FreqAI Models

A more complex machine-learning model is not automatically a better trading system. The model is only one component between data, target definition, execution and risk management.

Start with a baseline

Use a reproducible baseline before adding complexity. If a sophisticated model improves historical fit but fails to improve forward behavior, the extra complexity has not demonstrated value.

Model comparison framework

QUESTIONWHY IT MATTERS

Does validation improve?Reject complexity that only improves training fit.

Are predictions stable?Large changes after every retrain can create excessive strategy turnover.

Does net expectancy improve?Prediction metrics do not pay fees; trades do.

Does drawdown worsen?A higher final PnL can conceal a less usable risk path.

Does it survive new data?Forward behavior is the test of whether historical improvement transfers.

Classification vs regression

Classification can express discrete outcomes such as directional classes. Regression can estimate a continuous future quantity. The useful choice depends on the target and how predictions become trades.

Compare more than accuracy

Prediction metrics do not directly equal trading performance. Examine trade frequency, net results after costs, drawdown, stability across regimes and whether improvements persist in forward PAPER testing.

Version every material change

Changing model family, target, feature set or training window can create a different system. Preserve the prior record.

See our FreqAI experiment →