Predictable weather regimes at the S2S time scale

State-of-the-art Subseasonal-to-Seasonal (S2S) forecast systems correctly simulate the main properties of weather regimes, like their spatial structures and their average frequencies. However, they are still unable to skillfully predict the observed frequencies of occurrence of weather regimes after the first ten days or so. Such a limitation severely restrict their application to develop climate service products, for example to forecast events with a strong impact on society, such as droughts, heat waves or cold spells.

This work describes two novel corrections that can be easily applied to any weather regime classification, to significantly enhance the S2S predictability of the frequencies of the weather regimes. The first one is based on the idea of weighting the daily observed anomaly fields of the variable used to cluster the atmospheric flow by the Anomaly Correlation Coefficient (ACC) of the same variable, just before clustering it. In this way, the clustering algorithm gives more importance to the areas where the forecast system is better in predicting the circulation variable. Thus, it is forced to generate the most predictable regimes. The second correction consists in the ACC weighting of the daily forecasted anomalies before the assignation of the daily fields to the observed regimes, to give more importance to the grid points where the forecast system has more skill. Hence, the forecasted time series of the regimes is more similar to the observed one.

Two sets of four regimes each were validated, one defined by k-means clustering of SLP from NCEP reanalysis over the Euro-Atlantic region during lasts 40-years (1979-2018) for October to March, and another for April to September. Forecasts proceed from the 2018 version of the Monthly Forecast System developed by the European Centre for Medium-Range Weather Forecasts (ECMWF-MFS). Predictability was measured in cross-validation by the Pearson correlations between the forecasted and observed weekly frequencies of occurrence of the regimes, for each of the 52 weekly start dates of the year separately and for a 20-years hindcast period (1998-2017).

Results show that with both corrections described above, Pearson correlations increase up to r = +0.5, depending on the start date and forecast time. Average increase over all start dates is of r = +0.2 at forecast days 12-18 and r = +0.3 at forecast days 19-25 and 26-32. The gain is spread quite evenly along the start dates of the year.

Beyond the Euro-Atlantic region, these two corrections can be easily transferred to any area of the world. They may be employed to correct seasonal predictions of weather regimes too (results in progress). Besides, their application is straightforward and provides a significant skill gain at a negligible computational cost for potentially all S2S forecast systems and regime classifications. We foresee that they might also benefit forecasts of atmospheric teleconnections. For all these reasons, we warmly recommend the S2S community to take advantage of this 'low-hanging fruit'.