Interactive comment on “ What are the prospects for seasonal prediction of the marine environment of the Northwest European shelf ? ” by Jonathan

Dear Referee 1, Thank you for your review, and sorry for the slow response. I wanted to post a reply before making the changes and resubmitting the paper. The main aim of the paper was to provide a roadmap for North West European shelf seas seasonal forecasting, with a marine management, marine policy and the marine user etc. audience in mind, as well as a general science readership. With this in mind, we wanted to take the reader through the various approaches in a clear manner. We


Background
The Northwest European Shelf seas (NWS) are of wide economic, environmental and political importance. They support many ecosystem services and human activities, including fisheries, energy extraction and transmission (both renewable and non-renewable), shipping and waste removal. Most of these services and activities are sensitive to the variable environmental 5 conditions under which they operate, for example:  Shipping (transport and industrial) and offshore oil/gas and renewable operations are sensitive to wind/wave conditions and currents;  The capacity of sea-floor gas distribution networks is sensitive to bottom temperature (with capacity decreased in cold conditions when demand is likely to be highest; 10

Scope of this study
A number of scientific and technical developments have recently come together to enable us, for the first time, to assess the potential for seasonal forecasts for the NWS: first, the demonstration of skill of global seasonal forecasting systems in predicting key European climate indices several months ahead (e.g. Scaife et al. 2014); secondly the development of regional oceanographic models of the NWS, proven for use in operational prediction (O'Dea et al. 2012); and thirdly the combination 5 of the regional NWS models with historical observations, to produce a consistent estimate of the time-varying state of the NWS over recent decades ('regional NWS reanalysis', Wakelin et al. 2014). The NWS regional model allows us to investigate the prerequisites of the dynamical downscaling approach (C), while the reanalysis allows us to investigate in detail the mechanisms of year-to-year variability, and so evaluate which elements of the NWS state are likely to be predictable. For the mean climate, the dynamical downscaling approach (C) has been shown to add value to the simple 10 approach (A) of reading off variables from the underlying global climate model (e.g. Mathis et al. 2013).
In this study we use the above building blocks to evaluate the potential of seasonal predictability for the NWS. We address the following questions in turn: (Pathfinder) and (A)ATSR (ESA), although some AMSRE are used during the GHRSST period), sea-surface height (altimetry from AVISOv3 along track), sea ice concentration (OSI-SAF), and water column structure (ARGO floats).
To make seasonal forecasts, a set of GloSea5 simulations are run to form a forecast ensemble and re-forecast (hindcast) ensemble. Every day, 2 ensemble members are initialised and run forward for 216 days. The previous 3-weeks are combined into a 42-member ensemble to make a 6-month forecast, which is updated weekly. A re-forecast ensemble is also run every 25 week (with the same modelling system) to correct bias and drift in the forecasts. This includes 4 start dates for the relevant month, for each of the previous 23 years, run forward for 216 days. The forecast ensemble is used to predict how the following 6 months will compare to this climatology. The system is also run as a continuous reanalysis (the GloSea5 ocean and sea-ice global reanalysis) from 1990-2015 to provide initial conditions for the ocean component of the hindcast ensemble. The atmospheric initial conditions for the forecast are taken from the Met Office operational weather forecast 30 system, and the hindcast atmosphere is initialised from ERA-interim. GloSea5 shows improved year-to-year predictions of the major modes of variability compared to the previous system

