ml-strategy

Generate trading signals from OHLCV data by applying walk-forward machine-learning training to predict directional returns and outputting clean, clipped signals in [-1,1]. The process applies feature engineering, model retraining, and NaN/inf sanitation to per-symbol historical time series.

Walk-forward training in predictive trading models involves using expanding or sliding windows to retrain machine-learning algorithms on scheduled frequencies. This mechanism prevents future data leakage by ensuring models only learn from past data when predicting future directional returns.

Prevent look-ahead leakage on historical time series by implementing walk-forward training with expanding or sliding-window retraining. This approach strictly separates training data from prediction periods, ensuring machine-learning models never access future information during signal generation.

Yes, you can use scikit-learn to predict directional returns with RandomForest, GradientBoosting, and LogisticRegression models. The workflow applies standardized scaling, NaN/inf sanitation, and walk-forward retraining to produce valid per-timestep trading signals.

Yes, this machine-learning signal generation approach works for crypto, equities, and ETFs. It processes per-symbol historical OHLCV datasets, applying feature engineering and sliding or expanding window training to generate predictive signals across various asset classes.

Machine-learning trading signals contain NaN values due to unsanitized input data or feature engineering gaps. Resolve this by enforcing input validation, sanitizing inf/NaN values during preprocessing, and applying clipped output ranges in [-1,1] to ensure clean signal generation.