Second International Conference on Subseasonal to Seasonal Prediction (S2S) and
Second International Conference on Seasonal to Decadal Prediction (S2D)

Daily news

Day 3: Using subseasonal to decadal predictions in decision making

Marking the halfway point in the conference, the concurrent S2S and S2D themes covered forecasts for decision making, S2S predictability arising from stratosphere processes and variability, and innovation verification methods, observations, and metrics for hindcasts and real-time forecasts in the S2D timescale.
Below are some highlights from the meeting.

 Subseasonal to Seasonal Conference Highlights

S2S forecasts have many socioeconomic applications for supporting things such as early warning capabilities for high-impact events and information to help sectors make informed decisions on risk management. And if S2S predictions are to be useful, then research momentum should continue in collaboration with how users, decision makers, and stakeholders need and apply this information.

Chris White, University of Strathclyde in the UK and the Antarctic Climate and Ecosystems Cooperative Research Centre in Australia, was the morning keynote lecture, setting the stage by discussing applications of S2S predictions. He emphasized that the S2S timescale needs to be clearly defined, its sources of predictability identified, and its end-user needs met through judicious application. Broad-scale precipitation and temperature forecasts are useful, but not nearly as much as specific variables for specific regions with specific interests. When talking with various sectors, the research community should strive to understand how to help address some of their questions and challenges.

Later in the day, Steve Pawson, NASA Global Modeling and Assimilation Office, showcased some of NASA’s platforms and modeling capabilities on the S2S timescale. For example, using their GEOS model, they are able to model the stratospheric aerosols of a Mt. Pinatubo eruption, showing about a half of a degree of SST warming in the tropical region, where other areas experienced cooling. Pawson also described challenges and tradeoffs with model complexity (e.g., soil moisture, salinity, sea ice, aerosols) versus the representation (e.g., resolution, ensemble size, stochastic physics) of features, even though growth in computing power is pushing both to higher resolution. There are choices that need to be made on which direction. NASA has decided to move into a more complexity space because they have observations.

Amy Butler, CIRES/University of Colorado, Boulder, and NOAA Earth System Research Laboratory, gave a keynote on the role of the stratosphere in S2S predictability. She noted that the stratosphere is a valuable source of predictability on S2S timescales. Many models are not capturing the Quasi-Biennial Oscillation (QBO) because of many of the small processes, but if an initialized model does have the correct QBO phase, then this could be a source of predictive skill. She also showed how atmospheric waves from the troposphere can propagate into the stratosphere and break, causing the polar vortex to rapidly slow down and potentially split. If the vortex remains its zonal structure, it provides a high source of predictability for winter weather. However, models are unable to forecast tropospheric wave forcing beyond ~10-15 days, so wave-driven stratospheric extremes (like sudden stratospheric warming (SSW) events) are also not skillfully predicted beyond those timescales.

A couple of group discussions occurred during the day. Andrew Robertson, IRI/Columbia University, led the first discussion on the S2S Prediction Project’s phase 2 plans, which include the applications and communications part of the project, and what the community recommends for ways to promote the forecasts and use. A key point made was that successful development of S2S operational products requires a user-centered approach, such as a “user needs first” co-design. However, users’ needs vary greatly across sectors, and meeting these needs requires collaborative and interdisciplinary effort in the scientific and social sciences communities. Many ideas were also presented on ways to improve the database, including adding model hindcasts and ocean variables into the phase 2 of the S2S prediction.

The final discussion centered on the stratosphere forecasting abilities on the S2S timescale. Andrew Charlton-Perez, University of Reading, identified a number of key questions still being addressed by the community, including how well are key dynamical processes in the stratosphere represented in S2S models, how important is higher vertical resolution and the complexity of stratospheric physics, and where are the “windows of opportunity” for exploiting subseasonal skill in the stratosphere. Participants noted that there is a need for more data/diagnostics to diagnose model mechanisms and that case studies/attribution studies can be useful to highlight the role of the stratosphere in major events and analyse processes. Another idea was to look at false alarm predictions (predictions that didn’t happen) for SSW events, which could provide some interesting insight into the processes.

Some additional highlights:

• As an example of an application that meets end-user needs, presenters identified that it would be much more beneficial to a water resource manager to issue S2S forecasts for precipitation and streamflow for well-defined watersheds than for arbitrary regions such as states or countries. (Sarah Baker, University of Colorado, Boulder/Bureau of Reclamation; Rachel Bazile, University of Sherbrooke, Canada)
• Further examples of tailoring S2S forecasts to end-user needs include predictions of 100-m winds for wind energy production and atmospheric river activity along the US West Coast. (Dominic Büeler, Institute of Meteorology and Climate Research, Germany; Duane Waliser, NASA Jet Propulsion Laboratory)
• Useful applications of real-time S2S forecasts are being developed separately in different sectors: the S2S4E project – providing operational climate service for the energy sector; the quasi-operational excessive heat outlook system in the health sector; and the S2S hydrologic prediction in the water management sector. (multiple speakers)
• S2S prediction skill, and hence the applicability of S2S forecasts, strongly depends on the target. Further case studies and evaluation of model sensitivities are required to improve forecast applications.
• Extending prediction range of SSW events has associated potential to improve S2S skill. But challenges remain to improve the underlying models in ways that allow the S2S community to focus on processes not anomalies. (Joan Alexander, NorthWest Research Associates; Alexey Karpechko, Finnish Meteorological Institute)
• However, others showed that while SSW events are not necessarily predictable on S2S timescales, there is evidence of surface precursors of SSW events, including from modes of variability (e.g., MJO). (Daniela Domeisen, ETH Zurich)
• The stratosphere Northern Annular Mode (NAM) is more predictable than the troposphere NAM (20 days compared to about 10 days). But stratospheric forecasts tend to be over confident (in a signal-to-noise sense) out to about three weeks, and there is no evidence of under confident stratospheric forecasts. (Andrew Charlton-Perez, University of Reading)