Comparison between NWS Reanalysis and GloSea5
The GloSea5 ocean and sea-ice global reanalysis and CO5 NWS reanalysis differ in a number of ways. Both rely on the ocean model NEMO, but run with grids of different horizontal and vertical resolutions. The GloSea5 global reanalysis 5 system is a coupled global model system designed to capture the key components of the global climate system in order to make a seasonal forecast, having used data assimilation over a wide range of variables to constrain the model. Conversely, the CO5 NWS reanalysis is a regional reanalysis, where a higher resolution ocean model is (one-way) forced from ERA Interim atmosphere forcing, using SST-only data assimilation. Effectively, the CO5 NWS reanalysis has a higher resolution and better representation of the NWS physics, whereas the GloSea5 global reanalysis has global scope and assimilates a 10 wider range of observations. Both have high enough resolution to include an open Dover Strait allowing a route for Atlantic water into the southern North Sea which is important for simulating the local seasonal cycle of salinity.
The CO5 NWS reanalysis requires lateral boundary conditions (GloSea5 is a global model system and so doesn't need them) which are taken from FOAM before 1990 and from the GloSea5 Reanalysis thereafter. The change from FOAM to the GloSea5 Reanalysis led to a discontinuity that was important for variables such as sub-surface temperature and salinity in the 15 open ocean -these variables are not used in this study but may influence the NWS fields that are.
An important difference between the two simulations is that the CO5 NWS reanalysis is a shelf seas model that includes all the key shelf seas processes whereas the ORCA025, being a global model, neglects some important processes, including dynamic tides.
There are also key differences in the riverine forcings, and the treatment of the Baltic Sea. The CO5 NWS reanalysis uses 20 river forcings (with inter-annual variability) from the E-HYPE river model (which gives too much discharge), and treat the Baltic as an open boundary where T and S are relaxed to Swedish Meteorological and Hydrological Institute (SMHI) model forcings. GloSea5 uses a river climatology, and model's the Baltic explicitly (although at too coarse a resolution to accurately simulate the complex interaction between the Baltic and NWS shelf sea).

Observations 25
Much of the evaluation of the CO5 NWS reanalysis (Wakelin et al. 2014) focused on the mean state of the model. Here we are more interested in the modelled temporal variability, and so undertake additional evaluation. We compare the model to limited observed time-series to assess its performance at replicating several observed events. Here we describe the observed time-series.

Southern North Sea Ferry data
Ferries are well established vessels of opportunity for oceanographic measurements taking regular long-term samples of surface water while the ship is on passage between ports (Bean et al. 2017). Observations can be in the form of samples taken by crew for subsequent testing in a laboratory or more sophisticated "Ferry boxes" as packages of instruments that semi-autonomously monitor temperature, salinity and other water properties. We use the monthly salinity data from the ferry 5 on the Harwich to Hook of Holland route, which took quasi-weekly temperature and salinity samples at 9 standard stations between 1971 and 2012 (Joyce 2006) and is reported in the ICES Report on Ocean Climate (Larsen et al. 2016) andMCCIP Report Cards (Dye et al. 2013). We use point time-series from this dataset to compare to the model.

Western Channel Observatory (WCO)
The Western Channel Observatory (WCO) is an oceanographic time-series in the Western English Channel (Smyth et al. 10 2015). In situ measurements are undertaken fortnightly at open shelf station E1 (50.03˚N, 4.37˚W) using the research vessels of the Plymouth Marine Laboratory and the Marine Biological Association. We compare time-series of temperature and salinity from a range of observed depths to model output from the nearest grid box.

NAO
The North Atlantic Oscillation (NAO) is a climatic phenomenon in the North Atlantic Ocean of fluctuations in the difference 15 of atmospheric pressure at sea level between the Icelandic low and the Azores high. These fluctuations control the strength and direction of westerly winds and storm tracks across Europe (Hurrell 1995

Storm track latitude index
When analysing the relationships of the shelf SSS we find correlation patterns which suggest storm track latitude may be important. We therefore analysed the mean sea level pressure data to produce a Storm track latitude index, following a method adapted from Lowe et al. (2009). The 3-hourly ERA Interim mean sea level pressure data from all modelled latitudes 25 at 2°30'E were filtered with a Blackman band pass filter. The temporal variance of this filtered mean sea-level pressure was calculated for each month for each grid box at 2°30'E, and the latitude with the greatest variance was recorded as the storm track latitude. We consider the winter (DJF) mean storm track.

Analysis techniques
We use regional mean time series of model output (surface, bed and surface minus bed, temperature and salinity, SST, NBT, DFT, SSS, NBS, DFS respectively) from the reanalysis, adapted from the region mask from Wakelin et al. (2012) (Figure   1a). We calculated monthly, seasonal and annual means from these time series, as well as directly from the model. In addition to model output, we extract regional mean time series for the ERAI surface forcings, the GloSea5 ocean lateral 5 boundary conditions (using the masks in Figure 1b-d), and the E-HYPE river forcings.

