FreqAI Training

The training pipeline matters as much as the model. Time-series leakage or unstable retraining can make an apparently strong system useless.

Training window

A short window adapts faster but contains fewer market regimes. A long window supplies more observations but can dilute recent structure. Treat window length as an experiment parameter and version it.

Walk-forward thinking

PAST WINDOW→TRAIN→NEXT DATA→OBSERVE→RETRAIN

The important boundary is time. Training can use only information available before the prediction interval. The next interval is evidence, not material to improve the model retroactively.

Leakage test: ask whether changing a future candle could change any feature, normalization value, label or model-selection decision used for an earlier prediction. If yes, investigate leakage.

Retraining cadence

Retraining more often does not guarantee better predictions. It increases compute use and can increase model turnover. Choose a cadence that matches how quickly your features and target are expected to change.

Avoid future leakage

Every feature available at decision time must be constructed without information from the future. Splits and transformations must respect chronological order.

What to record at every retrain

RECORDPURPOSE

Training cutoffProves which observations were available to the model

Window lengthMakes adaptation speed reproducible

Feature/target versionSeparates model refresh from strategy redesign

Validation resultTracks whether model quality changes before trading results do

Forward tradesConnects each deployed model period to actual PAPER behavior

This creates an audit trail: when performance changes, you can distinguish a new model fit from a new strategy definition.

Evaluate forward

After historical development, freeze the version and observe it in PAPER mode. If features, target or training rules change materially, record a new version instead of merging the histories.

FreqAI setup →

From the lab

V5.1 showed why an adaptive label is not enough; V6 moved the next ML experiment onto a more structured training and execution framework.

V5.1 lessons →Building V6 →