Projects
Projects
Short-Dated Options Volatility Forecasting and Trading Strategy
This project develops an end-to-end framework that connects short-horizon volatility forecasting with systematic SPY options strategy design. I modeled annualized 5-day realized volatility using GARCH-family models, Random Forests, feedforward neural networks, and LSTM-based models, then compared the forecasts with at-the-money implied volatility to construct a volatility-edge signal.
The strategy component translates this signal into long- and short-volatility option positions, with attention to option structure selection, entry and exit rules, position sizing, and risk management. The full workflow was implemented and backtested in QuantConnect.
Methods: GARCH, EGARCH, GJR-GARCH, Random Forest, feedforward neural networks, LSTM, volatility-edge signal construction, options strategy design, backtesting
Tools: Python, QuantConnect, scikit-learn, arch, pandas
Chess Anomaly Detection with Metropolis-Hastings Sampling
This project develops a probabilistic anomaly detection framework for identifying unusually strong chess play by comparing observed move sequences against a counterfactual human baseline. The method uses Maia-based human move proposals and Stockfish-based centipawn-loss scoring within a Metropolis-Hastings sampling framework.
The project also includes a FastAPI service to make the sampler easier to run, reproduce, and test through structured API endpoints.
Methods: Metropolis-Hastings sampling, empirical p-value testing, Markov chain diagnostics
Tools: Python, FastAPI, Docker, Stockfish, Polars
Momentum-Based Algorithmic Trading Strategy
This project develops a systematic momentum-rotation trading strategy in QuantConnect using selected S&P 500 stocks. I combined technical indicators including SMA, RSI, Bollinger Bands, and 20-day momentum signals to identify long opportunities and construct a rules-based trading framework.
The strategy design included both universe selection and risk-management rules. I screened candidate stocks based on liquidity, sector diversification, recent momentum, volatility, and maximum drawdown, and selected PLTR, GEV, and TSLA as the final trading universe. I also implemented weekly rebalancing, a SPY 200-day market regime filter, fixed and trailing stop-loss rules, trend deterioration exits, and RSI-based exits.
To reduce look-ahead bias, I split the backtest into training, validation, and out-of-sample periods. The final out-of-sample backtest achieved a Sharpe ratio of 2.538 and a net profit of $2.77M on a $10M initial portfolio.
Methods: momentum rotation, technical indicators, universe selection, backtesting, risk management
Tools: Python, QuantConnect, pandas
Web-Scraped Recipe Dataset Construction and Interactive Analysis
This project builds an automated web-scraping pipeline to collect and structure recipe data from the official game Wiki for Stardew Valley. The pipeline extracts recipe structure, ingredient composition, and numerical buff effects from heterogeneous HTML pages.
The final output includes an interactive R Shiny application that supports recipe search, craftable recipe identification, and minimal-additional-ingredient recommendations.
Methods: web scraping, HTML parsing, rule-based filtering, interactive application development
Tools: R, R Shiny, rvest, dplyr, stringr
Insurance Subrogation Outcome Prediction
This project focuses on predicting insurance subrogation outcomes using claim-level information, loss severity indicators, and recovery-related attributes. I performed preprocessing and feature engineering, including categorical encoding, interaction terms, and distributional transformations.
I developed and compared multiple tree-based models, including XGBoost, LightGBM, CatBoost, and TabM. The modeling pipeline emphasized reproducible experimentation, cross-validation, and out-of-sample robustness.
Methods: classification, feature engineering, cross-validation, model comparison, ensemble learning
Tools: Python, XGBoost, LightGBM, CatBoost, scikit-learn, SHAP