Running FreqAI with Live Retraining and Sliding-Window Backtests
Summary
This guide explains how FreqAI trains and deploys adaptive machine-learning models in live or dry trading and in historical backtests. Live operation can retrain models as capacity permits, use the latest trained model for predictions, and apply limits on retraining frequency or model age. Saved models and predictions support recovery and reuse, while retention settings can remove older models. Backtests instead retrain across successive windows, so users must download enough earlier data to cover both the training period and startup candles.
The guide describes how training-window and backtest-window lengths determine the sequence of models, and how saved predictions can speed later strategy tuning. It warns that feature engineering must avoid future information, that short backtest windows can require many model trainings, and that prediction reuse assumes the strategy can produce equivalent features from a one-row dataframe. Hyperopt is best restricted to entry and exit criteria because feature and target construction cannot be optimized this way. The material is operational guidance, not evidence that any model or trading strategy is profitable; it also flags risks from loading untrusted saved model files.
Key ideas
- FreqAI supports adaptive model training in live trading and in sequential historical backtest windows.
- Live settings can govern retraining cadence, prediction age, and retention of older models.
- Backtests need historical data preceding the requested period so models can be trained before predictions begin.
- Feature engineering must avoid future information even though features are computed across the training range.
- Saved predictions can support entry and exit tuning, but feature and target parameters are outside the described hyperopt workflow.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.