Windows of Opportunity for Skillful Forecasts Subseasonal to Seasonal and Beyond

,

Processes that provide forecast opportunities The intermittency of opportunities for skillful long-lead predictions is already evident from long-lead weather forecasts, with those initiated from some atmospheric flow configurations being more skillful than others.Intrinsic predictability levels of different atmospheric states can be estimated from forecast ensembles. 2In reliable and well-calibrated 3 prediction systems, the spread among ensemble members quantifies forecast uncertainty and the underlying predictability.At times when the forecast uncertainties are relatively small we deduce that the atmosphere is more predictable; higher predictability means a window of opportunity for more skillful forecasts.If we can identify processes and conditions that lead to these predictable conditions, we can potentially provide more useful S2S and S2D forecasts.
Midlatitude atmospheric processes.Midlatitude atmospheric processes can lead to circulation patterns or weather regimes that affect predictability.For example, studies of flowdependent verification based on Euro-Atlantic midlatitude weather regimes indicate that European blocking is the regime associated with the least accurate medium range weather forecasts over Europe, while the negative phase of the North Atlantic Oscillation (NAO) leads to the most skillful predictions (Ferranti et al. 2015).In another example, forecasts for Europe made when the westerly jet across the Atlantic is displaced to the north (Atlantic ridge) tend to be less accurate (Frame et al. 2013).The physical processes that can reduce forecast skill are generally associated with atmospheric instabilities and weather regime transitions.At longer ranges, weather regimes play an increasingly important role.Their low frequency, large spatial scale, and sensitivity to tropical and stratospheric conditions point to processes that can matter for subseasonal predictions and beyond.For example, a persistent winter ridging was associated with California drought during 2013-16, even though the underlying causes have not been robustly established (e.g., Teng and Branstator 2017;Swain et al. 2017).
Quasi-stationary Rossby waves during boreal summer can also condition the skill of subseasonal predictions and impact the predictability of boreal summer climate extremes including heat waves, flooding events, and short-term droughts.Recent examples of such linkages include the 2003 European heat wave, the 2010 Russian heat wave, and the 2012 flash drought in the U.S. Great Plains (e.g., Schubert et al. 2011;Wang et al. 2014).Similar mechanisms are also at play in the Southern Hemisphere during its summer and similarly impact extremes such as the 2009 Australian heat wave (Parker et al. 2014).These waves often manifest as nearly circumglobal teleconnections in which the summer jets act as waveguides (e.g., Ding and Wang 2005;Beverley et al. 2019), sometimes resulting in a synchronization of extremes over distant parts of the globe.While such waves appear to be 2 A forecast ensemble consists of several multiple independent forecasts (ensemble members) with perturbed initial conditions and/or perturbed model physics. 3In a reliable ensemble, forecast probabilities match the observed frequencies.An ensemble system can be made more reliable through statistical calibration, which aims to relabel the forecast probabilities, so that they match the observed frequencies.

E611
mainly forced by submonthly vorticity transients with limited predictability (Schubert et al. 2011), there is mounting evidence that tropical processes (e.g., Newman and Sardeshmukh 1998;Watson et al. 2016), land forcing (e.g., Koster et al. 2014), or even internal atmospheric resonance processes (Kornhuber et al. 2017), can amplify and sustain these planetary waves.Such processes could increase the predictability of associated extremes, such as heat waves over the United States and other continental regions, well beyond 2 weeks.Taking full advantage of the potentially enhanced predictability associated with quasi-stationary Rossby waves at subseasonal and longer time scales requires better understanding of their growth and persistence along with improved simulation and prediction of associated processes.
Tropical processes.Tropical processes involving interactions within the atmosphere, and between the atmosphere and the ocean, provide a primary source of predictability on subseasonal time scales and beyond.Regions of organized deep convection in the tropics excite large-scale atmospheric teleconnection patterns through a variety of mechanisms: the linear dispersion of Rossby waves (e.g., Hoskins and Karoly 1981), modification of the background flow and Rossby wave breaking (e.g., Moore et al. 2010), and planetary wave propagation into the stratosphere (e.g., Smith and Kushner 2012;Plumb 2010).The potential for S2S predictability from tropical sources lies in the "slow" and predictable evolution of large-scale tropical convection patterns, the 1-2-week time scale for the extratropical atmosphere to respond fully to tropical convective heating (e.g., Matthews et al. 2004), and the several weeks for stratospheric-tropospheric interactions to occur.
