§0 New York, NY
Senior Data Analyst with 8+ years in utility analytics, applying causal inference, time-series forecasting, and NLP to problems that resist easy answers.
§1 About
I've spent eight years at Consolidated Edison turning New York City's messiest data into decisions — designing SQL/SAS ETL pipelines over 500M+ smart-meter records that feed demand forecasting and rate design models, and refactoring legacy workflows to cut manual processing time by 25%.
Along the way I completed an MS in Computer Science at NYU Courant (on top of a BS in Mathematics from NYU Tandon, with minors in Economics and Finance), and built a toolkit centered on causal inference — difference-in-differences, synthetic control — plus time-series analysis and NLP pipelines.
What drives me is the same thing that drew me to competition math as a kid: problems with genuine resistance. I'm looking for work where the questions are hard, the stakes are real, and the answer isn't in the back of the book.
§2 Selected work
P·1
Do FOMC announcements move markets? Difference-in-differences with clustered standard errors across 96 announcements (2014–2025) identified a statistically significant 12 basis point volatility premium in rate-sensitive sectors (p = 0.043), validated by parallel pre-trends. Synthetic control on the March 2020 emergency cuts showed financials underperformed their counterfactual by 13.6%.
P·2
An end-to-end NLP pipeline ingesting 2 GB+ of Reddit posts from four finance subreddits and every 119th Congress bill, scored daily with FinBERT and joined to S&P 500 sector returns. The finding: legislative sentiment is a 2-day lead indicator (p < 0.05 in every Granger-causality test at lag 2) — Reddit and news never were.
P·3
A hybrid recommender built on a 4,000-playlist slice of Spotify's Million Playlist Dataset, enriched with audio features pulled batch-by-batch from the Spotify Web API. Content-based filtering builds a user profile from a playlist's track features and ranks candidates by cosine similarity; collaborative filtering scores tracks by co-occurrence across playlists. Blending the two (α = 0.6) beat every baseline — popularity, logistic regression, decision tree — and each standalone method on its own.
§3 Contact
I'm exploring full-time opportunities in NYC (hybrid or remote). If your team works on hard quantitative problems, I'd love to connect.