hongjin-he/MicroWorld

What it solves

MicroWorld addresses the failure of traditional quantitative finance models (like factor mining and time-series ML) which rely on historical patterns that decay as soon as they are discovered and traded. It replaces pattern recognition with a "world model" that simulates the actual market participants—institutions, regulators, and retail investors—to predict the "denoised equilibrium price" rather than short-term price ticks.

How it works

Instead of analyzing price data residues, MicroWorld models the market as a multi-level, multi-objective game played on a fully connected network of agents. It uses a hierarchical mean-field system to simulate how different player types, each with unique information sets and constraints, interact to reach an equilibrium. The system identifies "behavioral wedges" (divergences between market price and equilibrium) and monitors the "institutional mean field" to signal when unstable regimes are likely to resolve into market unwinds.

Who it’s for

It is designed for mid-to-long-horizon holders (retail and institutional) who operate on timescales of weeks to months. It is explicitly not intended for high-frequency or intraday traders, as the model posits that behavioral noise dominates those shorter horizons.

Highlights

• Causal Attribution: Every prediction is derived from specific agent interactions and constraints rather than post-hoc statistical correlations.
• Information Asymmetry Modeling: Explicitly models the gap between what different agent types (e.g., institutions vs. retail) see and know.
• Denoised Price Prediction: Focuses on the equilibrium value of an asset once behavioral noise is stripped away.
• Regime Stability Detection: Uses a Lyapunov crisis indicator to warn of unstable market regimes before crashes occur.