REPLYOPSAI RESEARCH NOTE
FREQAI GUIDE

FreqAI Setup: From Installation to First Model

Treat the first model as an experiment, not a finished trading system. The useful workflow is reproducible training followed by forward PAPER observation.

Define the experiment first

Choose markets, timeframe, prediction target, training window and retraining cadence before optimizing results. Record these settings so later model versions remain comparable.

A setup that can be tested

LAYERDECISION TO FREEZE
Market scopePairs, timeframe and trading mode
FeaturesInputs available strictly before each decision
TargetExactly what the model is trained to estimate
TrainingWindow, split method and retraining cadence
Trading policyHow predictions become entries, exits and position sizes
EvidenceNet closed trades, drawdown and forward observation window

Changing several of these layers at once makes attribution difficult. Version the complete experiment, not just the model file.

Features and target

Features describe the information available to the model; the target defines what it is asked to predict. Leakage between future data and training inputs can make historical results meaningless, so time ordering matters.

Training and retraining

FreqAI can periodically retrain models as new observations arrive. More frequent retraining is not automatically better: it increases compute cost and can make behavior unstable if the data window or target is poorly specified.

Move to PAPER evaluation

Track net PnL after modeled costs, trade count, drawdown, win/loss distribution and behavior by market regime. Preserve failed versions instead of silently replacing them.

Example from our lab

Our current V6 configuration tests BTC, ETH and SOL on 5-minute data with a 30-day training window and six-hour retraining cadence.

Follow the experiment →