Factor models learn correlations. World models learn causation. The difference is everything.

I came to machine learning from mathematical physics — with the conviction that the right mathematical structure, once found, makes everything else obvious.

My work sits at the intersection of rigorous theory and empirical systems: I prove theorems and ship code. Sometimes on the same day. Usually not.

journey
    title Research Journey
    section Physics
      Olympiad Training: 9: Me
      Tsinghua Talent Program: 8: Me
    section University
      HKUST AI+Math: 7: Me
      UROP Hallucination Research: 8: Me, Prof
    section Exchange
      Stanford CS IHP: 9: Me
      WorldQuant Research: 8: Me
    section Building
      Alpha Flow Co-founder: 9: Me
      7 Theorems Paper: 8: Me

Working paper · July 2026 · HKUST + Stanford IHP

A mathematical framework connecting world models to quantitative finance. Seven original theorems — no hand-waving.

E-Game-C:  x_t → Encoder (VAE) → z_t → Game Module (MFG) → z_{t+1} → Controller π* → a_t
                                              ↑
                              replaces the RNN in Ha & Schmidhuber (2018)
                              with a mean-field game equilibrium operator

Key result (Theorem 3): prediction error has a fundamental lower bound of σ²_τ·h + λ_η·m²_η·h — irreducible regardless of model complexity. Know your enemy.

Independent research · PyTorch · WorldQuant BRAIN

Diffusion models don't find the best alpha. They find a distribution of good alphas — which is exactly what you need when the goal is a low-correlation portfolio, not a single factor.

Mean IC = 0.148   vs   random baseline ≈ 0.05   →   3× improvement
Training time: < 5 min on Google Colab T4

Honest caveat: no train/test split yet. Real-world IC is probably lower. The mechanism works; productionizing it is the next step.

UROP · HKUST Business School · Prof. Tony Cho · Target: Journal of Accounting Research

5,552 U.S. firms. 50GB+ SEC filings. The finding that keeps me up at night: all tested models (GPT-3.5 through GPT-5) fabricate confident, plausible-sounding answers when the filing contains no relevant information whatsoever.

We called it D-type hallucination: the model doesn't know what it doesn't know.

Scale:    5,552 firms · SEC 10-K/10-Q · HKUST HPC4
Result:   GPT-4o: +21.5% consistency (BERTScore), +54.8% (Prompt Score) vs GPT-3.5
          D-type: 100% fabrication rate across all models when no ground truth exists

Data visualization · Built before and during Stanford exchange

Nine interactive visualizations mapping Stanford's 7 AI labs — research trends, collaboration networks, and industry talent flows from 2018 to 2026. Built as a research planning tool, not just a visualization exercise.

Educational · Archived

An ambitious attempt to build a global stock trading platform in one semester. Failed due to data costs, infrastructure complexity, and the general hubris of thinking this was a weekend project. The post-mortem is honest and worth reading if you're considering the same mistake.

theory   = ["Stochastic Calculus", "Itô/Lévy Processes", "Mean-Field Games",
            "Convex Optimization", "Probability Theory", "Groupoid Algebra"]

ml       = ["Diffusion Models", "PPO/DDPG", "Transformers", "Flow Matching", "VAE"]

quant    = ["Alpha Factor Design", "IC/ICIR", "Backtesting", "Order Book Dynamics"]

systems  = ["PyTorch", "Python", "HPC/Slurm", "WRDS/SEC Analytics",
            "OpenAI API", "edgartools", "Git"]

Physics Olympiad → Tsinghua Physics Talent Program (rank 92nd nationwide, gold-medal equivalent) → HKUST AI+Math+Risk Management → Stanford CS Exchange

The physics training instilled one habit above all: ask why does this equation have this form? before asking how do I fit it? In markets, that question leads to stochastic dynamical systems — not factor regressions.