The Added Value of Statistical Seasonal Forecasts

Abstract

Seasonal climate predictions can assist with timely preparations for extreme episodes, such as dry or wet periods that have associated additional risks of droughts, fires and challenges for water management. Timely warnings for extreme warm summers or cold winters can aid in preparing for increased energy demand. We analyse seasonal forecasts produced by three different methods: (1) a multi-linear statistical forecasting system based on observations only; (2) a non-linear random forest model based on observations only; and (3) process-based dynamical forecast models. The statistical model is an empirical system based on multiple linear regression that is extended to include the trend over the previous 3 months in the predictors, and overfitting is further reduced by using an intermediate multiple linear regression model. This results in a significantly improved El Niño forecast skill, specifically in spring. Also, the Indian Ocean dipole (IOD) index forecast skill shows improvements, specifically in the summer and autumn months. A hybrid multi-model ensemble is constructed by combining the three forecasting methods. The different methods are used to produce seasonal forecasts (three-month means) for near-surface air temperature and monthly accumulated precipitation seasonal forecast with a lead time of one month. We find numerous regions with added value compared with multi-model ensembles based on dynamical models only. For instance, for June, July and August temperatures, added value is observed in extensive parts of both Northern and Southern America, as well as Europe.