A spatially-continuous neural network temperature model for weather derivatives evaluation

Abstract This manuscript develops a spatially continuous temperature model for pricing weather derivatives. We formulate temperature at any specific point as a function of its longitude and latitude coordinates. Using satellite climate data, we learn this function using deep neural networks and use it to model and forecast temperature. More specifically, our model allows for describing temperature dynamics at any point within the zone of interest, even when there is no available temperature data at that specific location. This approach enhances our ability to evaluate climate risk with greater precision across different regions. Furthermore, we design a neural network architecture that maintains model explainability, which is crucial for promoting the use of these methods in real-life applications. Through numerical experiments with NASA-MERRA-2 satellite data, we illustrate the model’s application in pricing Heating Degree Day (HDD) derivatives. Additionally, we explore potential extensions of the model to improve forecasting accuracy further.