The forecasts of WUSHP combine three elements, namely meteorological forcing from ECMWF's Seasonal Forecast System 4 (Molteni et al., 2011), bias correction of the meteorological forcing with the quantile mapping method of Themeßl et al. (2011) and simulations with the VIC hydrological model (Liang at al., 1994). The skill of the system was assessed with hindcasts. These cover the period 1981–2010, were initialised on the first day of each month and extend to a lead time of 7 months. The system is probabilistic (15 members), so each set of hindcasts consists of a total of 5400 runs (30 years × 12 months × 15 members). In addition a single reference simulation was performed in which VIC was run with a gridded data set of model-assimilated meteorological observations, namely the WATCH Forcing Data Era-Interim (WFDEI; Weedon et al., 2014). The reference simulation has a dual aim. The first aim is to create initialisation states for the hindcasts. Secondly, the output of the reference simulation, e.g. run-off, is used for verification of the hindcasts. This output will be named “pseudo-observations” here.

Due to the set-up of the routing module of VIC, the state of discharge could not be saved and loaded. Hence for spin-up discharge, each 7-month hindcast was preceded by a 1-month simulation with WFDEI forcing, which in turn was initialised with the model states generated in the reference simulation and zero discharge. All hindcasts and simulations were performed on a 0.5∘×0.5∘ grid in natural flow mode, i.e. river regulation, irrigation and other anthropogenic influences were not considered. VIC is run with a time step of 3 h. More details about the set-up of the system and the hindcasts can be found in the companion paper (Greuell et al., 2018).

In this paper we analyse hindcasts of run-off, discharge and evapotranspiration. Run-off is defined as the amount of water leaving the model soil either, along the surface or at the bottom, while we define discharge as the flow of water through the largest river in each grid cell.

Discrimination skill (briefly skill from now on) is measured in terms of the correlation coefficient between the median of the hindcasts and the observations (or pseudo-observations; R). We will designate R values as significant for p values less than 0.05. We also considered metrics designed for the evaluation of categorical forecasts (terciles), namely the relative operating characteristics (ROC) area and the ranked probability skill score (RPSS). The thresholds used for assigning individual (pseudo-)observations to terciles were determined from the (pseudo-)observations themselves. Similarly hindcasts were assigned to terciles by reference to themselves. Due to this strategy metrics are unaffected by biases, a desired property (see Sect. 1). In the companion paper skills in terms of the considered metrics were compared, and it was found that for all combinations of target and lead months the skill patterns in the maps were similar to a high degree. For that reason, we selected only one of them (R) for this paper.

Unless mentioned otherwise, prediction skill of the hydrological variables is determined against the pseudo-observations (see Sect. 2.1). These have the advantages of being complete in the spatial and the temporal domain and of being available for all model variables. We will refer to this type of skill as “theoretical skill”. In the companion paper theoretical skill for discharge was compared to “actual skill”, which is the skill assessed with real observations. For the determination of the skill of the meteorological forcing we used the WFDEI data.

To investigate the possible contribution of trends to skill, skill in the meteorological forcing and in run-off was determined both before and after removing the trend from both the (pseudo-) observations and the hindcasts. Data were detrended by first constructing time series (1981–2010) for each variable, target month, lead month and grid cell (30 values). We then removed the trend from each time series by first fitting a least-squares regression line to the time series and then subtracting the time series corresponding to the line from the original data. For the hindcasts, time series were constructed for the mean of the ensembles, and the resulting best fit was subtracted from each member individually.

Like in the companion paper, skill was analysed on a monthly and not on a seasonal basis with the aim of achieving a relatively high temporal resolution in the skill analysis. Attention was confined to consistent skill, which we define as skill that persists during at least two consecutive target or lead months. In accordance with Hagedorn et al. (2005), we designated the first month of the hindcasts as lead month zero.

In most result sections, we will first analyse and explain skill at the level of the entire domain. We will then take out the most noteworthy details of the summary plots and seek an explanation for them.

