CHAPTER TWO

History and Current Status of S2S Forecasting

Providing useful weather and ocean forecasts, as well as predicting other aspects of the Earth system, have significantly improved national capabilities for decision-making in sectors including energy, agriculture, transportation, insurance and finance, defense, emergency preparedness and response, and public security including health, water, and food. As discussed in Chapters 1 and 3, the ability to foresee environmental changes and disruptive events weeks and months in advance could have tremendous additional value because of the broad range of decisions that are made weeks to months in advance. As a prelude to developing a U.S. research agenda for advancing subseasonal to seasonal (S2S) forecasting, this chapter lays out the history and evolution of the S2S forecast endeavor and briefly summarizes current operational capabilities and research activities.

EVOLUTION OF THE FORECAST ENTERPRISE

Short- to Medium-Range Forecasts (Up to 14 Days)

Modern weather prediction evolved from the global weather observations obtained during World War II, the computers that followed in the wake of the war, and a working knowledge of equations that model the typical variations in the midlatitude atmosphere. Earth-sensing satellites, starting with the Television Infrared Observation Satellite Program (TIROS) in the 1960s, provided striking views of Earth’s changing weather patterns and contributed to the understanding of weather systems and to the improvement of routine weather forecasts.

With these improved data sources and modeling capabilities, purely subjective forecasts based on atmospheric synoptic maps, experience, and intuition gave way to a combination of computer-generated atmospheric and marine forecasts based on physics equations and a statistical interpretation of the forecast information. This know-how developed into highly capable systems operated by the civil and defense weather services. In the latter part of the 20th century, consumer interests in weather and the wide demand for specialized forecasts stimulated a vigorous private sector

operating alongside the public weather services (NRC, 2003). Similar trends are also occurring for ocean forecasting and applications.

Moving into the 21st century, the combination of greatly improved atmospheric and oceanic observations and accelerating computer power has produced increasingly accurate and reliable atmospheric forecasts. Computer-calculated forecasts of global and regional weather patterns are now as accurate at 72 hours as they were at 36 hours in the 1990s (Figure 2.1). Although this might suggest that lead times for useful forecasts could continue to increase indefinitely with further improvements

in observations, understanding, and computer capability, the discovery of mathematical chaos in nonlinear physical systems in the early 1960s by Edward N. Lorenz (Lorenz, 1963) challenged this assumption. Instead, Lorenz showed that unavoidable small errors in initial conditions will amplify during the computation, bringing a natural limit to the lead time that the “weather”—or any given natural environmental phenomenon—can be predicted, at least deterministically. Today, the emphasis is on improvement and extension of lead times through probability forecasts, created by averaging over space and time and running multiple cases to create ensembles of forecasts that reflect probabilities of variables or events at future times. Along with probabilistic ensemble forecasts, recent advances in weather prediction accuracy have come from improved understanding of the underlying processes and more realistically incorporating them into the forecast models, in part by increasing model spatial resolution and in part through better parameterization of unresolved processes. Furthermore, improved measurements and assimilation of those measurements to more accurately reflect the initial state of the environment have also greatly advanced the skill of weather forecasts.

Following the atmospheric community, short- to medium-range (up to 7 days) ocean forecasts have been routinely available for the past 10 years and provide predictions of ocean currents, temperature, and salinity (GODAE, 2009). These predictions are used by national agencies (e.g., Navy, National Oceanic and Atmospheric Administration, Coast Guard), the oil industry, and fisheries, among others, for various applications such as ship and submarine routing, search and rescue, deep ocean drilling, oil spill drift application, monitoring of open ocean ecosystems, fisheries management, coastal and near-shore resource management (e.g., Jacobs et al., 2009). The development of these ocean forecast systems has been critical to the development of high-resolution coupled ocean-atmosphere-ice-land prediction systems for improving short- to medium-range forecasts, and they are being used by the ocean community to develop subseasonal ocean forecasts (Brassington et al., 2015). With the addition of aerosol chemistry and biogeochemistry, such models are often referred to as Earth prediction systems. Advances in coupled model systems are central to extending lead times and furthering accuracy of short- and medium-term forecasting capabilities in the ocean and atmosphere, are the basis for advancing S2S forecasts, and are critical for developing a more expansive set of routinely forecast Earth system variables.

