Co-founder & CTO · Quantum Signals
Every trading desk answers the same questions by hand: what regime are we in, is today trending or mean-reverting, which features matter right now, how far back should we look? These judgment calls shape every intraday model, and they are usually made with rules and intuition rather than data. This talk presents our work using Temporal Fusion Transformers to let the model answer these questions itself. The architecture mirrors how a trader actually thinks about a moment in the market: a multi-day context for regime, an intraday context since the open for day-character, and a recent window for the immediate prediction. Attention over features and over time replaces hand-tuned choices about what matters. We share what this approach has revealed about US equity index futures microstructure and results from live models on ES, NQ, RTY, and YM.