As already pointed out in the introduction, a number of specific hindcasts were carried out with the aim of isolating the contributions of different sources to skill. The FullSHs, in which skill is due to both meteorological forcing and initial conditions, constitute the starting point. The specific hindcasts can be seen as restricted, in the sense of limiting the types of sources of skill, versions of the FullSHs. The following five sets of specific hindcasts, each consisting of 5400 computer runs, were produced:

The InitSHs isolate the skill due to both types of initial conditions considered here (soil moisture and snow). Like in the FullSHs, the annually varying initial conditions are taken from the reference simulation, while for each year the meteorological forcing is identical and consists of an ensemble of 15 S4 hindcasts. More specifically, we selected member 1 from the 1981 hindcasts, member 2 from the 1983 hindcasts, etc. By using identical meteorological forcing for all of the years of the hindcasts, skill in hydro-meteorological variables due to skill in the forcing is eliminated.

Forcings and initial conditions of all of these hindcasts differ among the calendar months so that the annual cycle is conserved. Hence, in the list above, the following apply:

“Identical for all years” means that the forcings (or the initial conditions) for all hindcasts starting in, for example, May are identical.

Thus, like the FullSHs, all specific hindcasts for a single starting date consist of 15 members, which is important, since ensemble size affects skill metrics (Richardson, 2001). Also, in all hindcasts the probabilistic character is exclusively due to the 15 members of the meteorological forcing, while initial conditions are deterministic. This consistency is important, since the main aim of the various specific hindcasts is to compare them with each other. A disadvantage of the small ensemble size is the sampling uncertainty (see Sect. 4.2 of the companion paper).

Discharge initialisation, a potential source of skill, is not considered. This has no effect on most of the analyses of the paper, since these are made in terms of run-off. Where discharge is analysed the effect of discharge initialisation is, due to the limited residence time of water in the rivers, restricted to the first lead month of the hindcasts (see Yuan, 2016).

In this sub-section, the skill of the meteorological forcing will be analysed. Attention will be limited to the three input variables of VIC that have the largest effect on run-off and evapotranspiration, namely precipitation, 2 m temperature and incoming short-wave radiation. The WFDEI data are used as a reference. Here the data after bias correction are considered. In Appendix B we will discuss the skill of the raw S4 data, which is the meteorological forcing before bias correction. Differences in skill between the bias-corrected and the uncorrected data are negligible for temperature and short-wave radiation and small for precipitation.

Figure 1 shows results of the skill analysis of the precipitation forcing. Figure 1a provides an example of the skill for a single target and lead month (January as lead month 0). A summary of the skill in the precipitation hindcasts is given in Fig. 1b, which plots the fraction of all cells within the domain with statistically significant R values, so Fig. 1a condenses into a single point in Fig. 1b. During the entire year, there is considerable skill for lead month 0 (on average in 61 % of the domain), but skill declines very rapidly to 6 % for lead months 1 and 2, just 1 % more than the percentage of cells in the case of no true skill at all. Hence, from lead month 1 on, skill is almost negligible. Regarding lead month 0, there is more skill in January, February and March than during the other months. For the lead month 0, hotspots of consistent skill, i.e. with a duration of significant skill of at least 3 target months, are situated on the Iberian Peninsula from November to March, in western Norway from January to April, in Greece and western Turkey from December to February, and in Scotland from December to March. All these occurrences of consistent skill are restricted to the winter half of the year and mostly to coastal regions (see Fig. 1a), suggesting them to be linked to the initial state of the sea surface temperature.