Seasonal Forecasts (3 to 12 Months)

Long-range and seasonal forecasts began in the mid-1950s as Weather Bureau forecasters noticed some identifiable large-scale patterns and relations between atmo-

spheric and ocean temperature anomalies in various locations (Hoskins and Karoly, 1981; Namias, 1953; Roads, 1999; Walker, 1924; Wallace and Gutzler, 1981). These early seasonal forecasts were made based on statistical methods. Dynamical seasonal predictions started in the early 1980s (Reeves and Gemmill, 2004), using atmosphere-only models with prescribed surface conditions. Often, the latest observed ocean anomalies persisted during the forecast, but other surface conditions, such as sea ice, snow cover, and soil moisture, were proscribed from climatology (e.g., average historical conditions). Such systems treated the surface as a fixed boundary condition and generally ignored the coupled dynamics with the surface that evolved over the forecast period (two-tier system). Focused on El Niño-Southern Oscillation (ENSO; see Box 1.3) prediction, the first coupled atmosphere-ocean forecasts were generated with simple dynamical or statistical models of tropical surface temperatures (Cane et al., 1986; Graham et al., 1987; NRC, 1986, 1991a, 1994; Shukla, 1998).

In contemporary seasonal forecast systems, many aspects of the Earth system are predicted in a coupled model involving the atmosphere, ocean, land, and cryosphere. These seasonal forecast systems seek to better exploit ENSO as a source of predictability, while also representing more recently discovered predictability sources originating from other natural modes of variability of the coupled ocean-atmosphere system; stratosphere-troposphere interactions; the slow evolution of the ocean, ice, land hydrology and biosphere; and radiative forcing from greenhouse gas (GHG) and aerosols and land use changes (see Chapter 4). In dynamic seasonal prediction systems, modeled Earth system components (i.e., atmosphere, land, ocean, and sea ice) are increasingly coupled numerically to represent the rapidly varying atmosphere exchanges of energy, water, and momentum, which give the system much of its predictability on timescales longer than a few days. Additional progress could be made: some components, such as the ocean, are more realistically coupled with the atmosphere, while aspects of coupling to the cryosphere and land are widely recognized to be oversimplified in today’s forecast systems (Doblas-Reyes et al., 2013). Recent research indicates that much of the seasonal predictability in some parts of the world derives from trends associated with GHG warming superimposed on natural variability, thus more realistic representation of atmospheric chemistry and biogeochemistry (e.g., GHG forcing, land use changes, aerosols) in seasonal prediction systems is also increasingly common (Doblas-Reyes et al., 2006).

Seasonal forecasting has improved over the past decade with efforts to reduce systematic model errors and with better understanding and representation of sources of predictability within the coupled Earth system. There are two other notable strategies for advancing the skill and utility of seasonal forecasts. One is the inclusion of quantitative information regarding uncertainty (i.e., probabilistic prediction) in forecasts

and probabilistic measures of forecast quality in the verifications (e.g., Dewitt, 2005; Doblas-Reyes et al., 2005; Goddard et al., 2001; Hagedorn et al., 2005; Kirtman, 2003; Palmer et al., 2000, 2004; Saha et al., 2006, among many others). This change in prediction strategy naturally follows from the fact that Earth system variability includes a chaotic or irregular component, and, because of this, forecasts must include a quantitative assessment of this uncertainty. More importantly, the prediction community now understands that the potential utility of forecasts is based on end-user decision support (Challinor et al., 2005; Morse et al., 2005; Palmer et al., 2000; Chapter 3), which requires probabilistic forecasts that include quantitative information regarding forecast uncertainty or reliability.

