A world model for market microstructure

This post introduces a predictive VAE variant trained with a modified objective that orients representations toward future states, achieving up to +90 events (roughly 9 seconds at 100ms snapshot frequency) of median lead time over realized volatility. It also details a systematic grid search that mapped the full precision-lead tradeoff across 20 model variants, showing precision is structurally bounded between 0.39-0.51, while lead time is the moveable lever. Three complementary models (standard VAE, balanced predictive variant, maximum lead-time variant) form a practical alert system. The post also clarifies two null results: the per-event KL signal does not flag individual adverse selection at the execution level, and the signal is episodic, silent on 29-40% of days and not tested on market stress events beyond January and March 2024.