Identifying relationships between the NWS response and the drivers.
We compare time-series of the shelf response to time-series of the drivers to identify important relationships. We note that a statistically significant correlation does not imply a causal relationship, and so we interpret the spatial patterns of the correlation coefficients to help interpret the underlying mechanisms behind the correlations. We consider it beyond the scope 10 of this study to undertake sensitivity studies to explore any mechanisms in detail.
We use the region mask of Wakelin et al. (2012) (Figure 1a), to create regional mean time series of the results from the NWS Reanalysis, including its atmospheric (downward radiative fluxes, surface air temperature and relative humidity, total precipitation, mean sea level pressure, winds) and riverine forcings. We average the oceanic T and S forcings around the boundary into 21 regions based on horizontal and vertical gradients to T and S (typically dividing the north, west and south-15 western boundaries into surface, mid-depth and deep layers (according to the typical modelled summer and winter mixed layer depths), and then divide the boundaries horizontally according to features within the data. A deep layer of salty, relatively warm water in the south-western and the southern part of the western boundary is identified as Mediterranean Intermediate Water and is treated separately. Most correlations have been found with the surface (0 -30 m) regions, and so this study focuses on these regions. These regions are shown in Figure 1b-d. We also use annual and monthly mean time-20 series of the NAO and Storm Track latitude.
Model and observed time-series are compared to one another with Pearson's correlations, and their significance is noted at the 95% level. Typically, we compare the annual mean time-series, but we also compare at the seasonal time-scale. We also investigate lag-correlations between the shelf response, and possible drivers and climate indices. For example, the time series of DJF NAO will be correlated with the DJF SST across the shelf (at 0-months lag). The DJF will then be compared to the 25 JFM SST (January-March; 1-month lag), FMA (February-April; 2-month lag) etc. (e.g. Figure 5).
For consistency we have used the same region mask (e.g. Wakelin et al. 2012) for the river forcings as for the shelf seas variables and surface atmospheric forcings. However, this mask was not designed for rivers and so several regions must be treated with care, or excluded. For example, the northern North Sea region combines the river flow from small sections of the Scottish and Danish coasts which does not make senseother regions to be excluded are the central North Sea, Shetland 30 shelf region, and the North Atlantic regions. Other regions combine river flow from different coasts, but in a more sensible mannerfor example the English Channel and Irish Sea regions combine river input from two coasts, but due to the smaller enclosed nature of these regions, this is sensible in terms of local salinity. In the modelling system, the rivers do not have a Ocean Sci. Discuss., https://doi.org/10.5194/os-2018-1 Manuscript under review for journal Ocean Sci. Discussion started: 22 January 2018 c Author(s) 2018. CC BY 4.0 License. specified temperature, and so assume the local temperature when they reach the sea. Therefore, rivers predominantly affect salinity, with only secondary temperature effects (associated with changes in density driven circulation and stability). As increased river flow reduces the local salinity, most correlations are expected to be negative.

Question 1: How well does the GloSea5 global seasonal prediction system represent the NWS? 5
Both GloSea5 and the CO5 NWS reanalysis assimilate similar SST observations, and so their simulated patterns of SST are relatively similar, and in agreement with observations. When looking at the near-bed temperatures (NBT) and the difference between the surface and bed temperatures (DFT), there are important differences between GloSea5 and the NWS Reanalysis.
DFT is an important diagnostic of stratificationwhen DFT > 0.5 °C the water column is considered stratified, and the DFT = 0.5 °C isotherm is indicative of the location of the modelled tidal mixing fronts. in GloSea5 more grid boxes are stratified in the summer than in the NWS Reanalysis: from April to September, the stratified 15 area of the shelf is ~20% more in GloSea5 compared to NWS Reanalysis. Figure 2b shows a map of the stratified regions for May (an exemplar stratified month). This shows that much of the southern North Sea, English Channel and Irish Sea that is modelled as being mixed in May in the NWS Reanalysis, is stratified in GloSea5. This highlights an important weakness in using the GloSea5 system to provide direct information on the NWS.
Further evidence is shown under Question 2 below that using GloSea5 NWS fields directly would be problematic. We 20 therefore conclude that this approach (Approach A) is not viable, and some form of downscaling of the GloSea5 fields would be needed to generate reliable NWS forecasts.

Question 2: How well does the CO5 NWS reanalysis represent inter-annual variability on the NWS?
While the ability of the CO5 NWS reanalysis to simulate the mean state of the NWS is thoroughly evaluated (O'Dea et al. 2012;Wakelin et al. 2014), its ability to simulate inter-annual variability has received less attention. Evaluation requires long 25 observed time-series, preferably of variables that are not assimilated into the reanalysis. Here we focus on two locations, the Harwich to Hook of Holland route in the Southern North Sea and the WCO (see Methods).
First, we compare the observed time-series of surface salinity in the southern North Sea to that from GloSea5 and the CO5 NWS reanalysis (Figure 3). The time-series exhibits multi-year oscillations and these are well simulated by the NWS Reanalysis (r = 0.89, p = 0.00), despite the fact that it does not assimilate salinity observations. There is a fresh bias in the 30 model (-0.20) and a slightly greater variation (standard deviation ratio of 1.11). GloSea5 does not capture a realistic multi annual variability (r = 0.2, p = 0.00 with standard deviation ratio of 1.81) and modelled salinity also shows a large fresh bias that increases due to a substantial salinity drift over the duration of the time-seriesfurther evidence that direct reading from NWS fields from GloSea5 would be problematic. As there are differences in the river forcings between GloSea5 and the CO5 NWS reanalysis we would expect differences in the modelled salinity. The CO5 NWS reanalysis uses E-HYPE river forcing (Donnelly et al. 2013) which are specified daily whereas GloSea5 uses a river climatology (Dai and Trenberth 2002;5 Bourdalle-Badie and Treguier 2006), and so exhibits no inter-annual variability.
Secondly, we compare the observed WCO temperature and salinity profiles to the nearest daily mean CO5 NWS reanalysis grid box (Figure 4). The WCO observations are not assimilated into the NWS Reanalysis, but the SST from complementary satellite products is. Unsurprisingly, the CO5 NWS reanalysis SST is in close agreement with the WCO observations, for both the seasonal cycle, and the year to year variations at the surface and at depth (30 m). The inter-annual variability is well 10 captured in SST for all seasons (r > 0.99, p = 0.000), and for most seasons at 30 m (typically r > 0.9, but September -November r = 0.59). There is little seasonal cycle, trend or inter-annual variability in the WCO salinity (e.g. compare the inter-annual variability in Figure 3 and Figure 4). Given the lack of salinity seasonality, we compare all the WCO-reanalysis data pairs, which has a significant correlation of r = 0.49 at the surface and r = 0.65 at 30 m (p = 0.000 for both cases).
Further evaluation of the CO5 NWS reanalysis against other long time-series is planned through the Copernicus NWS 15 regional Marine Forecasting Centre. These initial results suggest that the reanalysis can provide valuable information on inter-annual variability on the NWS, where there is a strong signal.
We now investigate empirical forecasts based on the response of the CMEMS reanalysis to the observed NAO, and then applied to the GloSea5 forecast NAO.

Question 3: Can predictable climate indices provide real predictive skill for the NWS? 20
In the literature, there are many empirical relationships between climate indices and various physical and biological responses. The CMEMS reanalysis (through data assimilation) combines observations with models to give the best possible state estimate of the NWS, and so provides a powerful tool to develop such relationships. We focus on the winter NAO, as it is an important source of year-to-year variability in the NWS, and GloSea5 has predictive skill for the winter NAO. By investigating relationships between the CMEMS reanalysis fields and the observed (NOAA) NAO, and then considering 25 how these relationships change when we use the GloSea5 forecast NAO, we can explore the empirical approach to NWS seasonal forecasting.
First we consider the correlations (and lagged correlations) between the winter (DJF) NAO and the surface forcings we that think will be important for the shelf temperatures ( Figure 5). We find a positive correlation of the NAO with the DJF surface air temperature and humidity (Figure 5a We now consider the surface forcings we think are likely to be important for shelf salinity ( Figure 6). The DJF NAO is strongly correlated with the DJF 10 m wind magnitude across the domain, and this persists into the third month (March-May) for the southern and central North Sea (Figure 6i-l). The correlation between winter (DJF) NAO acts in opposite ways 5 for winter (DJF) mean sea level and for total precipitation. The DJF mean sea-level pressure (total precipitation) is negatively (positively) correlated with DJF NAO in the northern regions and positively (negatively) correlated in the southern regions. These correlations persist for a few months in some regions (Figure 6a-d, e-h). River systems can give additional predictability by continuing to respond after the forcing, or can reduce predictability by having such long response times that they act as a low-pass filter. The river runoff forcings are highly correlated with the DJF NAO in the Norwegian 10 Trench (Figure 6m Having shown the correlations of the NAO with the important surface forcings, we now look directly at the relationship between the observed DJF NAO and the shelf response (Figure 7). We find a significant positive correlation between the winter NAO and the winter SST in the most southern and eastern shelf regions (Figure 7a). In most of these regions, the significance of these correlation persists for one month (Figure 7b), and in the English Channel and southern North Sea a second month (February-April (FMA) SST, Figure 7c). The NBT correlations also show significant correlations with the 20 NAO and having memory in some regions. The DJF NAO is also significantly correlated with the regional mean DJF SSS in the Skagerrak/Kattegat, which persists until FMA (Figure 7e-h).
The above results suggest that knowledge of the NAO index could provide some skill for important variables, at modest lead times of 1-2 months, even if the DJF NAO is only determined from observations (at the end of the December-February period). Because the GloSea5 system has skill in predicting the DJF NAO index from the previous November (Scaife et al. 25 2014), it is possible that the lead time could be increased by using the predicted rather than the observed NAO index. In of salinity anomalies. The difference in the SSS correlation patterns between the observed and GloSea5 NAO perhaps act as an error estimate to this approach, suggesting caution and further assessment is needed before relying on an empirical seasonal forecast for SSS. Overall, the results with the GloSea5 NAO suggest that real relationships exist between the forecast NAO and the observed NWS fields, and that further improvements in the seasonal NAO forecast would deliver higher levels of forecast skill and/or regional detail. 5 The correlations between the NAO and the shelf response describe how strong a linear relationship exists between the two.
Where there is significant skill (a significant correlation) this linear relationship can be used to predict shelf response from the NAO. Here we show an example for one of the stronger NAO/shelf response correlations: English Channel SST. We have forecast the temperature based on both the observed NAO and GloSea5 forecast, and plotted this against the modelled temperature in Figure 9 (both have been normalised) with zero lag and the following 4 monthly lagged seasons (forecasting 10 JFM, FMA etc. based on the DJF NAO). We also include the time series of the observed and GloSea5 DJF NAO for comparison. This figure shows how an NAO based seasonal forecast based on the relationships discussed in this paper would look.
While our example has focused on a region and variable with a relatively high correlation, the English Channel winter SST may have a direct application. For example, European Sea Bass (Dicentrarchus labrax) spawn between the southern North 15 Sea and the Celtic Sea, in February-April within the 9 °C isothermthis region expands in warmer years (Beraud et al.

2017). Sea bass is a high value fish that is exploited by commercial fisheries (ICES 2012) and is an important species for
recreational anglers. The English Channel SST forecast for JFM (Figure 9) would be directly applicable to Sea Bass spawning. Furthermore, as the GloSea5 based forecasts also have skill, these forecast may be made in November (when the GloSea5 DJF NAO forecasts are made), and so provide January -February forecasts as early as November, such results 20 could be used to inform precautionary management when needed.
Developing a range of empirical forecasts is a possible way of producing NWS seasonal-forecasts, especially if based on predictable climatic indices such as the NAO. Investigating the relationships between the climatic drivers and the shelf response within NWS climate control simulation (i.e. using a multi-century global climate model run with fixed climate forcings, to drive a multi-century NWS simulation (Tinker et al. in preparation)) may allow much subtler relationships to be 25 established than is possible with the relatively short modern observed period. However, there will always be limits to what is possible with empirical forecasts.

Question 4: What are the prospects for improving NWS seasonal forecasts?
With the maturity of seasonal forecasting systems, and shelf seas dynamic downscaling systems, it will not be long before seasonal forecast systems for the NWS, based on dynamical downscaling is technically possible. We can start to explore 30 whether such a system could have skill using the CMEMS reanalysis. For a dynamically downscaled seasonal forecast to work, the NWS variability should be strongly related to aspects of the boundary conditions. If there is only a weak relationship between the boundary and the interior of the shelf, any year-to-year variability modelled by GloSea5 will not Ocean Sci. Discuss., https://doi.org/10.5194/os-2018-1 Manuscript under review for journal Ocean Sci. Discussion started: 22 January 2018 c Author(s) 2018. CC BY 4.0 License. manifest in the NWS. The NWS is considered to be quasi-isolated from the North Atlantic ), however, it is a broad continental shelf sea and so interaction with the atmosphere is important. We now investigate the relationships between the state of the NWS in the CMEMS reanalysis and the boundary conditions that forced it, to see how much of the NWS variability is driven by the large-scale drivers, and how much is internally generated.
To do this, we consider the inter-annual correlations between the CO5 NWS reanalysis surface and open ocean boundary 5 forcings and the NWS temperature and salinity in the regions defined in Figure 1 (see Figure 10). Our interpretation of the correlations must be informed by physical insight, recognising that correlation does not imply causation. We examine processes influencing temperature and salinity separately, since temperature is likely to be strongly influenced by surface heat exchange, whereas advective processes may play a stronger role for salinity due to a lack of direct feedback from surface salinity onto the atmosphere. 10 NWS surface temperatures are significantly correlated with temperatures at the open ocean boundaries of most of the domain (the northern boundary west of 10° W being the exception; example shown in Figure 10a, b); however, this is likely to be due to common surface forcings acting on the open ocean and shelf seas temperatures. Shelf SST is strongly correlated with surface air temperature as expected with the use of CORE bulk formulae. This is very homogenous across the shelf, so that the annual mean surface air temperature over the central North Sea is significantly correlated with the SST in all shelf 15 regions (Figure 10g). This is also true of the humidity (Q2, Figure 10h) and the downward component of the thermal radiation (STRD, Figure 10i). The incoming solar radiation (SSRD) has smaller spatial scales, consistent with the synoptic spatial scale of the atmosphere (Figure 10m prescribed. Overall, variability in SST appears to be linked to large-scale drivers (which may be predictable), with some 25 contribution from less-predictable synoptic scale variability at the regional scale.
The salinity on the NWS is primarily a balance between salty water entering the shelf from the North Atlantic, and its modification due to water exchanges with the atmosphere (e.g. precipitation) and dilution from rivers (and exchange with the Baltic, which is not considered in this study). This leads to some intuitive relationships: the saltier the Atlantic is, the saltier the NWS; the greater the river flow and rainfall into the NWS, the fresher the NWS is. This can be considered the direct 30 mechanism that controls salinity. An important secondary mechanism is the rate at which the NWS water is exchanged with the Atlantic, which we consider later. correlate with the NAO (Winther and Johannessen 2006;Marsh et al. 2017;Sheehan et al. 2017). This positive correlation means more (relatively) high salinity is advected onto the NWS under NAO positive condition, which would tend to increase the NWS salinity. We do not explicitly look at the European Slope Current in this study, but understanding its predictability could provide an important mechanism for NWS salinity predictability.
Because the strength of the inflow of North Atlantic water onto the shelf (the indirect mechanism affecting the year-to-year 5 shelf salinity) and the direct mechanism (dilution from rivers and precipitation) are both correlated to NAO, but in an opposing manner, it is possible more information (than is contained within the simple NAO index) is needed to predict NWS salinity variability. Such a balance of opposing mechanisms may explain the relatively low NAO SSS correlations (Figure 7 i-l) despite the main drivers of salinity being correlated to NAO. This provides an example where dynamic downscaling may improve the predictability over an empirical NAO-based forecast. 10 The spatial patterns in the relationships in Figure 10 show that (some of) the variability on the NWS is strongly coupled to the large-scale boundary conditionsthis is consistent with the established view that the NWS is a boundary driven system (with the caveat that we are using a non-eddy permitting model for the NWS). Hence the overall concept of (empirical or dynamic) down-scaling based on large-scale boundary drivers that may be predictable by a global seasonal forecast system is plausible. This suggests that the key area for scientific effort is to evaluate and improve the predictability of the NWS 15 boundary drivers as produced by GloSea5, rather than needing a heavy investment in ensembles of the regional model.
Overall, we conclude that much of the year-to-year NWS variability is relatively tightly linked to the variability in the boundary conditions, which is a prerequisite for dynamic downscaling. We note that in some cases, there may be a balance of opposing mechanisms that may require more information that is encapsulated in the simple NAO Index -this may provide a pathway for additional predictability from dynamic downscaling when compared to empirical downscaling. Furthermore, 20 much of the temperature and salinity variability on the NWS is linked to large scale climate variations (including river outflow which integrates rainfall over a large area) rather than to more local effects (such as the direct effect of rainfall on the synoptic scale). This increases the prospect of useful seasonal predictions since global seasonal forecast systems are beginning to show significant skill in predicting large scale climate indices.

Conclusions and Prospects 25
Our preliminary investigation shows that despite the useful skill that GloSea5 has in predicting certain large scale climate indices, its output cannot be used directly for shelf seas seasonal forecasts because of limited resolution, missing shelf sea processes and simplified treatment of river runoff. However, we have shown evidence that many aspects of inter-annual variability on the NWS are driven by large scale variations in elements of the atmospheric, oceanic and riverine forcings that are closely linked to the winter NAO index, for which there is considerable predictive skill at a lead time of several months. 30 Indeed Figure 7 shows that a simple empirical downscaling approach driven by the forecast NAO index can provide significant skill in some variables/regions at a lead time of several months. Based on our results we can make a preliminary assessment of the three possible forecasting methods identified in the introduction: A. Read off the NWS marine variables directly from the underlying climate model. This is not feasible with current 5 generation seasonal forecast systems. Over the coming 5-10 years it is expected that such systems will move to higher ocean resolution, and may incorporate tidal processes and improved coupling with river hydrology (e.g. Holt et al. 2016). Hence this approach may become feasible in time, although the resolution of global climate models will remain coarse compared with what is achievable through regional models.
B. Empirical downscaling. Our results show that a significant level of skill can be achieved for a limited set of 10 variables at a few months' lead time. The very limited availability of long observed time-series on the NWS means that the empirical climate response functions will often need to be developed using reanalyses. While our evaluation of the CO5 NWS reanalysis has shown encouraging evidence of its ability to capture inter-annual variability, regional seas reanalyses remain in a relatively early stage of development and evaluation.
C. Dynamical downscaling. Our analysis of boundary drivers of NWS variability shows that in some cases there is not 15 a high correlation of the key drivers with the winter NAO index. There could be more information in potentially predictable drivers than is contained in the simple NAO index. For example, we have shown evidence that the shelf salinity is affected by two opposing mechanisms associated with NAO, with the dominance of each mechanism being controlled by a different factor. This suggests that two different years with the same NAO index may have anomalously high or low salinity (driven by the NAO) on the NWS depending on something not captured by the 20 NAO Indexin such a case there is potential for improved predictability from dynamical downscaling. The strongly boundary forced nature of the NWS circulation suggests that the downscaling step may be fairly tightly constrained by the boundary conditions. This would imply that uncertainties due to, for example, the initial conditions of the NWS may be small, allowing the available computing time to be concentrated in getting the most reliable possible predictions of the boundary conditions from the global seasonal forecast system, rather than 25 requiring large ensembles of the shelf model. Demonstrating a dynamically downscaled forecast is beyond the scope of the current paper but is planned for a future study.
Even the limited level of predictive skill we have shown here for some regions of the NWS may be useful for certain applications, e.g. SST forecasts for February-April may give early indications of increased risk of harmful algal blooms, and predictions of near bottom temperature and its impacts on the gas supply network may inform more resilient energy planning. 30 Further developments will be needed to deliver a seasonal prediction system for the NWS with sufficient skill and reliability to inform user planning decisions over a wide range of applications. Specific research priorities are:  In-depth assessment of regional reanalyses as a tool to develop empirical downscaling relationships  Assess potential added value of dynamical downscaling, initially through case studies  Identify the largest sources of uncertainty in downscaled predictions (e.g. seasonal forecast fields for specific drivers, downscaling model), to inform where to focus development effort  Assess predictability in seasons other than winter. To date seasonal forecasting systems have shown less skill in the 5 summer, but this is an active research area (e.g. Hall et al. 2017). It may also be possible to demonstrate some degree of memory in the shelf seas themselves, which would add to forecast lead times.
Many challenges remain before we can derive seasonal forecasts for the NWS that are accurate and reliable for a wide range of regions and variables. It is likely that some variables will prove to be inherently unpredictable to any useful degree. But the early results presented here show that current seasonal forecast systems can already provide meaningful information with 10 the potential for applications in marine operations and planning. As our understanding and capability develops, a close interaction between climate scientists, marine scientists and end users will be needed to bring the added value of seasonal forecast information into decision making in marine policy, planning and operations.

Data availability
The data used in this study is available from the Copernicus Marine Environment Monitoring Service (CMEMS) Northwest 15 European Shelf reanalysis (NORTHWESTSHELF_REANALYSIS_PHYS_004_009) available from their online catalogue.

Author contribution
JT designed the analysis, which was undertaken by JK and JT. JT and RW prepared the manuscript with contributions from all co-authors. RB, SD provided advice on policy and user relevance.

Competing Interests 20
The authors declare they have no conflict of interest.