The use of perturbed parameter ensembles represents a second strategy that is now commonly used to quantify uncertainty in the initial conditions of seasonal prediction systems, though the number of such ensembles in both the forecast and the retrospective forecast vary widely across different operational centers (Appendix B). Other techniques have been implemented to account for uncertainty in model formulation. Most prominent among these is the development of multi-model ensembles (MMEs). By combining the predictions from more than one model, MMEs quantify some of the uncertainty associated with individual model formulations and also tend to improve the forecast, probably because errors in one model may not appear in the others. With a few caveats, MMEs that include multiple operational and/or research models appear to achieve a better skill than individual models, by combining different approaches to data analysis, data assimilation, model parameterizations and resolutions (Kirtman, 2014; Kirtman et al., 2014; Weigel et al., 2008). Other techniques, such as perturbed physics ensembles or stochastic physics (e.g., Berner et al., 2008, 2011) have also been developed and appear to be quite promising for representing some aspects of model uncertainty (e.g., Weisheimer et al., 2011). Chapter 5 covers these developments in more detail.

Subseasonal Forecasts (2-12 Weeks)

A prevailing expectation is that subseasonal prediction in the 2- to 12-week range between short- and medium-range prediction and seasonal prediction poses serious challenges. This expectation arose from the perception that the subseasonal atmospheric forecast problem does not fit neatly into the simplistic paradigms of an initial-value weather forecast problem (because the lead times are too large and initial-value information can be lost) or the so-called “boundary-value climate prediction problem,” terminology associated with the early seasonal climate forecast systems that were driven by prescribed surface temperature anomalies. However, recent work indicates

the potential for predictability across all timescales (Hoskins, 2013; WMO, 2015a). There is evidence to indicate that the existing coupled ocean-atmosphere-ice-land Earth system forecast models, mentioned above, integrate the information from the initial conditions across the coupled system, including the slowly varying components (e.g., ocean, sea ice, and land hydrology), to produce subseasonal forecasts with realized skill in traditional weather variables often comparable to that of the seasonal forecasts (Dutton et al., 2013, 2015).

Predictability and prediction studies on intraseasonal tropical variability and the Madden Julian Oscillation (MJO; see Box 1.3) have further advanced the prospects of subseasonal forecasting (e.g., Lin et al., 2008; Vitart et al., 2007b; Waliser et al., 2006). However, it is important to note that within the subseasonal timescale, predictability and prediction in sub-monthly timescale is still relatively underexplored and underdeveloped compared to forecasts with lead times of a month to a season (Doblas-Reyes et al., 2013; Vitart et al., 2012). An important goal for subseasonal (and seasonal) forecasting is to move beyond multi-day averages of typical meteorological variables to prediction of the likelihood of important and disruptive events in all components of the Earth system, such as heat and cold waves, unusual storminess, ice cover, sea level, Gulf of Mexico Loop current position, etc.

CURRENT STATUS OF ACTIVITIES AND RECENT PROGRESS

This section provides a brief survey of current capabilities and ongoing activities in both seasonal and subseasonal prediction, along with recent progress at operational centers. This is a prelude to establishing a U.S. research agenda that will lead to improved S2S forecasts and better-informed decisions in both the public and private sectors.

Seasonal

Most operational centers have produced routine dynamical seasonal predictions for more than a decade. A majority of the centers utilize global atmosphere, ocean, land, and sea ice coupled models (one-tier systems) to predict climate anomalies out to lead times of 6-12 months. A few centers, such as the International Research Institute for Climate and Society (IRI¹), use so-called two-tier systems, in which the ocean component is predicted first, and then those predicted sea surface temperatures are used

___________________

¹http://www.wmo.int/pages/prog/wcp/wcasp/clips/outlooks/climate_forecasts.html, accessed January 27, 2016.

