Projects and code. More on GitHub.
Trade-level microstructure on Kalshi: three-class targets (down / flat / up) at 5–60 minute horizons on chronologically split data (~35M train / ~8.6M test rows), a 1D CNN baseline, six heterogeneous experts (LightGBM, LSTM, Mamba, Moirai, FT-Transformer, CTTS), and a learned PyTorch mixture-of-experts gate over expert softmaxes. Large parquet and checkpoints live on linked Drive; PDF report in repo.
Early-stage Python agent for travel analysis and planning workflows. Currently being built out, so the repo may move quickly.
Predict the Fed’s next rate decision (CUT / HOLD / HIKE) from FOMC statement text: FinBERT mean-pooled embeddings, shifted next-meeting labels, a 1541-dim feature stack (current + previous embedding, cosine drift, pre-meeting Fed Funds, hawk/dove keyword counts), and chronological train/val/test with SVM / logistic regression plus an optional 2-layer Transformer over the last eight meetings when enough data exist. Includes scraper, drift analysis, and `predict.py` for new statements.
Listening pipeline that runs on a schedule: pulls recent Spotify plays, samples catalogs from your latest artists, embeds tracks with a sentence transformer, and ranks unheard songs after Spotify retired the audio-features API — text similarity instead of sonic features. Recommendations land in CSV with deduplication.
Segmented difference-in-differences on U.S.–Brazil routes (2015–2020) to estimate the 2018 Open Skies shock while isolating the 2019 Brazilian recession. BTS T-100 international data, gravity controls, hemisphere seasonality; full paper in the PDF.
Tests whether a “green premium” shows up in domestic airfares by joining DB1B itineraries with T100 segment data and modeling price vs. CO₂ per seat-km. Log-log FE, GBDT, and GAM; notebook ships with pre-rendered outputs.
ESS Round 10 analysis of age and political ideology where two countries had recently shifted right. Distill R Markdown article with interactive HTML; figures and narrative in one place.