A0902
Title: Joint modeling of temperature mean and volatility for weather derivative evaluation: A neural network approach
Authors: Zelda Marino - University of Naples Parthenope (Italy) [presenting]
stefania Corsaro - Parthenope University (Italy)
Salvatore Scognamiglio - University of Naples Parthenope (Italy)
Vincenzo Di Sauro - Univeristy of Naples Parthenope (Italy)
Abstract: The impact of weather on business is significant. Many sectors, such as tourism and agriculture, face weather-related risks, highlighting the need for hedging. Weather derivatives offer an effective tool to reduce the impact of weather variability on operations and financial outcomes. These instruments are based on weather indices over time whose payoffs can depend on variables like temperature, rainfall, or snowfall. Most traded derivatives use temperature indices such as heating and cooling degree days (HDD, CDD), which reflect deviations from a base temperature. Thus, accurate temperature modeling is essential for pricing and risk management. The aim is to propose a neural network approach for calibrating temperature models in weather derivatives. Building on prior work using fully connected networks to estimate daily means, a subnet for variance is introduced. This joint model offers better insight into temperature uncertainty. Assuming a normal distribution, the mean and variance are jointly calibrated. For explainability, the output is designed to mirror known models. Using NASA's MERRA-2 data, it is shown how the model effectively captures seasonal fluctuations in temperature variability. Numerical comparison with existing approaches in the literature shows the effectiveness of the proposed method in terms of both point prediction accuracy and the coverage probability of interval predictions.
