I work on quantitative research where machine learning meets financial structure: models that stay close to pricing theory, use only point-in-time information, and are judged after transaction costs.
| Area | Problem | Methods |
|---|---|---|
| Constrained deep learning | Option-pricing networks that stay close to pricing theory | PDE-style residual penalties · positive-weight input embeddings · ConvLSTM and Transformer encoders · moneyness and temporal positional encodings |
| Cross-sectional alpha | Point-in-time factor research, return ranking, regime-aware risk scaling | point-in-time factors with dynamic universe masking · size and sector neutralization · walk-forward gradient-boosting rankers tuned with Optuna · hidden Markov regime models · formulaic-alpha search with a Transformer generator |
| Reinforcement learning | Hierarchical agents for minute-level trading | MacroHFT-based hierarchical agents · ATR and PPO baselines · Multi-Patch Former state encoder · evaluation pipeline for all strategy variants |
| Market microstructure | Exchange simulation and short-horizon signals | asynchronous exchange simulator · matching against live order-book snapshots · maker and taker fees · delayed market-data release · order-book imbalance features |
| Differentiable optimization | Optimization problems as network layers | OptNet quadratic-programming layers · KKT implicit differentiation · primal–dual interior-point methods |
| Repository | Summary | Stack |
|---|---|---|
Constrained-Deep-Learning-for-Option-Pricing |
Undergraduate thesis. One-day-ahead pricing of TAIEX index options with a dual-branch ConvLSTM + Transformer network and a PDE-style residual penalty; stored test MSE is 48.7% below the ConvLSTM baseline. The README states what each constraint does and does not enforce | PyTorch · ConvLSTM · MultiPatchFormer · FANformer · Kou / FFT benchmark |
MacroHFT (team project, fork) |
Hierarchical reinforcement learning for minute-level ETHUSDT trading, extending MacroHFT (KDD 2024) with a regime-aware Dynamic Hybrid coordinator. My part: ATR and PPO baselines, Multi-Patch Former integration and the evaluation pipeline | PyTorch · DQN variants · expert mixing |
async-exchange-simulator |
asyncio paper exchange: relays a live crypto-exchange feed, matches orders against the live order book, charges maker and taker fees, tracks hedge-mode positions, monitors P&L live. Built in a five-person trading training program; the README lists the modelling limits and what I did not write |
asyncio · WebSockets · matching engine · unittest |
Hangman-AI-Agent |
Hangman agents: PPO and DQN with curriculum learning, per-letter XGBoost and CatBoost classifiers combined with dictionary pattern matching, and an ensemble solver that routes each guess by word length and attempts left | PyTorch · Gymnasium · XGBoost · CatBoost |
twse-attention-stock-alerts |
Watches exchange disclosures of attention stocks, extracts EPS from free-form Chinese text, compares it with the previous quarter and pushes alerts to Telegram | Selenium · regex parsing · Telegram Bot API · Docker |
bb-keltner-squeeze-sar |
Rule-based squeeze-breakout strategy with next-bar execution and commissions, tuned by searching for parameter plateaus instead of single optima | TA-Lib · vectorbt · Optuna NSGA-II · HiPlot |
tw-stock-rnn-trading · tw-stock-rnn-selection |
2024 study that turns GRU / LSTM / ConvLSTM forecasts into trading rules, with a retrospective that reproduces the original backtests and measures how much of their profit came from look-ahead bias and missing costs | PyTorch · vectorbt · Optuna |
Current research code is proprietary and is not hosted here. The READMEs of my own repositories state what the code does and its known limitations; the research repositories also document their data and evaluation.
Languages 
Modeling 
Scientific computing 
Backtesting and infrastructure 
optimization constrained optimization · quadratic programming · Lagrangian duality
KKT conditions · implicit differentiation
statistics Gaussian processes · hidden Markov models · autoregressive models
finance option pricing · factor modeling · exposure neutralization
transaction-cost modeling · market microstructure · walk-forward validation
- Deep Learning-Driven CTA Investment Decisions: Stock Price Prediction and Portfolio Management (Medium, 2024)
- 使用 Bollinger Bands – Keltner Squeeze 配合 SAR 指標的順勢交易策略 (Medium, 2024, in Chinese): a trend-following strategy built on the Bollinger–Keltner squeeze with Parabolic SAR exits