A0283
Title: Data-driven optimal phase division for improved weather index insurance design
Authors: Jing Zou - Technische Universitaet Dresden (Germany) [presenting]
Ostap Okhrin - Technische Universitaet Dresden (Germany)
Martin Odening - Humboldt-Universitaet zu Berlin (Germany)
Abstract: Past research has demonstrated that the hedging effectiveness of weather index insurance can be improved by decomposing the vegetation period of plants into separate growth phases. We conduct a data-driven selection of optimal division points to improve further the division procedure to mitigate temporal basis risk. Various statistical and machine learning methods are applied and compared concerning their ability to model the weather-yield relation. Using farm-level winter barley yield data from 217 rural households in Saxony, Germany, and corresponding weather data and phenology information, we first separate the whole crop growth cycle of winter barley into four sub-phases according to the phenology reports, secondly, fix the start and end points while relax the internal three points to create 24804 combinations, thirdly search for the optimal division points by employing Polynomial Regression, Generalized Additive Model, Random Forests, Support Vector Machine, and Artificial Neural Networks to model the nonlinear relationship between 8 weather variables and yield variability. The results suggest that optimal division models achieve better performance than benchmark models, and the consistency in data-driven flexible points with the reference ones, especially in terms of Polynomial Regression, Generalized Additive Models and Artificial Neural Networks. In addition, the model fitting results of different methods indicate robustness.
