| ABSTRACT | Seasonal Forecasts are critical tools for early-warning decision support systems, that can help reduce the related risk associated with hot or cold weather and other events that can strongly affect a multitude of socio-economic sectors. Recent advances in both statistical approaches and numerical modeling have improved the skill of Seasonal Forecasts. However, especially in mid-latitudes, they are still affected by large uncertainties that make their application often complicated. The MSCA-H2020 project ARTIST aims at improving our knowledge of climate predictability at the seasonal time-scale, focusing on the role of unexplored drivers, to finally enhance the performance of current prediction systems. This effort is meant to reduce uncertainties and make forecasts efficiently usable by regional meteorological services and private bodies. An empirical forecast is here designed throughout a statistical model based on advanced Machine Learning (ML) techniques. Such an approach, in combination with the more classical dynamical one may become critical to improve climate forecasts. In fact, a hybrid model would combine the theoretical foundation and interpretability of physical modeling with the power of Artificial Intelligence (AI), that can reveal unknown or disregarded spatiotemporal features. ARTIST focuses on seasonal prediction of temperature hot/cold extremes in Europe, and here we present a first attempt to predict heat wave propensity across a target season. From a list of possible candidate drivers, a feature selection approach is used to identify the best variable subset for the prediction of seasonal extreme heat propensity. The solution of this selection problem relies to a Genetic Algorithm wrapped around a Random Forest, that repeatedly works with a different variable subset to minimize a cost function. Land-surface candidate predictors, often overlooked by previous literature, represent a large portion of the initial feature set. We also try to improve the physical interpretability of our results by focusing on the Mediterranean area, where we link predictors and targets throughout a regularized regression approach. Preliminary results are encouraging. Future works foresee the hybridization with dynamical predictions, that will hopefully help overcoming the problematics of purely dynamical seasonal forecasts for extreme events in mid-latitudes. |