Machine Learning Algorithms for Forecasting Weather Derivative Temperatures

Weather derivatives allow businesses to hedge their exposure to risk from variations in the weather that cause fluctuations in business volume rather than catastrophic damage. With increasing weather uncertainty from global warming this poses new challenges as previous work has focused on just a few locations worldwide due in part to the lack of consensus on a standardized weather derivative pricing model.

We propose machine learning (ML) algorithms as an alternative to the current state-of-the-art methods, mainly designed for modelling temperature for specific parts of the world. We compare the three main approaches widely used in previous work, namely the Ornstein-Uhlenbeck (OU) models used by Alaton et al. (2002) and Benth et al. (2007), 3 auto-regressive time series models and 10 machine learning algorithms to find the best for predicting daily average temperatures for the purpose of pricing weather derivatives. The models are rated on their out-of-sample performance across 64 locations around the world over different forecasting horizons used in temperature derivative pricing. Our findings show that machine learning algorithms outperform the other methods when predicting temperature more suited for the forecast windows for which temperature derivatives are priced.

Based on joint work with Michael Kampouridis and Panagiotis Kanellopoulos.