Figure 2 shows important aspects of skill in the 2 m temperature hindcasts. One aspect is the possible contribution of a 30-year trend, which could be related to greenhouse warming, to the skill. Figure 2a and b provide summaries of the skill of the un-detrended and the detrended data, respectively, whereas Fig. 2c compares these two types of data. For lead month 0, the hindcasts have significant skill in the largest part of the domain (Fig. 2a and b), and detrending has a small effect (Fig. 2c). At longer lead times, the percentage of cells with significant skill quickly drops towards the theoretical no-skill limit (5 %), but there are a few exceptions, namely the following:

For lead month 1, February and March temperatures are predicted with significant skill in a considerable part of the domain (44 % in February; 53 % in March). In both months the region with skill is more or less contiguous and comprises the Russian part of the domain, the Ukraine and the regions bordering the southern part of the Baltic Sea (Fig. 2d and e). In February the region of skill extends towards central Europe. In March it also comprises northern Fennoscandia. This skill hardly diminishes by detrending the data (Fig. 2b and c), suggesting that the skill is not related to climate change. Indeed, in February and March the observed trend (in the WFDEI data set) is insignificant across most of the domain (11 % of the domain in February and 18 % in March), and, more importantly here, it is insignificant in the regions with significant skill in the temperature hindcasts (Fig. 2g demonstrates this for March). We conclude that the temperature skill in February and March as lead month 1 must be due to initial conditions of the climate model (see also the discussion on Fig. 10).

Since short-wave incoming radiation is important for evapotranspiration, we finalise this sub-section with a short analysis of its predictability (Fig. 3). In terms of R, skill is considerable during the first lead month, with 58 % of the cells having significant skill, on average over the year. Months from March to September tend to have more skill than the other months of the year. Beyond lead month 0, skill settles around the no-skill line, except from April to July, but the fraction of cells with significant skill never exceeds 21 % (in May as lead month 1). Trends in the data hardly affect skill (Fig. 3b).

In these sub-section analyses the effects of the meteorological forcing and the initial conditions on the predictability of run-off and discharge (discharge is only considered in Fig. 4) are isolated. We first address the question of how much of the skill in the run-off hindcasts is linked to trends. To examine this question, the pseudo-observations and the hindcasts of run-off were detrended, and the skill was compared to that of the un-detrended data sets. We found that for lead month 2 and averaged over all months of the year, the fraction of cells with a significant R decreased from 58.7 % to 57.4 % due to detrending, a difference of 1.3 %. This difference is much smaller than the decrease for temperature (11.8 %). We conclude that trends contribute very little to skill in run-off. All analyses of this sub-section hereafter pertain to un-detrended data.

This section analyses skill in the hindcasts of evapotranspiration, because hindcasts of evapotranspiration are useful in themselves and evapotranspiration affects run-off (see Sect. 1), and in order to demonstrate the rich possibilities of the pseudo-observations, the specific hindcasts and the detrending to unravel the various sources of skill. In VIC, evapotranspiration is computed with the Penman–Monteith method (see Shuttleworth, 1993).

Figure 9a summarises skill in evapotranspiration in the FullSHs. Levels of predictability are higher than for precipitation (Fig. 1), similar to those for temperature (Fig. 2) and lower than those for run-off (Fig. 4a). Figure 9b isolates the diverse contributions to skill for lead months 0 and 2 by showing the skill for the FullSHs and three specific hindcasts. Averaged over the year, meteorological forcing (MeteoSHs) contributes more to predictability in evapotranspiration than the initial conditions, among which soil moisture (SMInitSHs) causes more skill than snow (SnInitSHs). Hence, comparing skill in run-off with skill in evapotranspiration, the most important source of skill shifts from the initial conditions of soil moisture to meteorological forcing.

In the FullSHs (Fig. 9b) and focusing on lead month 2, there is hardly any skill in the evaporation hindcasts from November to March (9 % of the domain, on average over these months), with the exception of January (18 %), when the region of skill (Germany and Benelux) is part of a larger region of skill in the temperature hindcasts for the same target and lead month. We blame the winter minimum of skill in evapotranspiration to the low levels of evapotranspiration and the low levels of skill in the temperature forecasts for the same period. The next month (April) exhibits the highest level of skill of all months (44 % of the domain), which is mainly due to meteorological forcing and has smaller contributions by the initial conditions of soil moisture and snow. From May to September there is some significant skill (23 % of the domain, on average over these months). Whereas in May forcing is still the most important contributor to skill, initial conditions of soil moisture form the main contributor from June to October. We speculate that this shift in the order of importance between forcing and soil moisture is due to the amount of variability in soil moisture. In Europe in spring (April and May), soil moisture variations are relatively small and hence hardly contribute to variations in evapotranspiration. Later in the year (June to September), soil moisture is often available in limited amounts, so variations are larger and hence contribute more to variations in evapotranspiration. Snow initial conditions contribute to skill only during the snowmelt season from April to July.