The MJO (Madden and Julian 1994), when active, is a tropical phenomenon that provides a major source of predictability on subseasonal time scales.During an MJO event, regions of large-scale convection propagate eastward along the equator on a 30-60 day timeframe (Matthews 2008), generating Rossby waves along the way that propagate poleward, producing a delayed response in the extratropics 1-2 weeks later.The MJO impacts over North America, in particular, are strongest and most consistent for MJO phases having an east-west dipole of convective heating in the Indian and western Pacific Oceans (Ting and Sardeshmukh 1993;Lin et al. 2010;Tseng et al. 2018) and may, when modulated by ENSO (Arcodia et al. 2020), potentially persist up to 5 weeks (Riddle et al. 2013).More broadly, the MJO impacts S2S predictability across the tropics, including the Maritime Continent; the Indian, Australian, and West African monsoon regions (e.g., Wheeler et al. 2009;Ventrice et al. 2011;Taraphdar et al. 2018); and parts of China (e.g., Song and Wu 2019).Especially promising is that current dynamical models now exhibit significant MJO skill beyond 30 days (Vitart 2017).Recent studies (Benedetti and Vitart 2018) show that there may be additional predictability associated with the MJO effect on atmospheric aerosols.While the MJO creates opportunities for skillful S2S forecasts, important caveats exist.In particular, it is now clear that not all phases of the MJO provide equal forecast opportunities, and only specific regions of the world are affected during certain times of the year (e.g., Garfinkel and Schwartz 2017).There is also evidence that the MJO effects are modulated by other climate phenomena (see below), and that models still exhibit significant weaknesses in their representation of the MJO and its teleconnections.
ENSO, a tropically coupled ocean-atmosphere climate phenomenon, is another leading source of predictability.Once active, its effects are felt on the S2S to multiannual time scales (DiNezio et al. 2017).During warm ENSO events, also known as El Niño, the eastern tropical Pacific is anomalously warm, while the eastern Indian Ocean and western tropical Pacific are anomalously cold.Opposite conditions exist during cold ENSO events, known as La Niña.These conditions evolve over the course of months to years and tend to be quite predictable, depending on the phase of ENSO (e.g., Zheng et al. 2016).Studies have shown additional predictability of El Niño events following large volcanic eruptions due to the ocean-dynamics thermostat mechanism described by Clement et al. (1996) and Eddebbar et al. (2019).Sea E612 surface temperature anomalies associated with ENSO induce patterns of organized anomalous tropical convection from which emanate global teleconnections via Rossby wave propagation, midlatitude jet alterations, and other mechanisms.ENSO predictability influences forecast skill from months to years, for example, influencing MJO teleconnections on subseasonal time scales (e.g., Leroy and Wheeler 2008), while modulating the midlatitude background flow on seasonal and longer time scales.ENSO-related predictability stems from the duration, strength, and location of specific ENSO events and the lags of remote teleconnections.ENSO impacts, and therefore ENSO-related predictability, are regionally and seasonally dependent, particularly outside the tropics.
It is particularly promising that current climate prediction systems are now able to predict the development of an ENSO event with some skill out to at least a year ahead depending on the initial conditions (Barnston et al. 2017).On the other hand, the ability of models to predict the details (e.g., spatial structure and amplitude of the SST anomalies) and exact timing of individual ENSO manifestations including the associated global teleconnections, is still limited at those forecast leads (e.g., Kim et al. 2012;Wang et al. 2009).As such, we are still far from taking full advantage of ENSO-related predictability.
Interplay of troposphere-stratosphere processes.Tropical processes providing opportunities for S2S/S2D predictability extend to interactions with the stratosphere.Specifically, the stratospheric quasi-biennial oscillation (QBO; Baldwin et al. 2001) can modulate the predictability of the wintertime MJO and its teleconnections (e.g., Marshall et al. 2016;Zhang and Zhang 2018) as well as ENSO teleconnections (Garfinkel and Hartmann 2010).The QBO oscillates at a longer time scale than the MJO, around 28 months, with alternating easterly and westerly equatorial wind states that develop in the upper stratosphere and propagate downward until dissipating at the tropical tropopause.The modulation of midlatitude S2S extremes by the MJO depends on QBO phase.This effect has been demonstrated in the context of U.S. West Coast atmospheric rivers (ARs; Baggett et al. 2017).Times when the QBO and MJO are both active provide windows of opportunity for improved forecasts of S2S extremes.Skillful empirical forecasts using only the MJO and QBO to forecast S2S extremes have already been realized (e.g., Mundhenk et al. 2018;Nardi et al. 2019, manuscript submitted to Wea. Forecasting).
