Feature selection for data-driven seasonal forecasts of European heatwaves

The early warning of heatwaves using seasonal forecasting systems has the potential to mitigate economic losses and risk to life. Because of the limited reliability and computational expense of dynamical forecasting systems, efforts in recent years have turned to exploiting Machine Learning. Here, an inexpensive approach to forecasting summer heatwaves over Europe is developed, using an optimisation-based feature selection framework to detect a combination of variables, domains and time-lags used to skilfully predict heatwaves. The purely data-driven forecasts are shown to match, and in places outperform, the skill of the state-of-the-art dynamical multi-model products. Moreover, low skill over Scandinavia, a long-term issue common to most dynamical systems, is improved in our data-driven approach. This work also highlights that the greatest contribution to skill comes from predictors at 4-7 weeks time-lag (e.g. mid-March), and identifies predictors which can form the basis for future studies on mechanisms.