The contribution of trends to predictability of evapotranspiration is summarised in Fig. 9c for lead months 0, 1 and 2. For lead month 2 and averaged over all target months of the year, detrending leads to a decrease in the fraction of cells with a significant R, from 17.6 % to 13.8 %, a difference of 3.8 %. The contribution of trends to skill in evapotranspiration is less than its contribution to skill in temperature (a difference of 11.8 %) but larger than its contribution to skill in run-off (a difference of 1.3 %). Trends contribute to skill in evapotranspiration during the part of the year when they also contribute to skill in atmospheric temperature (Fig. 2c), namely from April to September and in November (for lead month 0). However, whereas during the 3 summer months the skill in the temperature hindcasts is almost exclusively linked to climate change, a considerable part of the domain still exhibits skill in evapotranspiration after detrending.

To provide a deeper understanding of the skill in evapotranspiration, the skill in April and July is analysed in some detail. Figure 10 deals with April as lead month 2, showing the skill in evapotranspiration from the FullSHs in Fig. 10a and from the MeteoSHs in Fig. 10b. Regions of skill, mainly a piece of land from southern Fennoscandia to the Black Sea, are the same in the FullSHs and in the MeteoSHs, though skill is somewhat degraded in the MeteoSHs. This indicates that meteorological forcing causes most, though not all, of the skill. Indeed, Figs. 2e (March) and 10c (April) show that the temperature forecasts for these two months after initialisation on 1 February contain skill in the mentioned region. We conclude that much of the skill in evapotranspiration is due to skill in the temperature hindcasts. The remaining part of the skill is due to initial hydrological conditions. While Fig. 9b shows this for the entire domain, we also found limited amounts of skill in the SnInitSHs and the SMInitSHs for April in the stroke of land from southern Fennoscandia to the Black Sea (not shown here). This means that in that region, initial conditions of the hydrological model on 1 February provide some skill to the hindcasts of evapotranspiration for April. We like to note that this could be consistent with the conclusion in Sect. 3.1 that the skill in the temperature hindcasts of February and March in this same region are due to the initial conditions of the climate model. These initial conditions could be, for example, sea surface temperatures or also the local state of snow and/or soil conditions. In the latter case, the two types of predictability in the mentioned regions would have the same or a similar source. Initial conditions of snow and/or soil conditions in S4 would lead to skill in the temperature hindcasts of S4, while initial conditions of snow and soil moisture in VIC lead to skill in the evapotranspiration hindcasts of VIC.

During the summer months and for all lead times, skill in evapotranspiration occurs in two regions, namely the southern part of the Mediterranean and western and northern Norway. Figure 11 shows target month July for lead month 5 as an example. Whereas Fig. 11a is for the FullSHs, Fig. 11b–d depict the maps for three specific hindcasts (SnInitSHs, SMInitSHs and MeteoSHs), and Fig. 11e shows skill for the FullSHs after detrending. Since the SnInitSHs and the MeteoSHs exhibit hardly any skill, while the SMInitSHs have considerable skill in the Mediterranean (Fig. 11b–d), it can be concluded that the skill in this region is due to soil moisture initial conditions. So, in this particular case, knowledge of soil moisture conditions on 1 February still yields skill in evapotranspiration in July. This skill in the Mediterranean is not affected by detrending (compare Fig. 11a and e), so it does not have a climate-change component.

