Forecasting the Evolving Composition of Guest Origin Markets in Platform Bookings: A Bayesian Compositional Time Series Approach Using Airbnb Data

Battle Count: 67

The paper is primarily focused on tourism demand forecasting and destination marketing, with no direct application to financial markets or quantitative trading. However, the underlying methodology—Bayesian compositional time series modeling on the simplex with Dirichlet likelihoods—has structural parallels to portfolio weight forecasting, market share dynamics, and allocation problems in finance. The BDARMA framework could theoretically be adapted to forecast asset allocation shares or sector rotation probabilities, but the paper does not explore any such applications. The seasonal precision modeling and probabilistic forecasting approach could inform risk management frameworks, but the domain-specific context (tourism bookings) limits direct transferability.

Implementation requires expertise in Bayesian computation (MCMC via Stan), compositional data analysis (ILR transformations, Dirichlet distributions), and time series modeling (VARMA dynamics). The darmaR R package provides some infrastructure, but users must handle data preprocessing (aggregating to monthly compositions, selecting top origin markets, constructing ILR contrast matrices), specifying Fourier seasonal terms for both mean and precision, setting weakly informative priors, running multiple MCMC chains with convergence diagnostics, and implementing rolling-origin forecast evaluation with Diebold-Mariano tests. The seasonal precision specification and centered MA innovations add additional complexity. The proprietary nature of the Airbnb data further limits practical implementation outside the author's context.

The paper uses proprietary Airbnb booking data (2017-2025) that is not publicly available. The darmaR R package (https://github.com/harrisonekatz/darma) and Stan are referenced for implementation, and the BDARMA methodology is described in detail with equations. However, the core empirical analysis cannot be reproduced without access to Airbnb's internal bookings database. Model specifications, prior choices, and evaluation protocols are well-documented, but the data dependency severely limits reproducibility.