Machine Learning for Temperature Forecasting in Weather Derivatives

Weather derivatives enable businesses to hedge weather related fluctuations in volume and revenue. Pricing these contracts requires forecasting the underlying weather variable over the relevant risk window. Temperature represents the largest share of exchange traded weather derivatives, yet despite more than two decades of research, there remains little consensus on the most effective modelling approach across diverse locations, with most studies focusing on location specific performance. We evaluate whether machine learning methods can provide a more generalisable solution for temperature forecasting in the context of weather derivative pricing. Our study conducts an extensive comparison of 17 models across 65 global locations, including 12 machine learning approaches (linear models, simple tree-based models, sequential and randomised ensemble tree models, neural network models and support vector machines), 2 Ornstein–Uhlenbeck benchmarks, and 3 classical time series models. We also assess performance over the specific pricing windows used in real world contracts, offering practical value to traders and investors. Model accuracy is evaluated using absolute percentage error, along with root mean square error, mean absolute error, and weighted absolute percentage error. Our results show that machine learning models statistically and significantly outperform benchmark approaches at forecast horizons aligned with typical hedging needs; for example, Random Forest reduces cumulative seasonal temperature forecasting error predicted 4.5 months ahead more than any alternative model. We further document systematic geographical variation in model performance and demonstrate economic relevance through a pricing experiment.