The skill in Norway has a more complicated origin. The three specific hindcasts show that it is due to a mix of initial snow conditions (Fig. 11c) and meteorological forcing (Fig. 11d). The effect of the initial snow conditions (on 1 February) can be understood with the help of the analysis of run-off skill in the SnInitSHs (Fig. 7), which led to the conclusion that run-off skill caused by snow initialisation occurs at the end of the melt season, which is July in much of Norway. Therefore, in this country and in July the timing of the disappearance of snow cover varies from year to year. This then has a considerable effect on evapotranspiration, since bare soil has, compared to snow, higher surface temperatures and hence more evapotranspiration in summer. The contribution to skill by forcing (Fig. 11d) fades with, but is not removed by, detrending (not shown here), so it has a part that is related to climate change and a part that is unrelated to climate change. The climate-change-related skill due to forcing resides in the temperature hindcasts, which have significant skill in this region at all lead times (Fig. 2f). The non-climate-change-related skill in the MeteoSHs for July is likely an indirect effect of the skill in the forcing (especially precipitation) during the first lead month (February). This leads to skill in snow water that is equivalent towards the end of February, which fades but has not disappeared completely on 1 July (Fig. 11f) and then causes skill in evapotranspiration at the end of the melt season.

A remarkable result of our work is the reduction of the skill in run-off beyond lead month 1, when annually varying S4 forcing is used (FullSHs) instead of meteorological forcing that is identical for all years (InitSHs; see Fig. 4). This result is counter-intuitive but is, as we discussed, a logical consequence of forcing with inter-annual variation that has no or insufficient skill, such as the S4 forcing. Other studies compared the FullSHs (also called climate-model-based hindcasts) with ESP hindcasts, which are slightly different from our InitSHs (see Sect. 4.3) but like the InitSHs have uninformative meteorological forcing for each year. Some of these studies (e.g. Singla et al., 2012; Mackay et al., 2015) found little overall difference in skill between the FullSHs and ESP hindcasts. However, in a study of Canadian catchments, Bazile et al. (2017) broadly confirm our finding that beyond the first lead month, ESP is superior to the FullSHs, while the reverse holds for the first lead month. Arnal et al. (2018) compared FullSHs with ESP hindcasts and found that, in Europe, ESP has more discrimination skill (“potential usefulness”) than FullSHs, although there are exceptions both spatially and seasonally. These authors, however, do not mention any trend with lead time in the difference between FullSHs and ESP. In contrast with our results, in the studies of Yuan et al. (2013), Thober et al. (2015), Yuan (2016) and Meißner et al. (2017), skill is enhanced when using meteorological hindcasts. In those studies this also happens at longer leads. This contrast might be explained by more skill in the meteorological hindcasts of the mentioned studies than in the present study, which could be due to the type of meteorological hindcasts (only Meißner et al., 2017, used S4) or the investigated region (in the mentioned studies of the US, Europe, China and Germany, respectively). Europe is a region with relatively little skill in meteorological hindcasts (Kim et al., 2012; Scaife et al., 2014; Baehr et al., 2015). Effects of regional differences in the skill of the forcing on the relative skill of FullSHs and ESP are mentioned by Wood et al. (2005), who reported that FullSHs for the western United States have practically no skill improvement over the ESP, except for some regions and seasons with predictability of the forcing originating in El Niño–Southern Oscillation teleconnections.

The specific hindcasts of this study show that in Europe initial conditions of soil moisture are the largest source of skill in the seasonal run-off forecasts produced with WUSHP. Contributions to skill by the initial conditions of snow and by the meteorological forcing are mostly much smaller. To our knowledge, two other studies analysed sources of skill of hydrological seasonal forecasts for Europe with dynamical systems similar to those of the present study, namely Bierkens and van Beek (2009) and Singla et al. (2012). Comparing our results with those of Bierkens and van Beek (2009), both studies agree that initial conditions form the dominant source of skill. However, compared to the present study, Bierkens and van Beek (2009) find a larger contribution to skill by the meteorological forcing, at least in summer. This difference might be due to the quality of the forcing. Bierkens and van Beek (2009) developed an analogue events method to select, on the basis of annual sea surface temperature (SST) anomalies in the North Atlantic, annual ERA40 meteorological forcings, which they used as forcing for their hydrological model. One might speculate that in Europe their semi-statistical forcing is more skilful than the S4 forcing used in WUSHP. This suggests that there is room for improvement of climate-model seasonal forecasts, so if and when this improvement is realised, the relative contribution of the meteorological forcing to skill in hydrological variables would increase. As for the second study of the sources of skill, conclusions of Singla et al. (2012) are not directly comparable with those of the present study, as they used ESP and reverse ESP (see Sect. 4.3).

