FreqAI Configuration for Rolling Training, Features, and Model Fitting

Summary

This reference lists configuration options for FreqAI, Freqtrade's machine-learning feature. General settings cover rolling training and inference windows, model identification and persistence, retraining frequency, model expiration, and prediction statistics. Feature settings describe timeframes, correlated assets, future-label horizons, shifted candles, recency weighting, dimensionality reduction, feature importance, and outlier detection or protection.

It also covers training and testing splits, shuffling, model-specific arguments, and PyTorch fitting controls such as epochs or steps, batch size, and early stopping. The parameters help users configure data preparation, model training, and operation across backtests and live runs. The document flags that continual learning is experimental and may overfit or become stuck, and that reversing train/test order is unconventional. It is a parameter reference rather than trading research: it supplies no strategy results or predictive evidence, and the supplied text is truncated in the data-split section.

Key ideas

• FreqAI uses training and backtest window settings to retrain models as the evaluation period advances.
• Feature controls include timeframes, correlated assets, shifted candles, recency weighting, PCA, and outlier handling.
• Training settings cover data splits, model arguments, epochs or steps, batch size, and early stopping.
• Continual learning is described as experimental and potentially prone to overfitting or local minima.
• The parameter list is technical configuration guidance and does not establish trading performance.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.