freqtrade FreqAI: Add ML to Your Crypto Bot

freqtrade is already one of the most capable open-source crypto bots. Its FreqAI module adds machine learning on top, allowing strategies to adapt as market conditions change. This guide explains how to get started with FreqAI without drowning in complexity.

What Is FreqAI?

FreqAI is a machine-learning framework integrated into freqtrade. It trains models on historical features and generates predictions that your strategy can use. For example, FreqAI might predict whether the price of BTC will be higher in one hour based on recent indicators and order book data.

Key capabilities include:

• Automated feature engineering
• Model training and retraining
• Prediction generation for live trading
• Support for classification, regression, and forecasting
• Integration with scikit-learn, XGBoost, and LightGBM

When FreqAI Helps

FreqAI is most useful when:

• Market conditions change frequently
• Simple rule-based strategies underperform
• You have clear features that predict short-term moves
• You can validate models rigorously out-of-sample

It is less useful when:

• You do not have enough historical data
• Your strategy is already robust
• You cannot explain why the model makes predictions

Setting Up FreqAI

FreqAI is included with freqtrade. Make sure you have a recent version installed:

pip install -U freqtrade

Enable FreqAI in your config file:

"freqai": {
  "enabled": true,
  "identifier": "my_freqai_model",
  "feature_parameters": {
    "include_timeframes": ["5m", "15m", "1h"],
    "include_corr_pairlist": ["BTC/USDT", "ETH/USDT"],
    "label_period_candles": 24,
    "include_shifted_candles": 2,
    "DI_threshold": 0.9,
    "use_SVM_to_remove_outliers": true
  },
  "data_split_parameters": {
    "test_size": 0.2,
    "random_state": 1
  },
  "model_training_parameters": {
    "n_estimators": 100
  }
}

This configuration tells FreqAI to train on 5-minute, 15-minute, and 1-hour candles, predict 24 candles ahead, and use 20% of data for validation.

Creating a FreqAI Strategy

A FreqAI strategy inherits from IFreqaiStrategy instead of the regular IStrategy. The key difference is that you can access predictions from the model.

from freqtrade.strategy import IStrategy, merge_informative_pair
from freqtrade.strategy.interface import IStrategy
import talib.abstract as ta
 
class FreqaiExample(IStrategy):
    def populate_indicators(self, dataframe, metadata):
        dataframe = self.freqai.start(dataframe, metadata, self)
        return dataframe
 
    def populate_entry_trend(self, dataframe, metadata):
        dataframe.loc[
            (dataframe['&-s_target'] > 0.02),
            'enter_long'] = 1
        return dataframe
 
    def populate_exit_trend(self, dataframe, metadata):
        dataframe.loc[
            (dataframe['&-s_target'] < -0.01),
            'exit_long'] = 1
        return dataframe

The &-s_target column contains the model's prediction. In this example, the strategy enters long when the predicted return is above 2%.

Feature Engineering

FreqAI can generate features automatically, but the best results usually come from thoughtful feature design. Common features include:

• Lagged returns and volatility
• Moving averages and RSI
• Order book imbalance
• Funding rates
• Sentiment scores

Avoid creating hundreds of random features. That path leads to overfitting.

Model Selection

FreqAI supports several model types:

• LightGBMRegressor: Fast gradient boosting, good default
• XGBoostRegressor: Powerful but slower
• CatboostRegressor: Handles categorical features well
• SKLearnRandomForestRegressor: Simple and interpretable

Start with LightGBM. It is fast, accurate, and handles financial tabular data well.

Avoiding Overfitting

This is the hardest part of FreqAI. Common mistakes include:

• Using too many features
• Training on too short a history
• Predicting too far ahead for the model's signal horizon
• Ignoring transaction costs

Use these safeguards:

• Out-of-sample validation
• Walk-forward analysis
• Paper trading for weeks
• Limit the number of features
• Include realistic fees in backtests

Live Trading with FreqAI

FreqAI retrains models automatically on a schedule you define. In live mode, it:

1. Downloads recent data
2. Retrains the model if enough new data is available
3. Generates predictions for the current candle
4. Passes predictions to your strategy

Start with dry-run mode to make sure the training and prediction pipeline works without errors.

Monitoring FreqAI Performance

Track these metrics over time:

• Prediction accuracy vs actual returns
• Feature importance drift
• Model retraining frequency
• Live P&L compared to backtest

If prediction quality degrades, the market regime may have shifted and the model needs retraining or redesign.

When FreqAI Adds Noise Instead of Edge

Machine learning is not always helpful. FreqAI can hurt performance when:

• The feature set is too large and the model memorizes noise.
• The label horizon is too short and the signal is dominated by transaction costs.
• The market regime changes and the model has not been retrained.
• The strategy was already profitable without ML, and ML adds complexity without benefit.

If adding FreqAI does not improve out-of-sample Sharpe or reduce drawdown, remove it. A simpler rule-based strategy with good risk management is often better than a complex model that barely beats a coin flip.

Feature Selection Checklist

Before training a FreqAI model, review every feature:

• It can be computed at the exact time of the prediction, with no future data.
• It has an intuitive economic or behavioral explanation.
• It is not highly correlated with another feature already in the set.
• It adds value in a univariate test or feature importance analysis.
• It is stable across different market regimes.

Good features are usually simple: lagged returns, volatility, volume trends, and order-book imbalance. Fancy features often fail out-of-sample.

Bottom Line

FreqAI adds genuine power to freqtrade, but it is not a magic upgrade. Success depends on good features, careful validation, and disciplined risk management. Treat FreqAI as an adaptive layer on top of a solid strategy, not as a replacement for one.

Start simple, validate honestly, and scale up only when paper trading shows consistent edge.

Related reading: Freqtrade Beginner Guide | FinRL vs Stable Baselines 3 | How to Backtest Without Overfitting