When is extreme too extreme? A Bayesian approach to modeling NYC rainfall with PyMC

1:45pm - 1:56pm on Friday, October 6

Jorn Mossel



On September 1st 2021, hurricane Ida hit NYC, causing extreme rainfall resulting in deadly flash floods. The weather station at Central Park recorded its wettest hour in history with 3.15 inches of rain, shattering the previous record. One might wonder: is such an extreme event statistically expected to happen once in a while (say every 100 years), or is this an indication of a changing climate? In this talk we take a purely statistical approach to tackle this question. We start with the introduction of Extreme Value Theory, a branch of statistics devoted to extreme events such as floods, stock market crashes and reliability engineering. We apply this to NYC rainfall data with a simple python implementation. Next, we switch gears and use a Bayesian approach with PyMC, to incorporate data from nearby weather stations for better accuracy. Along the way we get a quick intro of hierarchical models and Gaussian processes.

Want to edit this page?