ML Backtest

Single-asset walk-forward pipeline that trains a calibrated logistic-regression classifier on technical features and routes its signals through a volatility-targeted execution engine with realistic transaction costs.

Strategy

• Data: Yahoo Finance downloader or CSV ingestor normalises timestamps and appends log returns.
• Features: Rolling volatility, RSI, slow-SMA distance, and Donchian-channel location (lagged returns and fast-SMA distance are scaffolded but presently disabled).
• Model: L1-penalised logistic regression with optional isotonic/Platt calibration, searching probability thresholds and minimum holding periods under exposure/turnover/return constraints.
• Execution: T+1 fills, EWMA volatility targeting, leverage caps, and basis-point cost model convert predictions to log-PnL and equity.
• Validation: Walk-forward rebuild of the full pipeline per fold with Monte-Carlo block-permutation Sharpe p-values and Wald–Wolfowitz runs tests on out-of-sample returns.

Summary of what this repository achieves

• Walk-forward evaluation that avoids look-ahead while aggregating fold metrics, positions, and returns.
• Metric suite covering Sharpe, Sortino, Calmar, drawdowns, VaR/ES, turnover, and exposure.
• YAML-driven configuration for reproducible experiments via configs/ml_lg.yaml.
• Modular codebase (ml_backtest/) split across data loading, indicators, feature engineering, execution, models/policies, and statistical tests.
• Notebook (main.ipynb) to inspect outputs interactively after a run.

Quickstart

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
# then run Jupyter notebook in the .venv