Currently, different long-term forecast products are proposed and tested. They are based on ensemble forecasts of standardized precipitation indices
(e.g., SPI-3 to SPI-12) with lead times reaching from 1 to 7 months. Note that for lead times that are shorter than the accumulation periods, the SPIs are calculated using a mix of observed precipitation during the period before the starting date and the forecasts thereafter (Dutra et al. 2014). This method allows seamless forecasts of SPI with 1 to 12 months lead time.
Two information types are derived from the ensemble system. The first one is associated with the probabilistic forecast. It is defined as the number of members that are associated with an SPI lower than -1 (drought) or larger than 1 (flood), or normal conditions (between -1 and 1). If there is a consistency between the members (e.g., more than 50 percent) to forecast a specific anomaly or normal conditions, then the information is reported. If none of these three specific conditions emerged predominantly (i.e., large uncertainties of the members), the SPI forecasted is considered not significant and no conclusion can be provided. This method does not quantify the intensity of a drought but the consistency and the spread of the ensemble can be related to the uncertainties of the forecasts.
The second information type is the ensemble mean of the SPI. In case of consistency within the ensemble members, the mean values, provided for the same accumulation periods and lead times as previously discussed, are reported and provide information on the strength of a forecasted event. Because of the use of the ensemble mean, this method tends to underestimate the intensities of observed events but allows distinguishing abnormal from extreme abnormal (dry or wet) conditions.
The predictability of the forecasts at such long lead times is obviously a big challenge, and there are only few papers in the literature that deal with this kind of assessment. Dutra et al. (2013) have assessed the prediction capabilities of the ECMWF Seasonal S4 model for different basins in Africa. Even if their model presents better scores than the climatology, we expect a positive effect of the underlying data construction, with observations added for the shorter lead times that will influence the results. In Lavaysse et al. (2015), the scores of the SPI forecasted have been quantified using different ensemble systems over Europe. At a monthly time scale (e.g., for SPI-1 with 1-month lead time), it has been shown that 40 percent of all droughts (defined as an SPI lower than -1) are correctly forecasted 1 month in advance. This score may not seem large, but according to the challenge and the difficulty to forecast precipitation at those lead times (one month of cumulated precipitation) and given the strict definition of an event, this score is clearly above the climatology (16 percent) and provides a first state-of-the-art predictability of these events. To increase the predictability scores of droughts, some studies suggest working with variables that have more persistency, such as the soil moisture (Sheffield et al. 2014). But to be comparable, the score related to the climatology also should be provided.
Ongoing works at JRC are testing the use of atmospherical predictors to forecast extreme precipitation anomalies. For instance, in Lavaysse et al. (2016), it has been shown that the occurrence of weather regimes (WRs) that classify atmospherical circulation patterns in predetermined anomaly patterns could improve the forecast scores in Europe. Over certain regions, such as Scandinavia, the precipitation anomaly patterns are strongly connected to blocking situations. These situations are generally better forecasted in the atmospheric models than the precipitation anomalies. It has been shown that using a simple best-correlation attribution of WR occurrences and precipitation anomalies, the forecast of WRs could improve the probability of the detection of droughts (i.e., SPI lower than -1) by more than 20 percent. Indeed, over Scandinavia in winter, 65 percent of the events are correctly forecasted 1 month in advance (comparing to about 40 percent using the forecasted precipitation). Obviously, this region and this season are well known to be strongly connected to the synoptic circulation; elsewhere or during other seasons, the forecasted precipitation could be better than this alternative method. To provide the best forecast product possible, an assessment of each forecast method has been made over Europe, and these results allow identification of the most accurate forecast for each region and each season. According to these past evaluations, the best forecast products can be chosen per region or grid cell in order to provide more robust information.