Machine Learning · 2025

S&P 500 Deep Learning Forecasting System

ML Research & Engineering

Temporal Fusion Transformer–based system for forecasting S&P 500 returns from mixed-frequency macroeconomic and market data, with ARIMAX and LSTM baselines.

Highlights

  • Built a full pipeline spanning 1991–2025 combining daily market indicators (VIX, yields, spreads) with monthly macro releases aligned via ALFRED vintage data to avoid look-ahead bias.
  • Implemented TFT, LSTM, and ARIMAX models, showing that weekly TFT and multi-horizon daily TFT achieve around 2% excess directional accuracy over naive baselines.
  • Analyzed failure modes such as prediction collapse and encoder–output gradient disconnect, and explored regime-aware attention and loss penalties for improved interpretability.