Bayesian Region Adaptive Feature Selection for High-Dimensional Climate Data in Hurricane Predictive Modeling

Atlantic hurricanes have been among the most damaging of all-natural and human-caused hazards afflicting the United States. Atlantic basin seasonal tropical activity forecasts prior to the hurricane season have received considerable attention due to the catastrophic and long-lasting damage to daily life and the economy.  Various statistical, dynamic, and hybrid models have been proposed for hurricane prediction in the past decades, however, the performance of previous forecasting procedures showed a notable variability, likely leading to inaccurate prediction. In this talk, we focus on understanding the climate system that leads to an active hurricane season, by extracting significant features among high-dimensional global field predictive factors. One of our main tasks is to select regions that are most relevant to Atlantic hurricane occurrence. We consider a Bayesian high-dimensional Poisson regression model to predict the number of hurricanes in each season. A novel region-adaptive feature selection prior built upon a dependent continuous shrinkage procedure is proposed, to enhance region selection of global field features. We demonstrate numerical results in various scenarios where the true active features contain different spatial structures.  Yearly prediction accuracy in hurricane counts is compared with other existing models using historical data from 1950 to 2013.

Bio
https://sites.google.com/view/shuangzhousomss