as boundary conditions for an atmospheric forecast with lead times out to 3-4 months. IRI has been issuing seasonal climate forecasts from this system since 1997 (Barnston et al., 2010). Examples of one-tier systems include the U.S. National Weather Service’s Climate Forecast System (Saha et al., 2006, 2010), which produces operational predictions with lead times of up to 9 months, and the European Centre for Medium-Range Weather Forecasts’ (ECMWF) seasonal climate prediction system, which is soon to be in its fourth generation.² Other nations have similarly developed seasonal prediction systems that include models developed specifically for this purpose, and the WMO Lead Centre for Long-Range Forecast Multi-Model Ensemble,³ coordinated by the Korea Meteorological Administration and the National Oceanic and Atmospheric Administration (NOAA), collects seasonal forecasts from 12 such seasonal prediction systems (Global Producing Centers) and combines them into multi-model seasonal forecasts that are used by regional and local climate centers around the world (see Box 2.1).

Seasonal prediction has been increasingly prominent at national and international operational centers for several decades. Almost all operational centers produce seasonal predictions at least once per month. Usually, deterministic and probabilistic forecasts of seasonal mean anomalies of surface temperature (atmosphere and ocean) and precipitation are issued for above, below, and near normal values. Seasonal outlooks and ENSO index predictions are also issued based on a combination of dynamical predictions, statistical models, and expert knowledge of teleconnection patterns.

In addition to the ensemble of model integrations into the future, seasonal forecasts require a historical series of model integrations over past decades (these are also called retrospective forecasts, reforecasts, or hindcasts). To create retrospective forecasts, the model configuration is integrating over a large sample of historical cases (i.e., forecasts with known outcomes). These are then used to calibrate future forecasts for biases and reliability as well as to evaluate model skill. An average (or another statistic such as an anomaly) over time is also required for the forecast to be meaningful. For seasonal prediction, this averaging period is usually a season (3 months). A common practice is to produce categorical (e.g., above normal, below normal, or near normal) probabilistic forecasts of seasonal mean anomalies of some basic variables such as surface air temperature and precipitation at monthly lead times (see Figure 3.2).

In addition to seasonal forecasts from operational centers throughout the world, collaborative international efforts aimed specifically at improving seasonal predictions have been critical for advancing forecasting capabilities. Many of these efforts

___________________

²http://www.ecmwf.int/en/forecasts/documentation-and-support/evolution-ifs/cycles/technicaldescription-seasonal, accessed February 3, 2016.

³https://www.wmolc.org/, accessed January 27, 2016.

have a focus on studying predictability and improving forecast skill via multi-model approaches. In addition to the MME seasonal forecasts issued by IRI, the Asia-Pacific Economic Corporation (APEC) Climate Center (APCC) provides routine seasonal MME forecasts to member countries, and the aligned Climate Prediction and its Application to Society (CliPAS) developed a database of retrospective forecasts for prediction and predictability research (Box 2.1). The North American Multi-Model Ensemble (NMME) is an S2S prediction and research effort involving universities and laboratories in the United States, the National Centers for Environmental Prediction (NCEP), and the Canadian Meteorological Center (CMC). NMME started producing seasonal multi-model ensemble forecasts in 2011 (Kirtman et al., 2014). The NMME-2 is now quasi-operational, with seasonal forecasts from the system issued in real time and used as part of NCEP’s operational prediction suite (Box 2.2).

In Europe, the Development of a European MME system for seasonal to interannual predictions (DEMETER) produced a comprehensive set of seasonal retrospective forecasts in order to evaluate MME skill (Palmer et al., 2004). The ENSEMBLES Program⁴ has built on DEMETER to assess how advances in individual seasonal forecast systems translate into reductions in ensemble mean error (Weisheimer et al., 2009). ENSEMBLES has attempted to objectively evaluate uncertainty in MME and other ensemble predictions at seasonal through decadal and longer timescales, including the relative benefits of different model and system configurations. Several model intercomparison efforts, including the WCRP Seasonal Prediction Model Intercomparison Project (SMIP-2),⁵ have also provided valuable insights into model predictive skill and predictability.