Seasonal to Decadal Conference Highlights

Wednesday’s morning session in the S2D prediction track continued from Tuesday to present research on "S2D ensemble predictions and forecast information.” Doug Smith, UK MetOffice, opened with a presentation on the question, “How skilful are decadal climate predictions?” Smith and other session contributors elaborated on the phenomenon that ensemble forecasting systems at S2D scales show limited levels of skill but are likely underestimating the actual levels of predictability – an effect that was, not without vivid discussion, referred to as the signal-to-noise paradox. Several talks explored recalibration, combination, and weighting of multi-model ensembles to help reduce areas with misleading negative skill and preserve skillful areas. Interest was high in the potential of such calibration and combination methods for operational settings. Dougie Squire presented CSIRO’s new system for seasonal to interannual prediction (“CAFE”) with a focus on interannual variability, as well as an integrated diagnostic and verification suite. Mikhail Dobrynin, CEN, University of Hamburg, showed how a recently developed sub-sampling approach holds potential to improve forecast skill for winter the North Atlantic Oscillation (NAO) based on observationally derived predictors.

Just like the S2D conferences overall aim to bridge gaps across timescales and move toward more seamless predictions in the temporal realm, increasing efforts among the science community – paralleled by strong demand from policymakers – aim to close gaps in the research-to-service delivery chain. In this spirit, the second half of Wednesday morning opened a session on “S2D forecasts for decision making.” In her keynote speech, Desiree Tommasi, NOAA Southwest Fisheries Science Center, gave an overview on “Climate predictions for fisheries applications,” mentioning the high influence of climate variability on the productivity and distribution of fish populations and pointing out present skill and future challenges in predictions for fisheries and ecosystem management.

In general, speakers noted increasing potential in extending the application of seasonal to decadal climate information for decision-making processes across various socioeconomic sectors. The contributions varied both in scale (international, national, individual user) as well as targeted sector (e.g., water management, marine safety), but all aimed to be of direct practical value for decision makers. Fundamental problems and limitations still need to be extensively addressed by the community. For instance, the evidence of reliability and effectiveness of incorporating climate information to decision making remains limited. The need for a common framework for verification of climate information was mentioned. Limited availability of observational data hinders the verification and validation of climate information, particularly at decadal timescales. Etienne Tourigny, Barcelona Supercomputing Center, presented an application of operational seasonal prediction systems for seasonal prediction of 2017 extreme wildfire events in California, Spain, and Portugal. Andrew Hoell, NOAA Earth System Research Laboratory, gave an overview on the Famine Early Warning Systems Network (FEWS-NET).

Wednesday afternoon continued on the theme of S2D forecasts for decision making, featuring a series of studies and initiatives to develop and/or deliver climate forecast information. For the western US, Sarah Kapnick, NOAA Geophysical Fluid Dynamics Laboratory, explained the potential value of seasonal forecasting of snowpack for water management, which is only starting to be evaluated in dynamical predictions. She showed that model-based predictions of March snowpack at 8-month lead generally outperform prediction skills based on climate indices at the time of initialization. In general, research in the US is showing increased interest in climate forecasts to inform decision making for water management on multi-year and seasonal timescales. This involves using model forecast fields as input to hydrological models and understanding how large-scale climate variability influences water resources. Anca Brookshaw, ECMWF, presented the seasonal forecast component of Europe’s Copernicus Climate Change Services (C3S), which is developing an array of graphical and digital multi-system forecast products to inform decision making for multiple sectors. The session overall underlined that the development of usable climate forecast information requires building relationships with potential users, opening dialogs and co-design over longer preparation phases.

The last session of the day was dedicated to the theme of “Hindcast and forecast quality assessment.” In the opening keynote talk, Timothy DelSole, George Mason University, presented on recent developments in forecast quality assessment, focusing on the general question of how to decide whether one model is better than another. DelSole first laid out problems of inapplicability of standard statistical tests based on differences in correlation (or mean square error) when the skill measures are computed from data over a common period or with a common set of observations, and, subsequently, introduced a simple and robust recent random-walk test together with relevant literature from beyond the field of climate science.

The following technical recommendations and findings were presented by other speakers in the session:

• Danila Volpi, CNRM/Météo-France/CNRS, introduced a method in which spatial covariances of a given measure were calculated to obtain the teleconnection patterns of two models. After calculating the pattern correlation in the observations against this teleconnection, the resulting difference can serve as measure of model performance.
• Reinel Sospedra-Alfonso, Environment and Climate Change Canada, used canonical skill analysis to extract the part of the forecast which is skillful, by finding the highest order EOF for which the skill reaches saturation.
• André Düsterhus, University of Hamburg, recommended to not use anomaly correlations as a measure of skill on non-normal distributed measures but instead a new metric based on the earth movers distance, which is a measure for the difference of two PDFs.
• Kristian Strommen, Oxford University, pointed out that the signal-to-noise paradox in NAO predictability can potentially be replicated by an idealised Markov process, for which it could be argued that coupled models do not show sufficient persistence when simulating the NAO.