Machine Learning Signals for OHLCV Data

Predict short-term price direction from multi-ticker OHLCV data, generate trading signals, and backtest a simple equal-weight strategy.

Setup

• Python 3.9+ with pandas, numpy, scikit-learn, matplotlib, seaborn.
• Data/config: market_data_ml.csv, tickers.csv, features_config.json, model_params.json.

How to Run

1. Train and evaluate models (uses configs automatically):

python train_model.py

Outputs land in outputs/ (metrics, predictions, confusion matrices, feature importance).

2. Turn predictions into signals (example uses the best model from the default run):

python - <<\"PY\"
from signal_generator import generate_signals
generate_signals('outputs/predictions_RandomForestClassifier.csv', threshold=0.55)
PY

3. Backtest the generated signals:

python - <<\"PY\"
from backtest import run_backtest
stats = run_backtest('outputs/signals_RandomForestClassifier.csv')
print(stats)
PY

Backtest saves equity curves (.csv and .png) plus metrics JSON under outputs/.

Files

• feature_engineering.py – builds lagged returns, SMA, RSI, MACD, and labels (direction from next-day return).
• train_model.py – trains Logistic Regression and Random Forest (optionally XGBoost if installed), cross-validates with TimeSeriesSplit, scores on a holdout set, and saves predictions/plots.
• signal_generator.py – converts predictions to long/flat (or long/short) signals and attaches strategy returns.
• backtest.py – simple equal-weight daily strategy vs. buy-and-hold baseline with equity curve plotting.
• tests/test_pipeline.py – unit coverage for feature creation, training output shapes, and signal/backtest plumbing.

Notes

• Feature selection and label name come from features_config.json.
• Model hyperparameters come from model_params.json.
• Change thresholds or allow shorts via generate_signals(..., allow_short=True).
• No transaction costs are included; signals use next-day returns derived during labeling.