purged CV, and CPCV. Entry point for validation.

Chapter 13: Synthetic Backtesting

Selecting trading rules on a single historical path overfits. Synthetic path ensembles from calibrated O-U processes test rule robustness.

• synthetic_backtesting — O-U calibration, PT/SL mesh search, stability diagnostics

Chapters 14-15: Diagnostics and Strategy Risk

Strategy risk (probability of failing a Sharpe target) is distinct from portfolio risk (VaR, ES, drawdown).

• backtest_statistics — Sharpe, deflated Sharpe, drawdown, holding period
• risk_metrics — VaR, Expected Shortfall, Conditional Drawdown Risk
• strategy_risk — strategy failure probability diagnostics

Chapter 16: Portfolio Construction

Hierarchical methods (HRP, HCAA) avoid covariance inversion fragility. CLA solves constrained mean-variance problems exactly.

• hrp — Hierarchical Risk Parity
• hcaa — Hierarchical Clustering Asset Allocation
• onc — Optimal Number of Clusters
• portfolio_optimization — mean-variance allocators (min-vol, max-Sharpe, efficient risk)
• cla — Critical Line Algorithm

Chapters 17-19: Microstructure, Dependence, and Regime Detection

Microstructure features capture liquidity and order-flow dynamics invisible in OHLC bars. Structural break detection flags regime changes that invalidate model assumptions.

• structural_breaks — Chow, CUSUM variants, SADF bubble detection
• microstructural_features — Kyle/Amihud impact, VPIN, entropy
• codependence — distance correlation, mutual information, variation of information

Chapters 20-22: HPC and Advanced

Atom/molecule parallelism scales independent computations. Streaming analytics maintain bounded-memory indicators for real-time early warning.

• hpc_parallel — partition and dispatch utilities
• combinatorial_optimization — integer optimization with exact baselines
• streaming_hpc — VPIN/HHI streaming with bounded-memory state

See the full per-module detail in the Module Reference Index.