2025 Projects.

We introduce the Cradle prior, a new global–local shrinkage prior that combines a half-Laplace local scale with a half-Cauchy global scale to tightly shrink small coefficients while accommodating moderately large ones. Empirical and theoretical results suggest that Cradle often outperforms existing methods (e.g., horseshoe, Bayesian Lasso), particularly in sparse settings with moderately large effect sizes, such as genomic applications.

We analyze daily lead-time distributions for two Airbnb demand metrics, Nights Booked (volume-based) and Gross Booking Value (revenue-based), using compositional data transformations and extensive modeling techniques. They find that revenue-based demand systematically diverges from volume-based demand in a mid-range horizon, with a parsimonious Gamma distribution providing consistently strong day-level fits for both metrics, even outperforming spline-based methods in terms of divergence measures.

We present the B-DARMA model for analyzing compositional time series, exemplified using daily S&P 500 sector trading value proportions. They provide a step-by-step tutorial in Stan, covering model specification, handling simplex constraints, incorporating ARMA and time-varying volatility terms, and performing posterior predictive checks and forecasting—methods that extend to various fields with simplex-valued, temporally dependent data.