Berkeley Brothers Investment
Quantitative Researcher Intern
From research hypotheses to intraday alpha.
- Developed an LLM-driven alpha-mining workflow connecting hypothesis generation, executable factors, statistical evaluation, and feedback-guided refinement.
- Built a Polars/Parquet pipeline for 3-second Beijing Stock Exchange order-book data, with a factor-expression DSL supporting approximately 50,000 candidate expressions per day.
- Combined hypothesis filtering and signal deduplication with IC/ICIR, partial IC controlling for mid-price movements, and tail metrics; retained approximately 200 candidate signals ranked by Pareto dominance.
- Trained LightGBM on 171 factors: cross-sectional/time-series ICs of 0.07/0.05 and an OOS annualized Sharpe of 3.78, net of transaction costs.
Backtest: Jul–Dec 2025. Fixed base holdings reset daily; daily returns measured as P&L relative to base-portfolio capital.