Abstract

Extreme-value theory provides parametric statistical models to describe the behavior of extreme events, such as floods or heat waves, but their application to complex non-stationary datasets, with non-linear and/or non-additive effects of covariates, has been very limited so far. By contrast, machine learning algorithms based on artificial neural networks usually have excellent predictive skills for “average events” (from the center of the distribution), provided we have enough data and we train the algorithms correctly, but they are not designed to model extreme events. The goal of this project is to develop a statistical modeling framework that combines the strengths of both approaches, in order to enhance the prediction of extreme events based on artificial neural networks, and to apply it to study extreme sea surface temperature data in the Red Sea.

Deliverables

1. Learning about extreme-value theory and deep learning
2. Using/Modifying/Extending existing R code to fit machine learning-based extreme-value models, in order to study the behavior of extreme Red Sea surface temperature data
3. Exploratory analysis of the data, and in-depth statistical modeling (site-by-site modeling, and joint modeling)
4. Writing report about results of the data application