The dominance of soil moisture initial conditions in terms of domain-lumped skill also extends to the hotspot regions and periods of skill (Table 1). The understanding of the skill linked to soil moisture can be deepened by another level, as in Shukla and Lettenmaier (2011). The underlying idea is that this type of skill increases with the inter-annual variability of soil moisture at the date of initialisation and that this skill is gradually eliminated during the course of the hindcasts by inter-annual variability in processes like rainfall and snowmelt. The question is the following: to what extent are the hotspots of skill (see Table 1) linked to soil moisture initialisation due to the cause of the skill, and to what extent they are due to a lack of inter-annual variability in the processes that eliminate the skill? Figure 12 helps answer this question for the skill found in the run-off hindcasts of August as lead month 2 with a simple method of analysis. Figure 12a shows the standard deviation of total modelled soil moisture (σSM) on the day of initialisation (1 June), taken from the reference simulation. Figure 12b depicts the standard deviation of total rainfall (σRF) during the course of the hindcast (June–August), taken from the WFDEI data set, which is the investigated skill-eliminating factor. These two quantities were combined into an estimate of the skill (Sest): Sest=exp⁡-σRF2σSM2. This estimate (Fig. 12c) needs to be compared with the skill of the hindcasts, mapped in Fig. 12d in terms of R. The two maps are not expected to be exactly equal, not only because of the simplicity of the estimation method but also because Sest is not a correlation coefficient. However, in the limits, Sest has the desired properties. It is equal to zero for the cases of constant initial amounts of soil moisture or infinite variability in rainfall. It is equal to 1 for the cases of infinite variability in soil moisture or constant rainfall. The correlation coefficient between the patterns in Fig. 12c and d is highly significant (0.67), and the hotspot regions of skill are the same in both panels, namely the northern part of Fennoscandia and the southern part of the Mediterranean. So, in the case of August as lead month 2 the estimation method is reasonably successful in computing the pattern of skill in the hindcasts with the simple means of the WFDEI data set and model calculations from the reference simulation. The merit of the estimation method is the deeper understanding of the cause of the skill in the two hotspot regions. Northern Fennoscandia is a hotspot, because the amount of inter-annual variability in initial soil moisture is larger than elsewhere (Fig. 12a). The southern part of the Mediterranean is a hotspot, because the amount of inter-annual variability in rainfall is lower than elsewhere (Fig. 12b).

This simple method of analysis helped to bring the understanding of the skill in northern Fennoscandia and the southern Mediterranean to a deeper level, but it was less successful for the other hotspots. A more thorough analysis along these lines and a deeper understanding of skill in the hindcasts is left for future work.

The specific hindcasts of this study are related to the well-known ESP (e.g. Wood and Lettenmaier, 2008; Shukla and Lettenmaier, 2011; Singla et al., 2012; Van Dijk et al., 2013; Harrigan et al., 2018). ESP is not only used as an experimental tool in science but is also widely used to produce forecasts in operational mode (Day, 1985). ESP used for scientific purposes can be subdivided into proper ESP (called ESP from now on) and reverse ESP.

ESP (hindcasts) is similar to the InitSHs of this study. In both types of hindcasts the initial conditions vary from year to year and are quasi-perfect, i.e. they are taken from a simulation like our reference simulation, while the meteorological forcing is uninformative, e.g. by being the same for all years (in the InitSHs and, for example, in the ESP of Shukla and Lettenmaier, 2011) or by varying randomly from year to year (e.g. in the ESP of Singla et al., 2012). This eliminates skill due to the meteorological forcing, so skill can only be due to the initial conditions. However, while in ESP the forcing is selected from historic observations, it is selected from the S4 hindcasts in InitSHs in order to retain an inter-member variability and other statistical characteristics of the time series similar to that in the FullSHs. An advantage of ESP is that its production is relatively cheap, because no climate-model forecasts are needed.

