kieranjwood/trading-momentum-transformer

Implements a Temporal Fusion Transformer (TFT) attention-LSTM hybrid architecture for futures trading, combining multi-headed attention for regime-aware learning with changepoint detection modules that enable the model to switch between momentum and mean-reversion strategies. Optimizes directly on risk-adjusted metrics (Sharpe ratio) and provides interpretability through attention weight visualization, revealing how the model focuses on momentum turning points and segments trading decisions by market regime. Trains on 600+ continuous futures contracts from Quandl, with optional online changepoint detection at multiple timescales to improve robustness during regime transitions.

603 stars. No commits in the last 6 months.