Extreme Value Prediction for Zero-Inflated Data

Abstract

Depending on the domain, there may be significant ramifications associated with the occurrence of an extreme event (for e.g., the occurrence of a flood from a climatological perspective). However, due to the relative low occurrence rate of extreme events, the accurate prediction of extreme values is a challenging endeavor. When it comes to zero-inflated time series, standard regression methods such as multiple linear regression and generalized linear models, which emphasize estimating the conditional expected value, are not best suited for inferring extreme values. And so is the case when the the conditional distribution of the data does not conform to the parametric distribution assumed by the regression model. This paper presents a coupled classification and regression framework that focuses on reliable prediction of extreme value events in a zero-inflated time series. The framework was evaluated by applying it on a real-world problem of statistical downscaling of precipitation for the purpose of climate impact assessment studies. The results suggest that the proposed framework is capable of detecting the timing and magnitude of extreme precipitation events effectively compared with several baseline methods.