Hong Kong · Stanford
Quantitative research · world models for finance · generative methods
My core research interests are deep reinforcement learning, world models, the robustness of alternative data, and LLM interpretability — at the intersection with quantitative finance. I'm currently an undergraduate research assistant at both HKUST and Stanford, and co-founder of Alpha Flow and quantitative research consultant at WorldQuant.
I'm an undergraduate at the Hong Kong University of Science and Technology, majoring in Risk Management & Business Intelligence with an Extended Major in Artificial Intelligence and an Honors Track in Mathematics. In 2026 I'm a visiting student at Stanford, taking stochastic processes, machine learning, and convex optimization while looking for research collaborations at the intersection of reinforcement learning, generative models, and decision-making under uncertainty.
My background started in physics — I placed 92nd nationally in the Tsinghua University Physics Climbing Program, roughly gold-medal level on the Chinese Physics Olympiad track — before moving toward stochastic calculus, convex optimization, and machine learning applied to financial markets.
The core project of Alpha Flow (650+ GitHub stars). MicroWorld treats market prices as the emergent equilibrium of many interacting, utility-maximizing investor-agents rather than a single stochastic process to be curve-fit — modeling causal structure (who moves the market and why) instead of correlation alone, which is what lets it target the factor-decay problem that plagues correlation-based alpha mining at its root rather than patching around it.
The framework's mathematical backbone connects to the 2026 Fields Medal awarded to Yu Deng (with Hong Wang) for the rigorous derivation of the Boltzmann equation from N-body Newtonian mechanics (Hilbert's 6th problem). The parallel is structural, not decorative: N particles under Newtonian mechanics → N investors under utility maximization; the N → ∞ limit in both cases is a McKean–Vlasov SDE; the Boltzmann equation's role is played here by a Fokker–Planck–Kolmogorov equation over market state; and thermodynamic equilibrium in a gas corresponds to Nash equilibrium in a market. MicroWorld applies this derivation paradigm to quantitative finance for the first time as a systematic framework, adding the structure finance requires on top — event operators, hierarchical multi-agent layers, behavioral noise, and a full engineering implementation.
In active technical conversation with faculty at institutions including the Fields Institute, and in early-stage dialogue with early-stage-focused U.S. investment firms.
Stanford undergraduate research project, advised by Jehangir Amjad. An SDE-based world-model formulation for markets — encoder, latent dynamics, and controller — extending the Ha & Schmidhuber world-model architecture with mean-field-game and stochastic-control structure. Covers Itô-integral state evolution, a hierarchical mean-field-game layer, and Lyapunov stability analysis of market regime transitions. Drafted as a theory paper targeting quantitative-finance journals.
Generative search over discrete factor structures for robust alpha discovery in noisy financial time series, submitted on WorldQuant's BRAIN platform. Achieved IC = 0.148 out-of-sample, with an emphasis on reducing spurious correlation and factor decay relative to correlation-based objectives.
HKUST UROP with Prof. Tony Cho. Studies how reliable LLM-based extraction of alternative-data signals from unstructured financial text (SEC 10-K filings, earnings releases) is as an input to quantitative models — quantifying hallucination and inconsistency as a source of signal noise, across 3,258 U.S. public companies and 180,000+ model responses. Re-implemented SelfCheckGPT (BERTScore, LLM-as-judge, n-gram perplexity) from scratch after auditing the inherited codebase and finding a measurement bug, and ran the full study on an HPC/Slurm pipeline. Variance decomposition shows question cognitive demand explains 31× more inconsistency than firm size — evidence that signal reliability is task-structural, not just data-quality-driven. Targeting the Journal of Accounting Research.
A modular Python framework for evaluating systematic trading strategies against live and historical market data, built to support rapid iteration on the alpha-mining and world-model work above.
I'm always happy to talk about world models, generative methods, or quantitative finance research — reach out by email or connect on LinkedIn.