Mathematics · Data Science · Optimization
I use statistical learning, optimization and quantitative research to study decisions under uncertainty.
数学、数据与决策：从风险估计，到约束下的资源配置，再到时间序列中的风险管理。
Operations Research & Decision Optimization
| Research | Rolling inventory decisions under demand and supply uncertainty |
|---|---|
| Methods | Mixed-integer programming · demand forecasts · paired Monte Carlo · sensitivity analysis |
| Evidence | 144 synthetic episodes: buffered MILP raises mean economic value by 3.06% / 10.31% / 8.33% across three regimes, while aggregate service stays below base stock. |
| Read | Formulation and results · Paired evidence |
Statistical Learning & Risk Decision
| Research | From credit-card default scores to probability and cost policies |
|---|---|
| Methods | WOE/IV logistic scorecard · calibrated trees · SQL · grouped evaluation · PSI |
| Evidence | 30,000 historical UCI records; validation-selected calibrated RF has test AUC 0.7917 / KS 0.4580 / Brier 0.1322. Hypothetical 1:5 normalized cost is 0.5387/customer versus WOE 0.5600. |
| Read | Protocol and results · Executed metrics |
Quantitative Research & Risk Management
| Research | Trend timing versus volatility sizing in GLD/cash allocation |
|---|---|
| Methods | Time-series signals · self-financing backtest · transaction costs · rolling selection · block bootstrap |
| Evidence | 2020–2025 volatility targeting: 12.28% CAGR / 15.97% drawdown; trend: 12.87% / 31.10%. Rolling selection did not beat fixed 50% exposure over 2015–2025. |
| Read | Research and results · Reference evidence |
Results are conditional research findings: synthetic operations, a historical credit cohort with a reused split, and retrospective single-asset prices. Each repository documents its data, baselines, negative results, assumptions and reproduction workflow.
| Work | Tools and methods demonstrated in these repositories |
|---|---|
| Data and experiments | Python, NumPy, pandas, SQLite SQL, Matplotlib |
| Statistical learning | scikit-learn, XGBoost, WOE/IV, probability calibration, grouped bootstrap |
| Optimization | SciPy/HiGHS MILP, rolling horizon, Monte Carlo policy evaluation |
| Quantitative research | Time-series signals, transaction-cost accounting, walk-forward selection, block bootstrap |
| Reproducibility | Git, configuration files, aggregate CSV evidence, hashes/manifests, numerical validation |
Decision science · operations research · statistical learning · risk modeling · time-series research.
Data → Prediction → Optimization → Decision: estimate uncertain quantities, translate them into an explicit objective and constraints, then test the decision against transparent baselines.