Subseasonal

Building on a number of research and experimental efforts over the past decade, subseasonal predictions began in earnest with the establishment of an MJO prediction metric and its uptake by a number of forecast centers (e.g., Gottschalck et al., 2010; Vitart and Molteni, 2010; Waliser, 2011). As of 2009, the outputs from 10 operational centers have been used in an operational manner to provide ensemble predictions of the phase and magnitude of the MJO.⁶ These systems all produce daily ensemble forecasts (sizes range from 4 to 51) with a lead time of 7 to 40 days. Some centers also produce single deterministic forecasts using high-resolution versions of their models. NCEP CPC receives the daily forecasts of zonal wind and outgoing longwave radiation (OLR) from these centers and calculates the predicted MJO index. The forecast products are delivered as plume phase diagrams of the predicted MJO index for each center. APEC’s BSISO forecasts are produced similarly (see Box 2.1).

Many operational numerical weather prediction centers have also recently implemented extended-range (10-30 day) prediction systems that provide building blocks for more useful subseasonal prediction systems (Brassington et al., 2015). Such forecasts are developed in three basic ways: (1) by using forecasting systems designed for seasonal climate predictions, but utilizing only the first 30 to 60 days of the forecast, and paying more attention to the daily or weekly variations rather than the mean monthly or seasonal variations within the forecast; (2) by running an air-sea-ice-land coupled model with a higher resolution than the seasonal system; and (3) by extend-

___________________

⁴http://www.ecmwf.int/en/research/projects/ensembles, accessed January 27, 2016.

⁵ Seasonal Prediction Model Intercomparison Project-2.

⁶http://www.cpc.ncep.noaa.gov/products/precip/CWlink/MJO/CLIVAR/clivar_wh.shtml, accessed January 27, 2016.

ing the lead times of an ensemble medium-range weather forecast using a numerical weather prediction (NWP) model out to lead times of 30 days or more. Methods 2 and 3 produce systems that are independent of the seasonal system. Current operational systems include a 4-times-per-day, 4-member, 45-day lead ensemble from the U.S. National Weather Service (NWS), a 2-times-per-week, 51-member, 46-day lead ensemble from the European Centre for Medium-range Weather Forecasts (ECMWF), a once-per-week, 21-member, 32-day lead ensemble from Environment Canada, and at least eight others. See Appendix B, Table B.2 for more detail on forecasts and the configuration of subseasonal forecast systems.

Many of the same statistical considerations and associated trade-offs cited above for seasonal forecasting (e.g., forecast lengths and averages, ensemble sizes, MMEs, verification periods) are relevant for subseasonal forecasting, although the shorter lead times for subseasonal prediction allow for increased verification instances for a given size observation period. A number of operational centers now compute retrospective forecasts (also known as forecast histories or re-forecasts) as part of the operational forecast process and provide them along with the forecast itself. Frequently computing retrospective forecasts has allowed for continuous improvement of some aspects of forecast systems and permits the calibration to take account of recent events in the current weather/climate regime.

As for seasonal forecasts, an important component of subseasonal forecasts is the retrospective forecast, which is performed over a few years to decades in order to calibrate the real-time forecasts. In contrast to operational medium-range weather prediction, there is a lack of standardization among the centers in producing subseasonal forecasts. For example, some centers produce the forecast once a month, some once a week, some twice a week, and some every day. Some centers start all the ensemble members from the same initial time, whereas others use a time-lagged method where they start with different initial times and therefore different initial analyses. The retrospective forecasts are also produced very differently, for example, some on-the-fly and some with a fixed model version, which may not represent the latest operational configuration. These differences can make it difficult for data exchange, performance inter-comparison, and research. Further details of current subseasonal forecast systems’ resolutions, lead times, ensembles, and other considerations can be found in Appendix B.

