Hong Kong · Stanford

Quantitative research · world models for finance · generative methods

My work sits at the intersection of deep reinforcement learning, world models, robust decision-making, and quantitative finance. At HKUST, I do research with Prof. Nian Si and Prof. Tony Cho, and run an independent project on limit order books. I am also the Co-Founder & CTO of AlphaFlow, where I lead the technical development of MicroWorld.

About

I'm an undergraduate at the Hong Kong University of Science and Technology, majoring in Risk Management & Business Intelligence, minoring in Mathematics, and pursuing an Extended Major in Artificial Intelligence. 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.

New to the idea? Read what a world model of finance is and how MicroWorld builds one.

Ten-year ambition

My goal is to turn world-model research into infrastructure that financial markets actually run on, and to grow AlphaFlow into a billion-dollar company before I turn 30.

Research & Projects

Publications

From Sharpe Ratios to Executable Evidence: A Reproducibility Case Study of a Financial DRL Ensemble

Hongjin He. A CPU reconstruction of the FinRL-Meta ensemble benchmark (ten seeds; mean annualized Sharpe 0.910 against 1.53 reported, with equal-weight buy-and-hold at 1.434) plus a ledger of eight published financial DRL claims, none of which has a verified complete binding between claim, code, data, configuration and output. Preprint on SSRN (2026); under review at Transactions on Machine Learning Research (TMLR).

reproducibilitydeep RLquantitative financebacktesting

Sharp Ranking Bounds for Frozen Binary Forecasts with Unresolved Outcomes

Hongjin He. When some outcomes in a forecasting panel are still unresolved, a ranking of two binary forecasters computed on the resolved questions may differ from the ranking on the full panel. The paper gives sharp lower and upper bounds on the paired Brier contrast under unrestricted binary completions, a width-reduction rule, and exhaustively verified numerical checks (maximum absolute error about 2×10−16). Under review at Journal of Forecasting.

forecast evaluationBrier scorepartial identificationreproducibility
HKUST · Research

Distributional Forecasting from Limit Order Books — Independent Project

Studies whether a DeepLOB-style model can forecast the full conditional distribution of short-horizon returns rather than only an up/flat/down label. The work has reproduced the FI-2010 classification baseline and built an end-to-end continuous-return density pipeline on raw order-book data. Current pilots show that strong unconditional heavy-tail baselines remain difficult to beat, so the next step is to isolate whether the bottleneck lies in the signal, target, backbone, or distributional head.

limit order booksDeepLOBdistributional forecastingcalibration

Robust Decision-Making under Distribution Shift — Prof. Nian Si

Develops and stress-tests ideas in distributionally robust reinforcement learning, average-reward control, and heavy-tail robustness. The research program uses preregistered experiments and null tests to separate structural results from trading claims; it includes calibrated uncertainty sets, certified computation, negative evidence on crisis-timing overlays, and a pilot on robust estimation under heavy-tailed rewards.

robust RLdistribution shiftaverage rewardheavy tails

Robustness of LLM-Extracted Financial Signals — Prof. Tony Cho

Studies how reliably language models extract decision-relevant signals from SEC filings and earnings disclosures. The project combines a large-scale model-response pipeline with semantic-consistency metrics and by-question statistical analysis, focusing on when hallucination and inconsistency make a plausible answer unsuitable as quantitative-model input.

alternative dataLLM evaluationfinancial disclosuresHPC
AlphaFlow · Company & Independent Work

AlphaFlow & MicroWorld — A Multi-Agent World Model of Finance

As Co-Founder & CTO of AlphaFlow, I lead the research and technology behind MicroWorld: a world model of finance in which every market participant is played by an LLM-driven agent that stands in for one real participant, with its real positions, leverage and limits, trading against the others under real market rules.

We build it in two stages. First on-chain, where every account's positions and liquidation prices are public, so agents can be calibrated and checked against ground truth. Then all of finance, where the same method carries over to markets whose positions are hidden. Today MicroWorld reconstructs about 3,100 real accounts and $7.8B of positions on Hyperliquid and replays how shocks cascade through forced liquidations.

MicroWorld's current demos are simulations on real positions; they are not yet evidence of forecasting performance or commercial validation.

multi-agent world modelsLLM agentson-chain marketsrisk infrastructure

A Mathematical Framework for World Models in Quantitative Finance

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.

SDEsmean-field gamesstochastic controlworld models

Formal Mathematics in Lean 4

Seven Lean 4 formalizations of Erdős-type results, including the Duffin–Schaeffer theorem, submitted to the Justin Sun Prize and currently under review.

Lean 4formal verificationnumber theory

Diffusion-Based Alpha Mining

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.

diffusion modelsalpha mininggenerative search

Real-Time Quantitative Backtesting Framework

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.

Pythonbacktestinginfra

Experience

May 2026 –
Present
AlphaFlow logo
Co-Founder & Chief Technology Officer
Lead the research and technical roadmap for MicroWorld, with an initial focus on multi-agent market simulation, institutional risk transmission, and decision-support workflows.
Mar – May
2026
Quantitative Research Consultant
WorldQuant
Design and submit systematic alpha signals on the BRAIN platform using diffusion-based generative search over factor structures.
Jun – Sep
2026
Visiting Student — International Honors Program
Stanford University
Studied stochastic processes, machine learning, and convex optimization while pursuing research collaborations in reinforcement learning, generative models, and decision-making under uncertainty.
2026 – present
Undergraduate Researcher
HKUST — Prof. Nian Si & Prof. Tony Cho
Work across limit-order-book distributional forecasting, robust decision-making under distribution shift, and the reliability of LLM-extracted signals from financial disclosures.
Nov 2025 –
Apr 2026
Co-Founder, Quantitative Finance Arena
GBA Redbird Entrepreneurs Alliance
Co-built a quantitative strategy platform with Cantonese voice-input support.

Honors & Talks

Tsinghua University logoTop 92 nationally, Tsinghua University Physics Climbing Program (gold-medal equivalent, Chinese Physics Olympiad track)Dec 2023 – Feb 2024
Rated "Excellent" (top tier), Tsinghua University Physics Bridging Program — earned direct-admission recommendationMar – Aug 2024
Future Science Prize logoHost & Panel Guest, Future Science Prize AI Roundtable Forum — joint discussion with faculty including Prof. Mengdi Wang (Professor, Princeton University; Director, AI Innovation Center), Prof. Wei-Ying Ma (President, Hong Kong Institute of AI and Science; Chair Professor, Tsinghua University; ACM Fellow and IEEE Fellow), and Prof. Tim Kwang-Ting Cheng (Vice-President, HKUST; Chair, Asian Young Scientist Fellowship 2025 Steering Committee) Nov 2025

Contact

I'm always happy to talk about world models, generative methods, or quantitative finance research — reach out by email or connect on LinkedIn.