FreqAI Feature Engineering, Feature Expansion, and Data Preparation
Summary
This documentation explains how to define model inputs and prediction targets in FreqAI strategies. It distinguishes base features, such as price indicators, volume, and time variables, from configuration-driven expansions across periods, timeframes, shifted candles, and correlated pairs. It also describes a standard feature function for inputs that should not be expanded, and a required function for defining model targets. An example shows how these choices can multiply the number of generated features.
The material also discusses data preparation: SVM and DBSCAN options for identifying outliers, plus principal component analysis for reducing feature dimensionality. These tools can change which observations or inputs reach a model, and the text describes their configuration at a high level. The excerpt is incomplete and includes examples rather than a validated trading system. It provides no forecast accuracy, trading returns, or guidance on avoiding leakage and overfitting, so feature choices and preprocessing still require careful evaluation on appropriate data.
Key ideas
- FreqAI features are defined in strategy functions and can include indicators, prices, volume, and time variables.
- Configuration settings can expand features across timeframes, candle shifts, indicator periods, and correlated pairs.
- Model targets use a separate required function and naming convention.
- SVM and DBSCAN options can identify and remove observations treated as outliers.
- PCA can reduce feature dimensionality, but the excerpt provides no evidence of trading performance.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.