The full financial machine learning pipeline
Every step from raw OHLCV to a live trading signal — implemented, tested, and tracked.
CUSUM Structural Breaks
Adaptive threshold event filter (h = mult × σ) detects regime shifts in log-return series — the same structural break technique used in quant funds.
Dollar Bar Constructor
Volume-based non-uniform sampling replaces calendar bars with information-driven bars, reducing serial autocorrelation before any ML step.
Triple-Barrier Labeling
Profit-take, stop-loss, and timeout barriers generate path-dependent labels that respect market microstructure — as described in AFML Ch. 3.
Meta-Labeling Classifier
A secondary XGBoost model gates the primary signal, correcting for the ~78% structural label imbalance inherent in financial time series.
Fractional Differentiation
ADF-optimal d parameter preserves maximum memory while achieving stationarity — avoiding the information destruction of integer differencing.
HRP Portfolio Optimizer
Hierarchical Risk Parity clusters assets by correlation graph — no matrix inversion, no instability. Rebalances your portfolio with institutional rigour.
CPCV Backtesting
Combinatorial Purged Cross-Validation with deflated Sharpe ratio corrects for multiple-testing bias — giving you honest out-of-sample estimates.
RL Trading Agents
PPO and DQN agents (FinRL) trained on fractionally differentiated features with Kelly-sized positions and HRP portfolio constraints.
MLflow Experiment Tracking
Every backtest is logged: Sharpe, deflated Sharpe, max drawdown, Calmar, fill rate. Reproducible runs with full parameter lineage.