A research & engineering implementation of end-to-end Temporal Fusion Transformers (TFT) designed specifically for multi-horizon quantitative volatility forecasting and dynamic risk budgeting.

Key Quantitative Capabilities:

  • Interpretable Self-Attention: Variable selection networks identifying dominant macro regime indicators and high-frequency order book imbalance drivers.
  • Quantile Loss Optimization: Simultaneous estimation of 10th, 50th, and 90th conditional volatility quantiles for tail-risk hedging.
  • Production Pipeline: Optimized GPU batch inference processing 5,000+ financial time series with sub-millisecond throughput.

View Technical Paper & Code on GitHub