Machine Learning & NLP
Quantitative economist and NLP data scientist specializing in turning unstructured economic and financial text into actionable signals. I build end to end NLP pipelines in Python and R. I leverage PyTorch and Hugging Face transformers to construct supervisory sentiment indices from large scale regulatory corpora. My work applies fine tuned transformer models (BERT, RoBERTa, FinBERT), machine learning, and econometrics to problems in financial stability, sovereign risk, and banking supervision. I am experienced in SQL based data pipelines, large scale text processing (7.5M+ observations), and communicating model trade offs clearly to non technical stakeholders.
Technical Skills
Selected Projects
- Constructed a sentence level supervisory sentiment index from 1,388 FDIC communications (1978–2025) using domain specific NLP methods, topic modeling, and transformer based language models, including FinBERT and RoBERTa.
- Developed an end to end NLP pipeline for automated data collection, text preprocessing, sentence classification, sentiment extraction, and index construction using Python, PyTorch, and R.
- Developed econometric methods to isolate unexpected innovations in supervisory sentiment from publicly available economic and financial information.
- Applied event study methods to establish that favorable supervisory sentiment innovations are associated with higher bank excess stock returns, with no statistically significant response in the broader equity market.
- Integrated text based supervisory sentiment measures with over 1.4 million bank quarter observations to examine deposit pricing, funding composition, market concentration, and financial stability.
- Applied local projections, Bayesian vector autoregressions, panel fixed effects models, and logistic regressions to examine the dynamic and heterogeneous responses of banks to supervisory communication.
- Found that favorable supervisory sentiment innovations are associated with increased uninsured deposits, lower deposit funding costs, and lower odds of subsequent bank failure.
- Used NLP methods to identify digital fiscal reform events across 46 emerging economies from policy and news text, combined with World Bank DAI and EMBI spread data.
- Built and calibrated a quantitative sovereign default model; simulated 25,000 paths to generate impulse responses, reproduce empirical moments, and produce model simulated datasets.
- Documented that digital fiscal reforms are associated with improvements in tax revenue, translating NLP identified reforms into macro-financial signals.
- Develop machine learning models that integrate text-derived supervisory sentiment with bank level financial data to assess whether regulatory communication improves the prediction of bank failures beyond conventional financial indicators.