Stratospheric processes associated with the winter stratospheric polar vortex also provide opportunities for enhanced S2S predictability.The polar vortex describes the climatological circumpolar stratospheric westerly winds that circle the pole in wintertime (Waugh et al. 2017).About once every other year the Northern Hemisphere polar vortex is disrupted by tropospheric wave forcing and also by internal resonance of stratospheric waves (Birner and Albers 2017).When this happens, it can either be displaced off of the pole or split into two smaller vortices, and the stratosphere rapidly warms in a few days in an event called a sudden stratospheric warming (SSW; e.g., Butler et al. 2017).The anomalous winds and temperatures then slowly descend to the tropopause via wave-mean flow interactions and ultimately influence the tropospheric jet stream over a period of weeks to months (Baldwin and Dunkerton 2001).Typically, these impacts manifest as an increase in synoptic cyclone tracks over Europe and a decrease over the Russian Arctic, a tendency for colder-than-normal temperatures over Eurasia and the eastern United States, wetter conditions over southern Europe, and warmer temperatures over eastern Canada and subtropical Asia and Africa (see Fig. 1).In the Southern Hemisphere, stratospheric polar vortex variations have an impact on Australian hot and dry extremes (Lim et al. 2019).Deterministic predictability of SSW events is limited to 10-20 days (e.g., Karpechko 2018).Probabilistic SSW prediction likely depends on the state of the MJO, QBO, and ENSO, among other factors (e.g., Domeisen et al. 2019).ENSO, for example, can modulate both the strength of the polar vortex via its stratospheric E613 teleconnection, and the impact of SSWs on surface weather over the North Atlantic via its tropospheric teleconnection (Domeisen et al. 2020).
Once a SSW has occurred, the associated extreme stratospheric conditions can constrain the evolution of the atmosphere through wave-mean flow interactions.Since recovery to climatological conditions after a SSW occurs on radiative time scales (several weeks in midwinter) the impacts of SSWs are long-lived, potentially allowing increased surface predictive skill 3-6 weeks after an event (Sigmond et al. 2013).When the anomalous stratospheric conditions are simulated in a numerical weather prediction model, the extended period of anomalous wave-mean flow interactions can also shift a forecast away from climatology (Scaife et al. 2016) and into weather regime states with enhanced predictability (e.g., Charlton-Perez et al.

E614
2018).Other types of extreme stratospheric variability, such as vortex intensification, may also provide opportunities for predictability (Tripathi et al. 2015).
Land-atmosphere interactions.The land surface responds to sustained anomalies in the atmosphere on S2S time scales and can thus amplify and sustain anomalies to enhance predictability (Koster and Suarez 2001;Dirmeyer et al. 2018).Positive land-atmosphere feedback processes may exacerbate associated extremes (Koster et al. 2014) and enhance the persistence of major droughts (e.g., Vautard et al. 2007) and heat waves (e.g., Ford and Quiring 2014).Beyond these extremes, the land surface plays a role in S2S climate variations.In fact, any time and place that three specific conditions are met, the process chain for land-atmosphere feedbacks offers an opportunity for enhanced predictability (e.g., Dirmeyer et al. 2015).The first condition is sensitivity of the atmosphere to land surface state variations via changes in surface fluxes; the second is sufficient magnitude of variability in land surface states to drive a significant atmospheric response.These first two conditions are both in force in the so-called "hot spot" regions of land-atmosphere coupling such as the Great Plains of North America, the Sahel of Africa, the Indus River basin, much of Australia, southern Africa, and the Pampas of South America (Koster et al. 2004).The third necessary condition for land-driven predictability on S2S time scales is memory or persistence of land surface anomalies.The greatest memory tends more toward arid regions, but memory can also persist and deliver anomalies across seasons when coupling emerges (e.g., Guo et al. 2011).Effects may be nonlocal, as the same circulation systems that can initiate anomalies in the land surface can serve to advect their effects downwind to other locations (Koster et al. 2014;Koster et al. 2016).When these three factors are combined and accounted for in the forecast models, there is opportunity for improved S2S prediction skill stemming from land-atmosphere interactions (e.g., Koster et al. 2011;Dirmeyer et al. 2018).
Mid-to high-latitude ocean-atmosphere interactions.In addition to ENSO and MJO related ocean-atmosphere processes, other oceanic dynamics and feedbacks outside the tropics more broadly provide opportunities for enhanced S2S predictability.For instance, Arctic sea ice anomalies have been linked with high-to midlatitude atmospheric anomalies (e.g., Alexander et al. 2004) and enhanced predictive skill (He et al. 2018).During years with low seasonal sea ice concentrations (when there is more heat loss from more exposed open water), the north-south differences in atmospheric temperatures across the Barents Sea are reduced.These conditions have been linked to wintertime cyclones traveling farther south into western Europe, instead of their tendency to move eastward toward Siberia, as well as more frequent cold winter extremes at midlatitudes (Petoukhov and Semenov 2010;Inoue et al. 2012;Mori et al. 2014).There is evidence for predictive skill for high-latitude climate due to midsummer sea ice extent anomalies (He et al. 2018).Predictability associated with sea ice processes, derives both from lead-lag relationships between the sea ice anomalies and the high-to midlatitude response, and also from our increasing capability to predict sea ice anomalies themselves a season ahead (e.g., Stroeve et al. 2014).
In another example, oceanic processes associated with western boundary currents, such as the Kuroshio and the Gulf Stream currents, are being explored as sources of atmospheric predictability (e.g., Smirnov et al. 2015).Such currents, bringing warm water from the tropics poleward, are associated with sharp oceanic temperature gradients.At the exit of such currents, are mesoscale eddies which carry anomalous warm SST as they propagate out of the formation region.Hence the evolution of currents and eddies in areas of sharp temperature gradients is associated with ocean-atmosphere heat flux anomalies that can potentially affect the weather on S2S time scales (Chelton and Xie 2010;Jia et al. 2019).For example, Haarsma et al. (2019) showed that resolving oceanic mesoscale features (eddies and fronts) near the

E615
North Atlantic storm track translates into enhanced seasonal prediction skill compared to forecast systems that do not resolve oceanic mesoscale features.
Beyond S2S.While our primary focus here is on the S2S prediction problem, the windows of opportunity framework similarly applies to longer time scales.There is of course no clear spectral gap between variability at S2S time scales and the variability at decadal and longer time scales.In fact, while some phenomena already discussed for S2S, such as the MJO, have spectral peaks at subseasonal time scales, several others have substantial impacts beyond one season (e.g., ENSO or QBO).There are other processes that operate on multiyear and even multidecadal time scales that are particularly relevant to prediction at these longer time scales.Processes that involve slow ocean dynamics or long-lived atmospheric constituents, which affect S2S forecasts, can also lead to enhanced predictability on longer time scales, provided we can adequately capture underpinning processes in forecast models.As an example, the Atlantic meridional overturning circulation (AMOC) evolves on multidecadal time scales, and may provide potential predictability if it can be adequately simulated in prediction systems (e.g., Zhang and Zhang 2015).Also, particular Atlantic conditions may be especially conducive to higher prediction skill.For example, when the Atlantic Ocean heat transport around 50°N is strong at the initialization of a hindcast, SST skill in the northeast Atlantic at lead years 2-9 is significantly increased (Borchert et al. 2019).Also, when focusing on decadal hindcast skill in the North Atlantic subpolar gyre, skill is higher during strong multiyear trends, especially during the warming period of the 1990s, and lower in the absence of such trends (Brune et al. 2018).
The sudden injection of aerosols into the troposphere and stratosphere by a major volcanic eruption can alter climate for years (e.g., McCormick and Veiga 1992), while also influencing shorter term weather (Reale et al. 2011;Reale et al. 2014).While volcanic eruptions are not predictable, once they occur and their impact is captured in the initial conditions, they can enhance predictability across a variety of time scales (Benedetti and Vitart 2018).In addition, anthropogenic aerosols may also present a source of decadal predictability (Bellucci et al. 2017).Similarly, long-term increases in greenhouse gas concentrations and associated warming trends are a major source of predictability for surface air temperature and other surface meteorological variables, affecting the skill of climate forecasts at S2S time scales and beyond (e.g., Luo et al. 2011;Boer et al. 2013).For instance, Arctic sea ice anomalies due to polar amplification of long-term surface temperature increase have been linked to changes midlatitude weather (e.g., Cohen et al. 2014;Zhang et al. 2012).More broadly, there is a growing body of research examining the effects of climate variability and change on the predictability of extremes across the globe (e.g., Herring et al. 2019).

From windows of opportunities to forecast tools
Knowing which processes (e.g., atmospheric circulation, climate patterns, ocean dynamics, etc.) offer predictability is relevant for forecast interpretation as well as the development of new forecast tools.In fact, this knowledge must be reflected in our forecast models if we are to use them to identify opportunities for enhanced skill.
A priori identification of forecast opportunities.Providing an objective measure of the expected forecast accuracy is very valuable (e.g., Kalnay and Dalcher 1987;Molteni and Palmer 1991).There are several examples of flow-dependent verification based on weather regimes.This approach has been applied to Euro-Atlantic weather regimes, as discussed above (Ferranti et al. 2015).Also, Vigaud et al. (2018a) used daily wintertime 500-hPa geopotential height fields to classify four weather regimes over the Pacific-North American sector that can be used to identify windows of high forecast opportunity in the midlatitudes by highlighting Unauthenticated | Downloaded 09/24/23 02:30 PM UTC E616 persistent episodes in real time and in the forecasts.A prototype system has been set up at the International Research Institute for Climate and Society (IRI) to do this on a daily basis using the NOAA CFSv2 model.Forecast characteristics are plotted according to target day and lead time, permitting the visualization of a large number of forecast runs in a way that emphasizes forecast timing and consistency from one forecast to another (Tippett et al. 2012;Carbin et al. 2016).Similarly, cluster analysis of daily circulation/convection types and related modes of subseasonal predictability, has also been applied over South Asia (Moron et al. 2012), the Maritime Continent (Moron et al. 2015), West Africa (Vigaud and Giannini 2018), the Caribbean (Vigaud and Robertson 2017), and South America (Muñoz et al. 2015(Muñoz et al. , 2016;;Doss-Gollin et al. 2018).
Tools are being developed to predict the skill of climate forecasts during windows of opportunity.Because of the fundamentally nonlinear and chaotic nature of the atmosphere, inherent nonlinear interactions act so rapidly that much of their effects are unpredictable on S2S time scales and beyond.What predictability remains can be well approximated by much simpler linear dynamics and a residual random "noise" (e.g., Hasselmann 1976;Penland and Sardeshmukh 1995).This approximation, empirically determined using the linear inverse model (LIM) technique (Penland and Sardeshmukh 1995), generates S2S forecasts about as skillful as those from operational dynamical models at both NCEP and ECMWF (e.g., Newman et al. 2003;Albers and Newman 2019).The LIM technique also allows identification of high skill cases or forecasts of opportunity ahead of time (Albers and Newman 2019; see example in Fig. 2).Such higher skill depends upon the similarity of the forecast initialization to patterns

E617
leading to the most rapidly amplifying anomalies within the LIM.For example, Newman et al. (2003) found that week 3 forecast skill of the 250-hPa extratropical streamfunction was largely related to how strongly the initialization matched varying combinations of only three patterns, which together represented ENSO, the MJO, and the Pacific-North American (PNA) teleconnection pattern.
North American week 3-4 temperature and precipitation.The combined influence of ENSO and the MJO significantly impacts the wintertime general circulation over North America for lead times up to at least 4 weeks.Based on these relationships, probabilistic 2-m temperature and precipitation forecasts over North America are now generated solely on the basis of the climatological linear trend associated with each variable along with their statistical relationships with the initial state of the MJO and ENSO (Johnson et al. 2014).Such forecasts, now used as part of the NOAA operational forecast suite (experimentally for precipitation), exhibit substantial skill for some regions and some initial states of the MJO and ENSO out to a lead time of approximately 4 weeks.The highest temperature forecast skill tends to follow MJO phases 2-3 and 6-7 owing to the strong extratropical response to the dipole convective heating anomalies associated with these phases (Lin et al. 2010;Tseng et al. 2018), although the precise MJO phases and regions with maximum skill are sensitive to the state of ENSO (Johnson et al. 2014).
Risk of severe cold spells over Europe.Forecast skill of the NAO has been shown to be conditioned on the presence of an MJO or SSW in the initial conditions (e.g., Sigmond et al. 2013;Tripathi et al. 2015;Ferranti et al. 2018).By inferring surface weather through large-scale flow patterns, forecasters try to identify the spatial extent of temperature anomalies that are predictable at subseasonal time scales.For example, reliable subseasonal forecasts of the NAO and blocking are used to assess the risk of severe cold spells over Europe (Ferranti et al. 2018).Current S2S models can deliver skillful forecasts for some large-scale patterns 2 weeks ahead, and longer in certain cases.This suggests that subseasonal predictions have the potential to support early warnings of severe cold events over Europe.The success of predicting weeks ahead the changes in large-scale flow leading to cold spells depends on the type of regime transitions.For example, the ECMWF ensemble provides reliable probabilities of cold spells associated with the establishment of Greenland blocking (NAO−) well beyond 1 week (e.g., Ferranti et al. 2018).The predictive skill of such events can be significantly enhanced during MJO activity via tropical-extratropical teleconnections.Using the ensemble spread rate of change 4 with lead time as a measure of predictability, it is possible to show that the NAO− circulation regime exhibits a higher level of inherent predictability than the European blocking regime (Ferranti et al. 2018).
Atmospheric rivers and extreme weather activity.As previously discussed, two modes of tropical variability offer windows of opportunity to forecast aspects of North America's extreme weather at subseasonal lead times.Specifically, the MJO has been shown to modulate springtime hail and tornado activity (Thompson and Roundy 2013;Barrett and Gensini 2013), along with west coast snowpack (Guan et al. 2010), while the MJO and QBO have been shown to modulate wintertime AR frequencies (Baggett et al. 2017).Baggett et al. (2018) developed an empirical model using the current state of the MJO to forecast hail and tornado activity over the U.S. plains and found skillful forecasts of opportunity out to 5 weeks.A similar empirical model was developed by Mundhenk et al. (2018) that utilized both the states 4 The ensemble spread, of a well-constructed ensemble prediction system, is an indicator of forecast uncertainties and the rate at which the spread grows is an estimate of predictability.
As the ensemble spread grows rapidly with lead time, so does the forecast uncertainty, indicating reduced predictability.Conversely, as the ensemble spread grows slowly with lead time, so does the uncertainty, implying a smaller reduction in predictability.
of the MJO and QBO to predict AR activity along the U.S. West Coast.This empirical model also produced skillful forecasts with lead times out to 5 weeks for specific combinations of MJO and QBO phases [see Fig. 3 for an example of derived forecast tools following Nardi et al. (2019, manuscript submitted to Wea. Forecasting)].Dynamical models, by comparison, show skillful forecasts only out to ~2 weeks for similar variables (e.g., Carbin et al. 2016).

Atmospheric rivers and western U.S. water availability. The California Department of Water
Resources and its research partners are exploring opportunities for skillful S2S forecasts to aid in the management of western U.S. water availability and related hazards under extreme conditions.Efforts have mainly targeted wintertime ARs as a major source of freshwater input and on ridging as a dominant way ARs are restricted.Relationships between weekly AR occurrence, ridging, and certain phases of ENSO, the Arctic Oscillation, the PNA, and the MJO (DeFlorio et al. 2018;Guirguis et al. 2018) have formed the basis for predictability considerations.An experimental prediction framework is being developed to predict weekly AR frequency (DeFlorio et al. 2018(DeFlorio et al. , 2019) ) based on hindcast and forecast information from the S2S Prediction Project (Vitart et al. 2017).Results indicate conditional useful skill 1-3 weeks ahead.Forecast tools under development focus on ridging types and persistence, as they relate to distinct characteristics and opportunities for enhanced predictability and skill; they also

E619
examine the integrated water vapor transport in a given week associated with ARs over the U.S. West Coast.Guirguis et al. (2018) identified circulation anomaly patterns associated with AR activity along the North American west coast and showed how seasonal variation in the prevalence of these weather types is associated with interannual variability in AR landfalls.Such knowledge is providing the basis for new experimental tools built on historical analogs.
Climate extremes.Operational centers currently produce forecasts for extremes or hazards on weather time scales, such as severe weather or heat extremes.Generally, such operational forecasts have not been extended into climate time scales.Examples of climate extremes include monthly or seasonal temperatures or accumulated precipitation amounts that have occurred infrequently in the climate record, which may be defined using the tails of the observed distribution.These extremes can be the result of internal variability and can also be modulated by forcings such as increasing greenhouse gases.An example of a forecast of opportunity for climate extremes is that of flooding probabilities in Peru conditioned on ENSO states, given the heavy rains that typically accompany El Niño events.These forecasts of opportunity are now informing the development of potential early action plans.At seasonal forecast leads, these early action plans would be triggered with a high degree of confidence if several forecast criteria are met, such as El Niño SST anomalies (Niño region "1+2") exceeding 3°C or ensemble probabilities for precipitation above the 90th percentile exceeding 40% (Bazo et al. 2019).Interestingly, some studies have identified that subseasonal and seasonal climate models can more skillfully predict events with greater anomalies than events with anomalies closer to the mean (Becker et al. 2013;Becker and van den Dool 2018).These results suggest that useful outlooks for extreme events could be developed.

Challenges and benefits
While experimental tools that build on windows of opportunity are being developed, such an approach is not without challenges.At the outset, it is a question of culture.Recognizing that a forecast tool has scientific standing and practical value even if it provides information only when the opportunity arises is not customary.On the user side, both forecast centers and forecast users are accustomed to expecting forecasts on time, all the time.Adapting to an approach that provides probabilistic forecasts with information only in certain conditions requires a significant mindset shift.In seasonal forecasting, the approach to skill intermittency has been to issue probabilistic forecasts calibrated in order to yield climatological probabilities in the absence of a forecast signal (Mason et al. 1999).This approach provides users with a forecast "all the time" which can be acted upon when and where the probabilities deviate from climatology, provided the forecasts are well calibrated to issue reliable probabilities (e.g., for tercile categories or exceedance/nonexceedance of user-defined thresholds; Barnston and Tippett 2014).This approach has recently been applied to subseasonal forecasts of precipitation and temperature (e.g., Vigaud et al. 2017aVigaud et al. , 2019)), and discussed in the context of monsoonal climates (Vigaud et al. 2017b;Robertson et al. 2019), and East Africa/West Asia (Vigaud et al. 2018b).On subseasonal time scales, the NOAA Climate Prediction Center generates probabilistic forecasts following similar approaches for the calibration of dynamical model output (Unger et al. 2009;Strazzo et al. 2019).However, for subseasonal forecasts of particular meteorological variables such as precipitation, calibrated probabilities will often deviate very little from climatology, dependent on the region, season, and lead time.Furthermore, such methods of calibration may underestimate reliable probabilities of forecasts during periods of enhanced predictability.These could potentially be targeted more purposefully as part of a forecast of opportunity framework approach.
Moving toward opportunity-based forecast approaches may not always be practically achievable.Even when predictive linkages between processes and meteorological outcomes E620 have been theoretically established, it is another matter to build forecast tools that exploit such relationships.Among other things, this requires a large enough sample of past events to build statistically sound relationships as the basis of the new prediction tools.Building forecasts of opportunity automatically decreases the sample size as only a particular subset of days is being considered as training data.Sample size issues become particularly challenging for forecasts of events that happen rarely.
Advancing an opportunity-based forecast approach requires both foundational and applied research.This includes furthering our understanding of the processes that can lead to opportunities for enhanced predictability and forecast skill.In turn, this requires having robust observations as the basis of research, including key surface variables such as ocean mixed layer depth, sea ice thickness, snow water equivalent, soil moisture, etc. (NASEM 2016(NASEM , 2018)), along with improved long-term reanalysis products that are consistent across Earth's climate system.Research outcomes could point to new predictability sources to help alleviate overreliance on well-established processes such as ENSO, which have limited relevance and applicability.Research is also needed to further develop prediction systems and statistical methodologies that build on observations, process understanding, and models for new opportunity-based forecasts.There is also the need to develop evaluation diagnostics and metrics that optimally evaluate such forecasts to inform the development of opportunitybased prediction tools.
Last, but importantly, communication strategies need to be tackled.Given the intermittency in the occurrence of forecast windows of opportunity, their format may differ from typical weather forecasts made by operational centers.However, there are already examples of this approach in weather forecasting, where predictions of the risk of weather hazards are issued only when that risk is enhanced using the "watch" and "warning" progression.Another possible format is currently in operation at the Climate Prediction Center.Probabilistic hazard maps of extreme temperature, wind, and precipitation are issued for a lead time of 2 weeks.Rather than pinpointing exact locations and times as typical weather advisories and warnings do, these maps subsume broad regions over several days where and when probabilities of hazardous weather are elevated.Forecasts of opportunity necessitate a different kind of response from users, who must understand the uncertainty associated with the forecast event.
The development of capacity in forecasts of opportunity should be accomplished through the collaboration with potential users to ensure that forecast information is actionable (White et al. 2017).
Despite the challenges, the potential benefits of exploring opportunity-based forecast approaches are tantalizing, especially considering the current alternative of having forecasts with limited skill beyond weather time scales (Malloy and Kirtman 2020).Applying a window of opportunity approach to the examination of observational data and forecast system output could reveal new useful relationships and skill when and where it was thought there was none.This in turn could open the door to a new set of forecast products.With an approach that explicitly builds on specific opportunities, communication, and dissemination tools, users' expectations could be more adequately met, as they are better aware of when and where to expect skill and if the prediction is actionable.
Given that there is still untapped potential, it is safe to expect that in the future the range of forecast opportunities will expand beyond what is currently available.There is much that is unknown.For example, it is conceivable that predictive processes exist that have not yet been detected based on observations or are not fully understood.Moreover, these processes may not yet be well captured in our models and prediction systems that have limited resolution and parameterized physics.There are experimental statistical forecasts tools with consistent forecast skill in certain situations and for certain quantities at a level not yet achieved in our dynamical models based on current process understanding (e.g., for western E621 AR activity; Mundhenk et al. 2018).This is not surprising and, most importantly, it is also suggestive that further dynamical models' skill improvement is to be expected and purposefully sought after.For example, there is much variability in the ocean beyond ENSO that we are still trying to understand, simulate, and exploit for prediction purposes.Indeed, there are recent studies pointing to coupled processes both in the tropics (e.g., the MJO, the diurnal cycle, and monsoons) and mid-to high latitudes (e.g., sea ice, mesoscale frontal interactions) as important to modulate intraseasonal variability (e.g., Taraphdar et al. 2018;Li and Sriver 2018;Alexander et al. 2004;Smirnov et al. 2015;DeMott et al. 2015;Chelton and Xie 2010;Jia et al. 2019).This is even truer when considering the broader set of Earth system components and interactions that could lead to enhanced predictability.It is logical to build opportunitybased forecasts considering the processes that we understand create favorable conditions for enhanced forecast skill and improve our dynamical prediction systems so they optimally represent them.However, empirical models or dynamical models based on forecast output models (i.e., hybrid models) have already shown promise (e.g., Dobrynin et al. 2018) and may continue to be a synergetic innovative way forward for exploring predictability potential, until underpinning mechanisms can be more fully understood and better simulated by dynamical forecasts systems.

Fig. 1 .
Fig. 1.SSW events affect (a)-(c) cyclone track frequency and (d)-(f) surface air temperature anomalies for at least a month after they occur.Cyclone track frequency (cyclones per week) is calculated using Hodges (1994) tracking algorithm within 250 km of a given location for subsets of 1980-2015 January, February, or March days in the MERRA-2 data.(a) Cyclones that occurred within 30 days after a SSW event and (b) cyclones not within 30 days after a SSW event; (c) the difference between (a) and (b) with significance hatched for regions excluded by the 80% confidence interval smoothed by a 9-point smoother.(d)-(f) As in (a)-(c), but for daily 2-m temperature anomalies (K), with significance in (f) hatched at the 95% interval.Significance tested with n = 1,000 bootstrap resampling without replacement.

Fig. 2 .
Fig. 2. The LIM empirical forecast model applied to identify MSLP winter forecasts of opportunity in the Pacific basin for up to 6 weeks lead time.The figure shows median hindcast skill (1999-2010) of wintertime Pacific basin MSLP weekly-averaged anomalies predicted by operational NCEP CFS and ECMWF IFS models and by the LIM model, for forecast leads ranging from 1 to 6 weeks.For each forecast system, two categories are shown: one where the LIM expects skill will lie in the upper 10% of all hindcasts (darker bars) and a second consisting of the remaining 90% of hindcasts (lighter bars).Whiskers denote uncertainty due to the small hindcast sample size (based on bootstrap confidence intervals).Skill is measured by the pattern correlation of the forecast MSLP anomaly with its verification in the region 20°-60°N, 120°E-120°W.[Adapted from Albers and Newman (2019).]

Fig. 3 .
Fig. 3. MJO conditions opportunities to forecast atmospheric river activity with 3-4 weeks lead time.The figure shows the frequency of forecasts of opportunity calculated over all instances when the MJO is active at initialization during January-March.Opportunities are defined as phase and lead combinations for which the skill is significantly better than a random forecast at the 95% confidence level.Frequencies are calculated over all MJO phase and lead combinations from week 3 through week 4 (lead days 15-28).Atmospheric rivers are identified using a modified version of the Mundhenk et al. (2016) scheme (see Ralph et al. 2019), with the predictand defined as the 5-day forward running mean atmospheric river anomaly from climatology.The empirical prediction scheme is detailed in Nardi et al. (2019, manuscript submitted to Wea.Forecasting) and based on Mundhenk et al. (2018) and Johnson et al. (2014).Users can create similar figures using the web app: http://barnes.atmos.colostate.edu/S2SPredictionModel/.(Figure courtesy of Kyle Nardi.)