FreqAI Retraining: Cadence, Drift and Forward Evidence
How to reason about FreqAI retraining cadence, rolling windows and model drift without turning every new market move into another optimization cycle.
Why retrain at all
Market relationships change. A rolling model can incorporate newer observations, but retraining also changes the system that produces predictions. That makes model lineage and timestamps important.
Faster is not automatically better
Very frequent retraining can chase short-lived noise, while slow retraining can leave a model stale. The appropriate cadence depends on timeframe, sample density, target horizon and model stability.
Record every boundary
For each retrain, record the training cutoff, window length, feature and target version, model version and the forward trades generated afterward. This makes later diagnosis possible.
Our current experiment
V6 currently uses a 30-day training window and a six-hour retraining cadence on a 5-minute trading timeframe. These are experiment settings, not universal recommendations.
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.