FreqAI Features and Targets: Design Without Leakage
A practical framework for designing FreqAI features and prediction targets while preserving time ordering and separating model quality from trading quality.
Start with the decision timestamp
Every feature must be reproducible using only information available when the trade decision would have been made. This simple rule is more important than feature count.
A target is not a strategy
The target defines what the model learns to estimate. Entry, exit, sizing and risk rules decide how that estimate becomes a trade. Good prediction statistics can coexist with bad trading economics.
More features can make the problem worse
Large feature spaces increase the opportunity to fit noise, create unstable relationships and hide leakage. New features should have a hypothesis and survive unseen data.
Our V6 context
V6 keeps model training and trading policy conceptually separate. We evaluate the complete PAPER system with realized outcomes rather than presenting model output as profitability evidence.
A practical validation checklist
Keep chronological boundaries, record the exact feature/target/model version, separate model metrics from trading metrics, and evaluate the frozen system on unseen forward data. A change made after observing the result belongs to a new experiment.