Similarly, reverse ESP (see Wood and Lettenmaier, 2008) resemble the MeteoSHs of this study. In both types of hindcasts the meteorological forcing varies from year to year, while the initial conditions are identical for each year. This eliminates skill due to the initial conditions, so skill can only be due to the forcing. However, while in reverse ESP the forcing of each year is made up of the observations of that year, it is made up of the S4 hindcasts in the MeteoSHs. Moreover, in reverse ESP, ensembles are built by using differing initial conditions, whereas they are built by using differing meteorological forcings in the MeteoSHs.

In ESP and in the InitSHs, if all skill due to the meteorological forcing is indeed removed, the remaining skill, which is due to the annually varying initial conditions, should logically be the same in both types of hindcasts, since the initial conditions are the same. To test this expectation we produced ESP and compared its skill with that of the InitSHs. Indeed, skill from these two types of hindcasts is almost identical, as demonstrated in the Supplement (Fig. S8). We conclude that skill produced with specific hindcasts with a forcing that does not vary from year to year is not sensitive to the choice of that forcing, perhaps with the exception of forcings that deviate strongly from being realistic. We like to note here that, in odd years, one of the ESP ensemble members is identical to the pseudo-observation used for verification. This is a concern, but we deemed this less important than the requirement of identical forcing for all years, which is crucial for the explanation of the skill reversal (Sect. 3.2.1).

This similarity of the InitSHs and ESP is in sharp contrast with the skill resulting from reverse ESP and MeteoSHs, which are expected to be totally different. Keeping in mind that in both types of hindcasts, skill is caused only by skill of the meteorological forcing, this is the skill of the S4 hindcasts in the MeteoSHs. The present study showed that in Europe there is a small contribution to skill in the run-off hindcasts by the forcing and that this contribution tends to decrease with time. This differs from reverse ESP, in which skill is small at the beginning and then increases with lead time to reach perfect skill at very long leads (see Wood and Lettenmaier, 2008), because the meteorological forcing is quasi-perfect (i.e. identical to the forcing in the reference simulation), while the influence of the initial conditions, which are non-informative in reverse ESP, decreases with time.

We plan to launch an operational version of WUSHP. That version might include a post-processing procedure with the aims of removing biases in discharge and making the system more reliable. This could perhaps be done with statistical calibration (e.g. Gneiting et al., 2005; Schepen et al., 2014), a technique that, contrary to quantile mapping, considers information that is available from correlations between hindcasts and observations (see Wood and Schaake, 2008; Madadgar et al., 2014).

The superiority of the InitSHs (and the ESP) with respect to the FullSHs for hindcasts beyond the first 2 lead months raises the question of whether one should, in an operational version of WUSHP and for these lead months, issue forecasts like the InitSHs (or ESP) and not forecasts like the FullSHs. The logical answer is “yes”, but such a strategy should then be reconsidered when the meteorological forcing is taken from a new, possibly improved version of the climate model or from another, possibly better type of climate model.

The applied methods of analysis are not suitable for giving quantitative advice on what would be the best investment for increasing the amount of skill of WUSHP. However, since initial soil moisture is the dominant source of predictability, a large gain of skill could possibly be made by assimilation of soil moisture observations into the modelled state of soil moisture (see e.g. Draper and Reichle, 2015). In addition, observations of a snow-water equivalent could be assimilated into the modelled state of snow (see e.g. Griessinger et al., 2016). Improving the calibration of VIC would be another obvious road towards improvement of the seasonal predictions discussed in this paper. This should lead to higher actual skill but not necessarily to more theoretical skill (see the discussion section of the companion paper).