Although wintry conditions have been lingering across northern Europe and the northern USA, winter 2023-24 is largely in the rearview mirror for energy markets, and that’s a relief for many industry meteorologists. It was a very challenging winter for long-range prediction, with speculation and hype running rampant from the earliest winter forecasts to the week-to-week variations throughout the season.
The WCS Winter Forecast
The main theme of the winter, of course, was the El Niño event that developed in summer 2023 and became one of the strongest episodes on record by mid-winter, at least as measured by SSTs in the equatorial Pacific Ocean. The Oceanic Niño Index exceeded +1.9°C, a level only reached in a handful of events since 1950.
With such a strong El Niño episode in play, one might imagine that the winter forecast would have been easier than normal, because of the overwhelming and well-understood influences of such events. However, El Niño’s evolution was unusual, and the configuration of global SST patterns was unlike prior El Niño events, leading to uncertainty in whether the canonical impacts would be observed. In particular, statistical predictors and historical analogs strongly suggested a heightened risk of patterns favoring more mid-latitude cold outbreaks than would be typical of a strong El Niño. These factors were discussed in the World Climate Service winter forecast, and the subjective forecast confidence was relatively low.
As it turned out, the WCS subjective winter forecasts for Europe and North America proved to be fairly successful – although more so for precipitation and wind, and somewhat less so for temperature. The figures below show a qualitative comparison of forecast (above) and observed (below) conditions.
Europe
North America
Subseasonal Forecasts
The World Climate Service provides a range of tools to facilitate subseasonal forecasting, including dynamical model forecasts and historical analog analysis capabilities. One of the mainstays of subseasonal forecasting is the ECMWF extended ensemble model, which (since June 2023) runs daily with 101 ensemble members to 46 days in the future. These “EC46” forecasts are closely scrutinized by forecasters and provide a baseline for market sentiment at times.
One of the interesting and challenging aspects of the EC46 forecasts in winter 2023-24 was the unusual degree of persistence in the forecasts for subseasonal lead times. Much more so than usual, the forecasts tended to show the same signals week after week, and forecasters searching for potential pattern changes were forced to rely on other tools like historical analogs.
To illustrate the unusual behavior, the red line in the figure below shows the correlation of the week 4 and week 6 forecasts of 500mb height over the past year, averaged over the Northern Hemisphere north of 30°N. The black line shows the average correlation over 20 years of EC46 hindcasts, showing the typical model behavior in other years. Clearly, the week 4 and week 6 forecasts have been much more strongly correlated than normal since June 2023 – which happens to be when El Niño emerged.
We see the same thing in the EC46 temperature forecasts:
The maps below show just one example of what this looked like in a recent EC46 forecast for weekly average 500mb height (top) and 2m temperature (bottom) anomaly. Week 4 is shown in the left column, week 5 in the middle column, and week 6 on the right, from the same March 14 initialization. Clearly there’s a remarkable degree of similarity in the consecutive weekly forecasts, in both Northern and Southern Hemispheres, and not just over the ocean (where it is expected, at least for temperature, owing to persistent SST anomalies).
It’s worth noting that this high degree of persistence is unrealistic, in the sense that real-world weather patterns usually vary much more from week to week. Consecutive weekly averages of 500mb height and 2m temperature tend to have a spatial anomaly correlation of roughly +0.4 on average in Northern Hemisphere winter, for 30-90°N latitude. Based on the EC46 hindcast results shown above, the model is usually “in the ballpark”, but this winter was a major exception.
ENSO Impacts on Persistence and Skill
It seems very likely that the strong El Niño episode was to blame for the unusually persistent EC46 forecasts, but do the model hindcasts provide evidence to support this? Yes: the figure below show the relationship between the Dec-Feb Oceanic Niño Index and the Dec-Feb average forecast correlation for week 4 versus week 6. The two strong El Niño winters in the hindcast period (2009-10 and 2015-16) also produced significantly above-normal persistence, so this is clearly a typical model response. However, the sheer tenacity of this winter’s forecast persistence was unlike anything seen before.
Does El Niño have the same effect in the real world, i.e. does it tend to increase the correlation between consecutive weeks? Based on data from the same hindcast period (2004-2023), the answer is “perhaps”, with 2009-10 and 2015-16 both having above-normal persistence (see below). However, this winter was again a major exception – in the other direction! Remarkably, this winter in fact saw relatively low persistence from week to week, and especially for 500mb height:
The obvious conclusion is that this winter’s EC46 forecasts were really unhelpful: they showed extremely high persistence from week to week, but the outcome was actually the opposite, with more week-to-week variability than usual.
Not surprisingly, then, we can also show that the subseasonal forecasts were less skillful than normal this winter, and much less skillful than in the two previous big El Niño winters. This was a disappointment for forecasters: despite pronounced “boundary condition forcing” from the strong El Niño, the EC46 subseasonal forecasts were poor.
Comprehensive Winter Forecast Tools
The poor performance of the ECMWF’s subseasonal model this winter highlights the need for additional, independent forecast tools, and given that all dynamical models tend to behave similarly, the potential value of statistical and analog forecasts is clear and compelling. For more than 15 years, the World Climate Service has relied on historical analog and statistical methods to augment the dynamical model forecasts, and World Climate Service users are able to explore their own analog-based predictions with WCS tools. Success is found most often with an intelligent blend of diverse tools and guidance.
Looking ahead, machine learning and artificial intelligence offer tantalizing prospects of future improvements, as recent spectacular advances in medium-range forecasting have spurred efforts to extend the AI models into subseasonal and seasonal lead times. The winter forecasts of 2023-24 were indeed challenging, but the future is bright for long-range forecast innovation.