Configuring FreqAI Features, Targets, and Model Predictions

Summary

This documentation explains how to configure FreqAI within a Freqtrade configuration and strategy. It outlines core settings for training and backtesting periods, model identification, timeframes, correlated pairs, shifted candles, indicator periods, labels, and data splitting. It also describes where strategy code should call the FreqAI pipeline and how feature columns and target columns are designated, including automatic feature expansion across timeframes, periods, shifted candles, and correlated pairs.

The examples cover price-derived and indicator features, target construction, prediction statistics, and prediction trust signals based on outlier checks. Later material also illustrates defining a PyTorch regression model and classifier labels, plus optional compilation for performance. These are implementation examples rather than evidence of trading performance. The document cautions that features belong in designated engineering methods, that feature counts can grow rapidly, and that compilation can make errors harder to diagnose; model quality still depends on suitable data, labels, and evaluation.

Key ideas

• FreqAI setup requires both configuration parameters and strategy integration.
• Feature columns use a designated prefix and are created in feature engineering methods.
• Target columns are defined separately and provide the labels the model learns to predict.
• Configured timeframes, indicator periods, correlated pairs, and shifted candles can multiply feature counts.
• Prediction metadata includes statistics and signals that can help assess outliers and reliability.

Tags

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