Building a Profitable FreqAI Plugin with Institutional ML Techniques

The author shares how they built a profitable FreqAI trading plugin by addressing the flaws in standard FreqAI approaches, including misaligned labels, feature overfitting, and lack of signal quality filtering.

Details

The author identified three key issues with standard FreqAI approaches: labels that do not match real trades, feature overfitting due to lack of pruning, and no signal quality filtering. To address these, they implemented several institutional-grade techniques. First, they used Triple Barrier Labeling, which determines the first barrier hit (profit target, stop loss, or time expiry) rather than just predicting future price. They also adapted the barrier sizes based on market regime (bull, bear, sideways) to fix the short bias. Next, they leveraged SHAP feature selection to reduce the feature set from 200+ to 30 relevant ones, refreshing the selection periodically to adapt to regime changes. A second meta-model was trained to filter low-confidence predictions, improving precision at the cost of lower recall. The author also used purged walk-forward cross-validation to eliminate lookahead bias, and added sequential memory features to provide temporal context to the model. The resulting system achieved 63% accuracy in both backtesting and live trading, a significant improvement over the standard 52% live performance.