Selected work
A collection of things I've built, tested, and explored across machine learning, causal inference, finance, and beyond.
Causal Inference / Policy
Estimated BNPL’s causal impact using doubly robust AIPW on the 2022 Survey of Consumer Finances. Compared AIPW, IPW, and Causal Forests for heterogeneous effects—strongest among adults under 35—while handling survey weights and multiple imputation with Rubin’s rules across debt, late-payment, and liquidity outcomes.
Quantitative Finance
Built and backtested a market-neutral strategy for energy equities, using clustering and cointegration to identify 50+ mean-reverting pairs per rebalance window from 2019–2025. With transaction costs, sector neutrality, and leverage capped at 2×, the strategy achieved a 1.38 Sharpe ratio and 4.2% maximum drawdown.
Sports Analytics
Scraped and engineered performance and demographic features from Europe’s top leagues, then built L1/L2 regression models reaching roughly 65% out-of-sample accuracy. A K-means and value-performance framework identified high-potential players and assembled cost-efficient squads for recruitment decisions.
Experimental Causal ML
Designed a synthetic-data benchmark of six causal inference and matching methods: PSM, IPW, MALTS, Genetic Matching, Lasso Coefficient Matching, and Causal Forests. Evaluated accuracy, bias, robustness, and computational scalability across sample sizes, feature dimensions, and heterogeneous treatment effects.
NLP / LLM
Architected hybrid retrieval over policy and economics papers with BGE embeddings, BM25, Reciprocal Rank Fusion, cross-encoder reranking, and Qdrant. Fine-tuned DeBERTa-v3-base with QLoRA on 185K FEVER pairs to classify cited claims as supported, refuted, or inconclusive before answers reach users.
NLP / Healthcare
Engineered a pipeline that turns ClinicalTrials.gov records into plain-language summaries and translates them for multilingual access. Automated sentence segmentation, medical terminology normalization, ROUGE, and BERTScore evaluation across 100 clinical trials and 10 languages.
Causal ML
Applied Uplift Random Forests to 13.9M Criteo ad exposures, improving AUUC by 29% over propensity targeting. A rule-based targeting policy, validated on a randomized holdout with bootstrapped confidence intervals, reduced audience size by 80% while retaining approximately 97% of incremental conversions.
Credit Risk
Built a framework for probability of default, loss given default, and expected credit loss using LendingClub loans from 2007–2015 and FRED macroeconomic indicators. Validated calibration and population stability, then stress-tested portfolio losses under adverse macroeconomic scenarios.
Market Risk / Deep Learning
Built a stacked LSTM with 0.954 out-of-sample volatility forecast correlation across seven asset classes and estimated VaR and Expected Shortfall for a $1M portfolio. Benchmarked against GARCH variants with Basel III backtests and measured 3.88× COVID and 10.08× hypothetical-crash VaR uplifts.