Welcome!

 

I lead Forecasting Data Science at Airbnb. My research develops Bayesian workflows for structured forecasting, where predictions must respect known relationships such as summation, bounds, hierarchies, and fulfillment delays. I focus on coherence by construction and empirical calibration, with open diagnostics and reproducible case studies. Applications include finance, political economics, and capital markets. I earned a Ph.D. in Statistics at UCLA under Robert Weiss and Ying Nian Wu.

Recent news

  • [Nov 2025] New note: Impact by design: translating “Lead times in flux” into an R handbook with code. Monitors full lead‑time distributions and turns them into a pickup‑risk bound (link).

  • [Nov 2025] New arXiv preprint: Centered MA Dirichlet ARMA for Financial Compositions. A minimal centering fix improves density forecasts on H.8 bank-asset shares with tied point accuracy and cleaner HMC diagnostics (link).

  • [Oct 2025] My paper with Thomas Maierhofer on forecasting renewable energy mix is now published in Renewable Energy Forecasting: Innovations and Breakthroughs! (link)

  • [Sep 2025] My paper with Erica Savage on the changes in Airbnb trip length distributions is now published with MDPI: Tourism & Hospitality! (link)

  • [Jun 2025] My paper with Liz Medina and Rob Weiss on shrinkage priors for the B-DARMA model is now published with MDPI: Forecasting! (link)

Workflow

These are the forecasting problems I work on most often.

  • Allocation: forecasting how a total splits into parts, with predictions that add up.

  • Delay: modeling and evaluating lead times and target shifts so metrics align with truth.

  • High dimension: pooling and shrinkage for many related series.

Featured Work

  • The foundational Dirichlet ARMA framework for forecasting category allocations with density coherence, multi-horizon simulation, and empirical calibration.

    Links:

  • Time‑varying precision for Dirichlet ARMA improves density forecasts when dispersion clusters.

    Links:

  • Two‑part method for time‑shifted metrics that evaluates on the correct time axis.

    Links: