Experiments with Seasonal Forecasts of ocean conditions for the Northern region of the California Current upwelling system

Resource managers at the state, federal, and tribal levels make decisions on a weekly to quarterly basis, and fishers operate on a similar timeframe. To determine the potential of a support tool for these efforts, a seasonal forecast system is experimented with here. JISAO’s Seasonal Coastal Ocean Prediction of the Ecosystem (J-SCOPE) features dynamical downscaling of regional ocean conditions in Washington and Oregon waters using a combination of a high-resolution regional model with biogeochemistry and forecasts from NOAA’s Climate Forecast System (CFS). Model performance and predictability were examined for sea surface temperature (SST), bottom temperature, bottom oxygen, pH, and aragonite saturation state through model hindcasts, reforecast, and forecast comparisons with observations. Results indicate J-SCOPE forecasts have measurable skill on seasonal timescales. Experiments suggest that seasonal forecasting of ocean conditions important for fisheries is possible with the right combination of components. Those components include regional predictability on seasonal timescales of the physical environment from a large-scale model, a high-resolution regional model with biogeochemistry that simulates seasonal conditions in hindcasts, a relationship with local stakeholders, and a real-time observational network. Multiple efforts and approaches in different regions would advance knowledge to provide additional tools to fishers and other stakeholders.

Scientific RepoRts | 6:27203 | DOI: 10.1038/srep27203 ROMS. The Regional Ocean Modeling System (ROMS; Rutgers version 3) is configured for the Washington, Oregon, and British Columbia coasts, using the Cascadia domain 30 . The domain (Fig. 1) extends from 43°N to 50°N with a horizontal resolution of 1.5 km, and 40 vertical levels. In the hindcast simulations, 16 rivers are based on observed streamflows 30 , but in the reforecast and forecast runs, the rivers are forced using a climatology of local river discharge data over eight years (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007). Tides are included. The biogeochemistry model is described in previous studies 24, 31 .
Water entering the domain at the southern and western boundaries is specified by CFS. Biogeochemical boundary conditions were implemented as described elsewhere 24,31 , using local empirical relationships with salinity. We applied these relationships to generate the initial and boundary conditions as predicted by CFS salinity fields. Experiments. Regional   The hindcast anomaly is on top, the January initialized forecast anomaly in the middle, and the April initialized forecast anomaly on the bottom. Each column displays a different model field. The first column shows bottom oxygen (ml/l), the second, surface chlorophyll, the third is SST, and the last column is bottom pH. Below the maps, a time series of the 8-day upwelling index (calculated following methods from Austin and Barth 2002) is plotted from the hindcast (black), the January initialized forecast (blue) and the April initialized forecast (green). The vertical dotted lines on the time series bracket the upwelling season over which the maps are averaged. Figure generated using Matlab version 2015b (http://www. mathworks.com/products/new_products/latest_features.html) and Adobe Illustrator CS5 (http://www.adobe. com/products/illustrator.html).
were test from different years. Forecasts were nine-month projections, requiring six days of computer time on 96 cores. In hindcast mode, the forcing was derived from the CFS-R reanalysis (for 1979-2009), which includes data assimilation, and the CFSv2 analyses (for 2010-present). An important implication of the resolution and the data assimilation is that the simulations using the reanalysis product will be superior to those based on the coarser fields available for forecast forcing from the CFS model.
In this study we applied interpolated CFS results directly to ROMS as surface forcing, boundary conditions, and initial conditions, without bias correction. As a result, our regional model will inherit some of the biases of the large-scale CFS simulations. It is recognized that various methods exist for bias correction, and future versions of J-SCOPE may employ such methods.
January and April were chosen as initialization months in order to test whether 2-4 month predictions could forecast critical periods for fish stocks. Specifically, January forecasts of the spring transition have the potential to predict the onset of upwelling and recruitment success of salmon and many 'spring spawning' rockfish, while April forecasts have the potential to forecast the extent of northern migrations of major commercial species such as Pacific whiting (Merluccius productus) and Pacific sardines (Sardinops sagax).
Based on the ROMS output of oxygen and temperature, we applied regional proxy relationships 32 that predict aragonite saturation state. Results were compared with observations made on coastal cruises performed in the summer of 2013 33 . From the modeled fields we calculated the percent of the upper 100 meters of the water column that was undersaturated with respect to aragonite. The upper 100 meters was specified because it is the approximate vertical range of the pteropod, a calcifying zooplankton affected by ocean acidification in this region 34 . Skill Assessment and Predictability of CFS in the J-SCOPE domain. As a prelude to assessments of J-SCOPE forecast skill, we examined six-month forecast skill of CFS forcing within the J-SCOPE domain. For this comparison a subset of the CFS reforecast data was compared with corresponding values from the CFS Reanalysis (CFS-R). For the reforecasts, we utilized four realizations of the forecast system from the 15 th of January of each year, spanning the period 1997-2009 (the period during which ARGO data were assimilated), and examined results of the six-month forecasts of July as compared with the corresponding CFS-R values (52 realizations). We considered four monthly mean quantities in this comparison: SST, density at 40 meters depth as a metric for upwelling, shortwave radiation, and alongshore surface winds. We compared persistence skill with model forecast skill, in order to assess whether the CFS model outperforms one that assumes an average seasonal cycle with temporal autocorrelation in the anomalies. For SST and density at 40-meters, we also correlated reanalysis tendency (i.e. change) from January to July with predicted tendency over that period. The 40-meter density metric was chosen for three reasons: 1) subsurface values are a better metric of upwelling than near-surface values, as vertical velocities are greater at depth; 2) 40 m is largely insulated from diabatic seasonal change; 3) CFS-R has a sea surface salinity bias due to an excessive relaxation to climatology 35 .
In each case we compared average forecast July values with average July reanalysis values (mean bias), correlated forecast July values to July reanalysis values (potential forecast skill), and correlated July reanalysis values to January reanalysis values (persistence skill). For SST and density at 40 meters, we also correlated reanalyzed change from January to July with predicted change over that period.
Skill Assessment and Predictability of J-SCOPE. Hindcast skill and predictive skill assessment of the J-SCOPE model were based on comparisons to observations from in situ cruise data spanning 2009-2014, and at two moorings from the Olympic Coast National Marine Sanctuary (OCNMS) in 42 and 15 meters of water. In 2013 an additional mooring (NANOOS Cha'ba) in 90 meters of water off the northern tip of Washington's outer shelf was available and we included it in comparisons summarized in the SI ( Figure S4). Both the model and the observations were smoothed with a moving average filter of 30 days to focus on seasonal timescales.
To quantify model performance, the model climatology was compared to the co-located observational climatology. Time-series comparisons are presented at the locations of two OCNMS moorings on the Washington shelf (CE042, CE015) as well as three profiles along the Newport line from the inner, mid, and outer shelf of Oregon and temperature time series from NH10 36,37 . The moorings were deployed during a portion of the year, and so the comparisons among climatologies are limited to the period the observations were available (May through October). The OCNMS mooring records were averaged over 2004-2014 for temperature and 2006-2014 for oxygen. The Newport line observations (2009-2014) are made bi-weekly year-round, and here we present CTD measurements of temperature and oxygen at three locations on the Oregon shelf: NH03, NH10, and NH25. In addition, we compare with SST (2 m) and BT (70 m) temperatures from NH10 (2009-2014). The model fields were extracted from the same location as the observations from each year, and then those time series were averaged together to make a model climatology for that location.
We use the hindcast simulation of 2009-2014 as a proxy for a long-term climatology. We tested this assumption by comparing the physical forcing from the long-term climatology from CFS-R to the six years used in this study. The shorter climatology is representative of the longer-term climatology for SST and alongshore winds in the model domain ( Figures S1 and S2).
To quantify predictability of the forecasts, the forecasted anomalies were compared to the hindcasted anomalies from 2013 and 2009 at the same OCNMS and Newport locations. The residual time series convey qualitative information about how model skill declines with time. Anomalies were generated by subtracting the model climatology results from the forecasts and hindcast over the same time period. We calculated three skill metrics when comparing the observed and simulated climatologies and the forecasted and hindcasted anomalies (see SI for equations): the normalized unbiased Root Mean Squared Deviation (RMSD 38 ), the correlation coefficient (R-value), and the normalized Bias. The RMSD is influenced by both the phasing of the series and how well the hindcast (forecast) variability compares with the observed (hindcast) variability. RMSD is normalized by the standard deviation of the reference field (either the observations or the climatology) (Eq. 2, SI). As defined here, Scientific RepoRts | 6:27203 | DOI: 10.1038/srep27203 Table 1. Summary of statistics for "Performance" and "Predictability" for the forecast and hindcast simulations of the J-SCOPE forecast system averaged over the upwelling season (April-September). The RMSD is indicated in bold, the R-value is indicated in italics, and the bias is in parentheses. The results are shown for four locations: the Washington shelf as represented by the coarse-scale CFS model (CFS), Washington mid shelf (CE042), Washington inner shelf (CE015), and Oregon outer shelf (NH10). The column 'Model Clim vs Obs Clim' is a metric of "Performance", comparing the hindcast simulated climatology to the observed climatology on the same timeframe as the samples (see Sup. Table 1). Other columns are measures of "Predictability" comparing forecast anomalies to the hindcast (or reanalysis for CFS) anomaly. For the profiles, the anomalies were averaged over depth to get one value for each observation, and then correlations were made based on these depth-averaged anomalies. Each row displays a different model field (SST, bottom temperature (°C) or bottom oxygen (ml/l)). Teal shaded boxes indicate performance or predictability better than our skill thresholds (− 1< RMSD< 1 and R > 0.5). Grey shaded boxes indicate that either R or RMSD are better than the skill threshold, but not both. the normalized total RMSD value also informs whether the model's standard deviation is larger (RMSD > 0) or smaller (RMSD < 0) than the standard deviation of the reference field RMSD. As the R-values approach 1, the phasing between the two temporal signals are in agreement; note however that this metric alone does not indicate the correspondence between the magnitudes of the two signals. Finally, we report bias between the model and reference fields. Summary statistics averaged over the upwelling season (April -September) are reported in Table 1 and equations for the skill metrics are available in the SI. Additional information about the data sets used is available in SI Table 1.

Results
J-SCOPE forecast results are available on NANOOS website bi-annually for stakeholders and the public. Results from the 2013 forecasts and hindcast are shown in Fig. 1 for SST, bottom oxygen, chlorophyll, and pH anomalies over the upwelling season, along with the wind forcing. The forecasted anomalies all indicate the right direction and in some cases, spatial patterns consistent with the hindcast anomaly. A summary of the skill statistics is given Mean fields and skill metrics for sea surface temperature (SST, °C), density at 40 meters (σ 40m , kg/m 3 ), north/ south winds at the sea surface (V wind , N/m 2 ), and downward shortwave radiation at the sea surface (SW, W/m 2 ). Boxed region indicates the J-SCOPE domain. Shown from left to right are: 1) mean July forecast (initialized from previous January); 2) mean July reanalysis; 3) "forecast skill" (correlation between July reanalysis and July forecast); 4) "skill over persistence" (forecast skill minus correlation between January reanalysis and July reanalysis); 5) "tendency forecast skill" (correlation of the change from January to July with predicted tendency over that period). The first two columns use the same colorbar (on the first column). The latter three columns use the colorbar in the 3 rd column. Figure generated using Ferret version 6.93 (http://www.ferret.noaa.gov/ Ferret/) and Adobe Illustrator CS5 (http://www.adobe.com/products/illustrator.html).
in Table 1 and below we discuss the performance and predictability skill of these forecasts. These skill assessments help to convey the level of uncertainty in the forecasts, and how this varies for each ocean variable.
Hindcast and reforecast skill of CFS, for our region. J-SCOPE forecasts rely on output from CFS for the climate forcing, so it is worthwhile considering the quality of CFS predictions for context. CFS has measurable skill in predicting SST in the Northeast Pacific for lead times of less than six months 26,39 , and for the spatial mean SST of the CCS, skill is typically greater than simple persistence when predicting spring temperatures from previous months. Much of this skill resides with the region's systematic response to both local and remote effects of ENSO associated variability 26 . Prior work 39 noted that model skill in the CCS rarely exceeded persistence with 90% significance; however, those spatial averages include both the Pacific Northwest (PNW) (where figures 26 suggest skill greater than persistence) and southern California (where figures 26 suggest skill less than persistence).
Here we take a regional PNW perspective; other studies 26,39-41 provide more quantitative results for the CFS model over an extended period and a broader spatial scale. Consistent with prior results, CFS had substantial skill for the years 1997-2009 and 2013, 2014 when measured against a reanalysis product (CFS-R) (Fig. 2, Table 1). For SST, July forecasts from the previous January have a ~2 degree warm bias near the coast, but achieve useful skill (r~0.4-0.6) in the J-SCOPE domain. Persistence skill is substantially lower in the J-SCOPE domain (r~0.0-0.4, i.e. forecast skill is as much as 0.4 greater than persistence skill in that region; Fig. 2). Forecast skill degrades further south along the California coastline, consistent with the pattern obtained in other studies 26,39 for seasonal predictions of the North Pacific. The PNW may gain some predictability via atmospheric teleconnections with other areas in the Pacific Ocean basin. Our forecast skill values for the Jan-July change in SST (i.e. the 6-month average rate of change of SST over that period) are higher than the forecast skill values for July SST.
For density at 40 meters, the forecast values near the coast are slightly greater than the reanalyzed values (Fig. 2). Forecast skill is appreciable throughout the PNW domain, and extends further south than was the case for SST. Persistence skill is substantially lower (and even negative) around the Columbia River outflow; hence the forecast model adds significant value there relative to simple persistence of observed anomalies from January. Forecast skill for the Jan-July change in density is roughly of the same order as the July forecast skill, achieving values of 0.6-0.8 in the PNW domain. This suggests that the CFS is serving as a useful predictor of seasonally integrated upwelling.
Shortwave radiation, which is the solar energy that warms the ocean surface and drives photosynthesis by phytoplankton, is biased high in CFS, and this bias is inherited by J-SCOPE. For the incident shortwave values, the forecast is biased significantly higher, and a modest correlation exists between reanalysis and predicted July values along the PNW coastline, which exceeds simple persistence (Fig. 2). In 2013, the forecasted shortwave radiation was biased high by about 50 W/m 2 when averaged over the modeled region (Fig. 3a), and by nearly 100 W/m 2 in 2009 (not shown). This suggests significant interannual variability in the bias of the CFS, which has not been characterized in our region.
For alongshore winds, the July forecast has a slight increase in skill over anomaly persistence, and is slightly biased towards stronger northerly winds in the PNW, with weak forecast skill at the 6-month lead time (Fig. 2). It remains to be quantified how skillful the CFS forecasts are at capturing the time-integrated wind stress over the   upwelling season, as compared to these 6-month forecasts of July winds. Based on observed winds from 2013, the observed summer upwelling season of 2013 began on April 8 (damp.coas.oregonstate.edu) 42 , which is typical, but the season ended 20 days earlier than average. The CFS forecasted winds captured the onset of the upwelling season with the spring transition, and the January forecast was close to the observed cumulative upwelling value for August (not shown). However, the upwelling season was too long and the winds were too strong, without enough relaxations (Fig. 3b).
In summary we find that CFS has measurable skill (relative to a null model of persistence) for two key metrics of ocean conditions that are relevant to our upwelling-driven ecosystem: integrated upwelling and alongshore wind stress. All models are biased and imperfect; however, two key biases are evident in the 2013 CFS forecasts that subsequently influence the J-SCOPE forecasts: a. Shortwave radiation (too much solar radiation) b. Wind Events (upwelling winds persist too late into the fall and too strong with minimal relaxations throughout the upwelling season).
These patterns in shortwave radiation and winds affect predictions, particularly the ability to forecast late summer and fall ocean conditions. J-SCOPE skill: performance and predictability. Below Table 1 and the discussion of the results below.
Model performance. When compared to local observations, J-SCOPE predicts the trends and seasonality of the observed SST, bottom temperature, and bottom oxygen on both the Washington and Oregon shelves: CE015 (Fig. 4), CE042 (Fig. 5), and Newport (Fig. 6). The model captures the seasonal cycle -SST is warmest during the summer, bottom temperature is coldest during the summer upwelling season on the shelf, and bottom oxygen declines over the course of the summer upwelling season. The shaded regions in Figs 4-6 indicate the interannual To quantify the agreement between the modeled and observed climatologies, normalized RMSD as well as R-values and bias were computed for the time series and profiles from Figs 4-6, and the results are compiled in Table 1. Additional winter and fall profiles are available in the Supplementary Information (Figure S3). The model climatology captures the seasonal variability in the observed climatology as indicated by the normalized RMSD values (< 1), and has significant R-values for all variables and locations tested except for SST at the shallowest location (CE015, Table 1).
The model simulates bottom temperature more realistically than SST, when compared to observations (Table 1) Skill assessment on individual years, such as 2013, is available in the Supplementary Figures S4 and S5. Results of the skill assessment presented here are similar to the RMSD values from prior hindcast simulations with higher resolution atmospheric forcing 31 . Because the model can simulate the observed seasonal cycle in all three ocean conditions with reasonable skill as a hindcast, the model performs well enough to be tested in forecast mode and its predictability can be tested using similar metrics.
Model Predictability. The results above focus on performance of the hindcasts in reproducing the seasonal cycle; the general strengths and weaknesses evident in the hindcasts are also apparent in terms of model predictability, assessed by comparison of hindcasts to forecasts of the same period. Qualitatively, the evolution of model forecasts for 2013 matches trends at the two moorings ( Figure S4). While these results are promising, it is important to show that J-SCOPE has skill beyond simply reproducing the seasonal cycle. The following discussion will focus on the predictability of SST, bottom temperature, and bottom oxygen concentrations at all three locations discussed previously (Figs 7-9), as well as within the entire model domain (Fig. 10), over the upwelling season (April-September), specifically the skill of the forecasted anomalies from monthly climatology. Note that the January initialized forecasts have a longer lead time than the April initialized forecasts.
While most of the SST anomalies are poorly forecasted by J-SCOPE, the re-forecasted anomalies for SST during 2009 were well correlated with hindcast anomalies (R > 0.5) on the shelf throughout the upwelling season  Table 1), which was probably due to the bias in the shortwave radiation known to exist in this large-scale climate model (Fig. 3). The model performed the worst in the April 2014 forecast. A map of the modeled performance comparing the forecasted anomalies to the hindcasted anomalies over the upwelling season, plotted as R, confirms these patterns of skill throughout the domain (Fig. 10). The regional model forecast was able to reproduce the variability, and in most forecasts, perform on par with CFS in the region (Table 1).
Consistent with the hindcast simulations, forecast model performance was better for bottom conditions than for SST. Forecasted bottom temperatures were biased cold in most forecasts by 0.4-2.4 degrees C (Table 1, Figure S4). Despite this bias, the forecasted bottom temperature anomalies were well correlated with the hindcast anomalies (R > 0.5) on the outer shelf during the upwelling season (Figs 7 and 8, Table 1), except for the April 2014 forecast. Some of the forecasts had R values better than our significance threshold (R > 0.5), but had non-significant RMSD values indicating that the forecasts had the correct phasing but with a different amplitudes than the hindcasts (Table 1). Bottom temperature forecasted anomalies perform better on the mid-shelf (CE042) than at the shallower site (CE015, Table 1). This pattern emerges in the spatial maps of R between the forecasted and hindcasted anomalies of bottom temperature, as well (Fig. 10).
On the Oregon shelf, the temperature profile comparisons along the Newport Line were forecast skillfully (RMSD < 1, R > 0.5) for half of the realizations (Fig. 9, Table 1). The vertical structure of the water column is captured by the forecast when averaged over the entire upwelling season (Table 1), in addition to seasonal anomalies Figure 9. A comparison of forecast versus hindcast anomalies from the NH10 location (mid-shelf) along the Newport Line on the Oregon shelf. All anomalies were calculated as the difference from a climatology based on the average of 2009-2014 hindcasts. January forecast is in blue, the April forecast in green, and the hindcast is in black. Temperature (left) and oxygen (right) profiles are shown from spring (April-June) and summer (July-September), fall (October-December) and winter (Jan-March). Statistics summarized in Table 1. (Fig. 9). As discussed, the fall transition is difficult to forecast on these timescales, so the lack of skill in the forecast during this time is not surprising.
The model demonstrated the most predictability for oxygen, both in terms of phasing (R) and overall variability (RMSD). Bottom oxygen forecasted anomalies were reasonably well simulated for all the locations (Table 1). Similar to the other ocean conditions, R > 0.5 suggest that phasing of the forecasts is correct, even if RMSD > 1 suggests that the variability (e.g. amplitudes) differ ( Table 1) for some of them. One of the forecasts had significant RMSD values, but non-significant or negative R-values indicating that the forecasts had the right variability but was not in phase with the hindcasts. In general, the forecasts were biased low. This bias was most likely due to strong, persistent forecasted upwelling-favorable winds, as relaxations have been shown to be important to relieve hypoxia 43,44 and the forecasted winds from this year did not experience the observed frequency of relaxations (Fig. 3). Overall, the forecasted bottom oxygen anomalies were skillfully represented throughout the model domain (Fig. 10). While all the forecasts performed the worst in 2014, and in the shallowest region of the shelf, they still maintained broad areas of skillful forecasts (Fig. 10, Table 1).
On the Oregon shelf, the forecasts of vertical profiles of oxygen at the mid-shelf location of the Newport line (NH10) resemble their observed counterparts in terms of averages over the entire upwelling season ( Table 1). The model performs the worst in the spring and fall months (Fig. 9). The fall transition is difficult to predict on seasonal timescales.
When the residuals of the forecasted anomalies from the hindcast anomalies are plotted in time, for all four forecasts shown in Table 1, the predictability of ocean conditions over seasonal timescales becomes more evident. Both the January-and April-initialized forecasts have minimal residuals (< 1) for bottom temperature and bottom oxygen until September (Day of Year 270, Fig. 11). For SST, the residuals vary more over the time series at this mid-shelf location (CE042).
Overall, J-SCOPE does have skill in forecasting beyond the seasonal cycle, but the predictive skill of J-SCOPE depends on the parameter of interest, the time of year, and the location in the domain. The results from these four forecasts are promising and indicate that seasonal forecasts in this region of these ocean conditions for fisheries are possible.
Aragonite saturation forecasts. Empirical relationships 32 using temperature and oxygen to estimate aragonite saturation state were employed to create aragonite saturation maps and cross-sections from the April forecasts of the August cruise period from 2013. Forecasts are compared to aragonite saturation observations from the cruise ( Figure S5).
Over the upwelling season, the smallest amount of undersaturated water is present at the onset of upwelling, and it evolves over the ensuing months until it reaches a maximum prior to the fall transition. Despite small biases discussed in the SI, the forecast is able to simulate the spatial gradients of the increasing percentage of the upper water column that is undersaturated when compared with observations (Fig. 12). The shelf water feeds the estuaries, which feature shellfish aquaculture throughout the upwelling season. Regionally, Heceta Bank (44-45° N) in Oregon and the region corresponding to the Juan de Fuca Eddy, just outside the mouth of the Strait of Juan de Fuca, both experience the highest percentage of undersaturated water in the upper 100 m of the water column. These regions correspond to hot spots for respiration associated with retentive zones 31 . In addition, they correspond to regions that experience the most intense dissolution of pteropods (calcifying zooplankton) collected during a cruise in 2011 34 .

Discussion
The J-SCOPE forecast system relies upon the following components: a real-time observational network, a working hindcast simulation of the seasonal patterns for the region complete with biogeochemistry, a region with significant skill from a global seasonal forecast system like CFS, and an identified group of stakeholders with products designed in mind for them. For J-SCOPE, NANOOS provides a portal for real-time regional observations. Regional hydrodynamic hindcast models have been developed to understand the dynamics on the shelf with success, such that biogeochemical models can be designed and linked to them as well 24,30,31,[45][46][47] . NANOOS and the California Current Integrated Ecosystem Assessment 3 bring linkage to and feedback from resource managers and other stakeholders with interest in these oceanographic and biogeochemical models.
J-SCOPE results indicate the forecasts have predictive skill on the order of several months into the future from late winter conditions through much of the upwelling season. The best skill lies on the outer shelf and near the bottom, mostly because of the biases at the surface, specifically due to excess solar radiation and wind in CFS on these timescales. Sensitivity tests to the shortwave forcing performed on the hindcast simulations revealed that a 20% reduction in the shortwave forcing reduced the SST by 2-3 degrees C on average, which is enough to account for the bias in 2013. The strengths of the model include the bottom temperature and bottom oxygen forecasts on the mid to outer shelf, and the spatial variability of those conditions over the upwelling season. Certain dynamics are less predictable on these timescales, and these limitations should be kept in mind. Specifically, the fall transition, which brings the end of the upwelling season, appears to be poorly predicted. This recurring problem means that the cumulative upwelling index will be too high, duration of hypoxic or ocean acidification events will be too long, and the severity of events will be exaggerated. This problem could be rectified through a combination of forecasts with increasing time resolution combined with an observations network. The resulting toolbox could prove useful for management decisions. A complete characterization of forecast skill would require a more extensive forecast suite and longer time series consisting of moored observations that span the entire year.
As an example of the potential utility for J-SCOPE, a recent study 22 used it to forecast the spatial distribution of Pacific sardine, whose migration and spawning area respond directly to surface temperature and other ocean conditions 48,49 . The study fit generalized additive models relating 2009 ocean conditions (SST, salinity, and chlorophyll), as reforecast by J-SCOPE, to observed spatial distributions of sardines. Notably, the authors 22 found that using the 2009 reforecasts initialized in January had moderate ability to predict May-August sardine distributions from three surveys. Their method accounts for the warm bias in surface waters by fitting sardine presence to J-SCOPE reforecasts before using the forecasts to predict future sardine distribution. Ongoing research aims to extend the forecasting to other species known to respond to ocean conditions such as tunas 50 , Pacific hake Merluccius productus 51 , and salmon Oncorhynchus spp. 52,11 . J-SCOPE development is ongoing, and some technical improvements target some of the limitations identified above. Specifically, CFS contains a combination of biases in the wind stress and cloud fields leading to over-estimation of insolation 53 . Future efforts might consider dynamical downscaling with a high-resolution coupled atmosphere-ocean model 53,54 in order reduce the biases outlined here. We will continue to test the forecasts in order better understand our capabilities and to convey uncertainties in the forecasts. In addition, we are including ensembles of multiple model runs in our future forecasts to better gauge uncertainty.
Having begun experimenting with seasonal forecast capabilities, we see strong potential to translate ongoing J-SCOPE forecasts into products relevant to US West Coast commercial fisheries, focusing on the phenomena that trigger stakeholder decisions and for which J-SCOPE forecasts have the most skill. In particular, we see potential utility for fisheries targeting pelagic species that respond strongly to temperature (e.g. hake, tuna, and mackerel in addition to sardine), since fishermen typically follow water temperatures when making decisions about fishing location 55 and in at least one case are already using short-term (48 hr) forecasts (http://nvs.nanoos. org/TunaFish).
Our results suggest that J-SCOPE has skill to predict the location and onset of hypoxia. The onset of hypoxia is sufficient information for state and tribal crab managers to prohibit or warn against the setting of traps similar to closures already triggered by harmful algal blooms (http://wdfw.wa.gov/news/aug0415a/). Additionally, the availability of empirical relationships for aragonite saturation 32 allowed us to relate ocean conditions to aragonite saturation state that will provide useful predictions for shellfish operations. Specifically, our results suggest that J-SCOPE has skill to predict corrosive bottom-waters on the shelf just outside estuaries with shellfish operations. Shellfish aquaculture is particularly sensitive to corrosive conditions. Further work with commercial fisheries and fishery managers will require direct engagement with stakeholders to understand these critical decisions and time scales for these decisions 21 . We expect seasonal forecast systems to have broad applicability to stakeholders and managers in other regions, similar to existing forecast systems for algal blooms and pathogens 56,57 .

Conclusions
Experiments with J-SCOPE suggest seasonal forecasting of regional ocean conditions for fisheries and managers is possible with the right combination of components. Those components include regional predictability on the seasonal timescale from a large-scale model of the physical environment (e.g. SST, winds), a high-resolution regional model with biogeochemistry that simulates regional seasonal conditions in hindcasts, a working relationship with local stakeholders, and a real-time observational network. Through the experiments described here, we discovered that biases in the shortwave radiation and the winds from the CFS model present challenges to forecasting on seasonal timescales. Despite these challenges, results suggest J-SCOPE forecasts have skill on timescales up to four months depending on the variable and time of year. We are working through NANOOS to socialize such results with regional managers and fishers to gain their feedback on the utility of such information.
One objective of this paper is to encourage other research teams to make seasonal forecasts of direct relevance to marine ecosystems, by coupling CFS (or other seasonal forecasts) to the many existing regional models of physical oceanography (e.g. ROMS, HYCOM, FVCOM), partnering with a real-time observational network, and seeking out local stakeholders and fishers of interest. The vision for the Global Ocean Acidification Observing Network (GOA-ON) highlights the mutually beneficial relationships among real-time observing, forecasting, and stakeholder communities 58,59 . Regional models that include plankton dynamics and biogeochemistry facilitate predictions beyond physics, reaching additional users and stakeholder needs. Multiple efforts and approaches, and in particular the lessons learned from different systems and regions, will be the best way to make progress in this arena.