Common targets for subseasonal prediction beyond intraseasonal tropical variability (e.g., the MJO and BSISO forecasts mentioned above) include tropical cyclones and extratropical weather. A number of centers provide long-lead information for 1- and 2-week outlooks depicting the probability of United States–based hazardous weather,

as well as tropical cyclone frequency and tropical-midlatitude teleconnection impacts. (e.g., NCEP CPC’s Global Tropical Hazards and Benefits Outlook and U.S. Hazards Outlook⁷). These products are graphically highlighted areas with expected persistent above- or below-average rainfall and regions favorable or unfavorable for tropical cyclogenesis in weeks 1 and 2. The outlooks are based on expert combination of various statistical and dynamical forecasts, including the MJO forecast mentioned above. More recently, the private sector has also become active in developing commercial subseasonal forecasts; for example, the NWS and ECMWF subseasonal forecasts have been combined in a commercial MME by the World Climate Service (Dutton et al., 2013, 2015). Many centers produce forecasts at various lead times of the mean values and probabilities of anomalies averaged over periods of a week or a month. Research efforts to support the further development of subseasonal forecasts are also beginning to develop. The World Meteorological Organization’s (WMO) World Climate Research Program (WCRP) and World Weather Research Program (WWRP) have jointly developed a new initiative on S2S prediction (the S2S Project) (Robertson et al., 2015; Vitart et al., 2012; see Box 2.3 and Chapter 6 for additional information). The thrust of the S2S Project is on improving the subseasonal prediction of extreme weather, such as droughts, heat waves, tropical cyclone development, monsoon precipitation, and subseasonal prediction in polar areas. To do so, the project collects forecasts and retrospective forecasts from a number of operational modeling centers into a common database and disseminates them in delayed mode for research purposes to the science and applications communities.

In addition to the seasonal forecasts discussed in the previous section, NMME-2 (Box 2.2) is beginning to focus on further developing subseasonal forecasts and retrospective forecast databases for the purposes of predictability research and model and forecast system improvement (see Chapter 6 for more details).

Recent Progress in Advancing S2S Forecast Skill

There has been substantial progress in improving the skill of both subseasonal and seasonal forecasts in recent years. Generally, forecast skill for traditional atmospheric variables is still low (see discussion below), but the skill of forecasts of indices of coupled ocean-atmosphere modes of variability is often higher. For seasonal prediction, the Niño 3.4 index, a major indicator of the ENSO, shows useful skill up to 1 year in some models (e.g., Jin et al., 2008; Stockdale et al., 2011). ECMWF System 4 and NCEP

___________________

⁷http://www.cpc.ncep.noaa.gov/products/predictions/threats/threats.php and http://www.cpc.ncep.noaa.gov/products/precip/CWlink/ghazards/index.php, both accessed January 27, 2016.

CFSv2 also capture the year-to-year ENSO variability with fair accuracy, as well as the main ENSO teleconnection pattern in the tropical and extratropical regions (e.g., Kim et al., 2012).

Prediction skill of extratropical modes and patterns such as the North Atlantic Oscillation (NAO) has also recently improved. Scaife et al. (2014a) show good prediction skill

of the winter NAO from the United Kingdom Met Office (UKMO) system, with a correlation in excess 0.6 between ensemble mean and observed NAO index for December-February for forecasts from the start of November (Figure 2.2). They also show that the model is capable of capturing at least qualitatively the observed influence of ENSO on the NAO, as well as the influence of Atlantic heat content, sea ice from the Kara Sea, and the Quasi-Biennial Oscillation (QBO) on NAO seasonal predictability. The performance of this model relative to what was possible a few years ago was likely achieved primarily through reducing biases in the model atmosphere and ocean, leading to an improved model climate (Scaife et al., 2011). Increases in model resolution were also likely important.

