The goal of this thesis is to produce probabilistic foehn forecasts by applying model output statistics (MOS) to numerical weather prediction (NWP) data. These forecasts were developed for south foehn and north foehn at two locations in the Gotthard region in Switzerland. The forecasts are valid for +15 h (15 UTC same day) and +39 h (15 UTC next day) lead times.
Multiple statistical models for producing probabilistic foehn forecasts were created and their performance tested. The input data consists of direct model output (DMO) of the NWP models European Centre for Medium-Range Weather Forecasts (ECMWF) and Consortium for Small-Scale Modeling (COSMO-2), foehn observations and seasonality. The forecasting models are trained and verified based on a time series of automatically diagnosed foehn. As the occurrence of foehn is a binary event, a logistic regression was used to model the link between the predictors (DMO data) and the predictand (foehn). The best suited predictors are selected from different sets of possible predictors using automatic selection methods. Two different methods were tested, namely the Stepwise Selection and the Shrinkage Method (Lasso).
The Shrinkage Method outperforms the Stepwise Selection in relation to both the computational efficiency and the performance of the created forecasting models. The performance of the forecasting model for the next day (+39 h) decreases, yet still achieves a high percentage of successful forecasts (∼ 90 %). With the verification measures used, the models based on the global model ECMWF have shown to have a higher forecast skill than models based on the limited area model COSMO-2. The best possible model for predicting south foehn is based on ECMWF data and achieves a hit rate of 73 %, a false alarm rate of 1 % and an overall correct percentage of 97 % for a +15 h lead time, and a hit rate of 52 %, a false alarm rate of 2 % and an overall correct percentage of 95 % are obtained for a +39 h lead time. Similar scores are achieved for north foehn forecasts, although the occurrence of foehn is more frequent with 31 % in Piotta compared to 5.5 % in Altdorf. Including current observational data as additional predictors significantly improves the skill of the shorter range forecasts (+15 h).