Similar progress in forecasting indices has also been made on subseasonal timescales. About 15 years ago, dynamical models had some MJO forecast skill out to 7-10 days (Waliser, 2011), but they performed worse than empirical models that use statistical methods to predict MJO (e.g., Hendon et al., 2000; Jones et al., 2000). Recently, skillful MJO forecasts have been achieved beyond 20 days (e.g., Kang and Kim, 2010; Rashid et al., 2010; Vitart and Molteni, 2010), and the ECMWF in particular has made substantial strides in its MJO forecast (particularly at longer lead times) (Figure 2.3). These advances are due largely to improvements in the representation of physical processes and coupling in the models, better initial conditions, and the availability of better

quality and longer periods of retrospective forecast data to calibrate the forecast (e.g., Vitart, 2014; Vitart et al., 2014). Through better model representation of the teleconnections (i.e., vertical profile of tropical heating and better stratospheric processes and stratosphere/troposphere interactions), improvements in the prediction of tropical phenomena such as MJO and ENSO have also led to some increase in prediction skill in the extratropics and also of ocean variables and phenomena such as tropical cyclone counts (e.g., Vitart et al., 2007a).

Despite this progress, traditional measures of S2S forecast skill such as anomaly correlation and root mean square error when applied to systems such as NCEP’s CFSv2

(Climate Forecast System) or GEFS (Global Ensemble Forecast System) indicate little skill from week 2 and beyond, even with the application of temporal averaging. As mentioned above, MME forecasts have improved forecast skill of such traditional atmospheric variables in some cases, but even for seasonal forecasts, large gaps persist across specific regions and seasons, especially for precipitation. For example, skill of ENSEMBLES multi-model forecasts of boreal winter conditions is good in the tropics and over oceans, particularly for temperature; however, skill over land, especially outside of the tropics, is limited (Figure 2.4). Although the quality is slightly better over land areas with strong ENSO teleconnections (e.g., North America in boreal winter), skill is still low in certain areas, for example, over most of Europe during the winter (Doblas-Reyes et al., 2013).

Enhanced forecast skill is sometimes possible during specific windows of time in specific regions. Skill to 20 days is possible, for example, during specific MJO phases (Lin et al., 2010; Rodney et al., 2013). Such contingent improvements in forecast skill, along

with generally low skill for traditional atmospheric forecast variables over large areas and time windows, highlights the importance and promise of so-called forecasts of opportunity (see Chapter 4).

The recent extreme low sea ice extent in summer in the Arctic prompted the Sea Ice Outlook⁸ to begin gathering seasonal forecasts in 2008. The first published skill of a retrospective forecast of sea ice extent appeared soon after, in 2013 (Merryfield et al., 2013b; Sigmond et al., 2013; Wang et al., 2013a). Dynamical models exhibit skill at these lead times, but their skill is still substantially below estimates of perfect-model forecast skill (see also Figure 4.2). Furthermore, no study has yet published an evaluation of S2S forecast skill at the regional or local scale for sea ice variables such as concentration, thickness, or ice-type, which are likely to be useful to forecast users.

In summary, over the past two decades substantial progress has been made understanding some of the physical drivers for S2S prediction, and operational centers have made some progress in improving S2S forecast skill. While prediction skill for indices of climate modes such as the MJO and ENSO has improved more dramatically, current operational skill is low for many traditional weather and climate variables. S2S forecasts for Earth system variables outside traditional weather and climate forecasts are less well developed, but have also been advanced by the development of coupled Earth system prediction systems.

The growing interest by the science community and operational forecast centers to develop and implement many of the projects and experiments described above, in addition to recent progress in S2S predictability research and operational predictions, illustrates the research priority and expectations associated with S2S timescales. However, an associated U.S. national research agenda aimed at strengthening the contributions of S2S forecasts to public and private activities has not yet emerged.

Finding 2.1: Although there has been considerable progress in S2S forecasting over the past several decades, many opportunities for improvements in S2S forecast skill remain.

___________________

⁸http://www.arcus.org/sipn/sea-ice-outlook, accessed January 27, 2016.

This page intentionally left blank.