How extremes can help: improved accuracy for 10-year forecasts

3 July 2018, by Stephanie Janssen

Photo: CC3.0 B.Abis, MPI-M

Leonard Borchert is a doctoral researcher at Universität Hamburg.

Leonard Borchert’s goal is to make weather forecasts for the next 10 years more accurate. To do so, he has observed the North Atlantic, which strongly influences the weather in Europe. Borchert’s findings have now been published in the Journal of Climate. As a PhD candidate at CEN, he also knows when good forecasts simply aren’t possible.

You work with climate models, the tools of the trade that all climate researchers use in order to make prognoses about the future. How would you describe climate models in just three words?

Playground—large—complex. They allow me to play with the laws of nature without doing any harm, and to learn something in the process.

Who needs 10-year forecasts?

They are highly relevant for politicians and farmers, as well as insurers. For example, less than average rainfall could cause a drought, or there could be a higher risk of hurricanes.

To date, long-term forecasts have proven difficult. Why?

10-year forecasts are also known as decadal forecasts, and they are situated between weather forecasts and climate projections. For a good weather forecast, it is important to know exactly what the current weather is like. This information can be used together with parameters like wind direction to predict the weather for up to five days into the future—but not much more.

Climate predictions are a different matter entirely ...

Yes, here timespans of 30 years or more are interesting. Large-scale physical processes are included in the climate models. What is the proportion of greenhouse gases in the atmosphere? How are land and sea ice developing? For a 10-year forecast, I need to combine the two different approaches. This is technically possible—the problem is the results: the forecasts aren’t particularly accurate.

How can I check how good a forecast is?

We test the models using retrospective forecasts, also known as “hindcasts.” To do so, we have the models forecast the weather for multiple timespans in the past for which we already have reliable meteorological records. We then compare how well the forecasts measure up to the recorded data. On this basis, we can calculate a “quality index” of sorts. The index currently ranges between 0.5 and 0.6 for the 10-year forecasts—which makes them slightly better that guessing. A 1.0 would be a perfect score. My latest was 0.8, which is practically sensational.

How did you manage that?

The problem is the quality index. It consists of several components, such as water temperature, air temperature, and wind currents, which are combined to assess the quality of the forecast. But, for example, the effects of the individual factors could cancel each other out.

I deconstructed the index and considered only one factor: heat transport in the North Atlantic. The weather in Europe is greatly influenced by the Gulf Stream, which transports heat here on its way north. And as we can see, this factor also has a major impact on the quality of forecasts.

How does heat transport influence forecasts?

My calculations show that good forecasts can only be made when the ocean either transports an especially large amount of heat—or an especially small amount—at the beginning of the forecast period. If the water temperatures are moderate, the forecast won’t be very accurate. The same applies to the amounts of water transported: if it’s particularly large or small, the forecast quality will be good, but with more moderate values it will likely be poor.

So extremes can be good for forecasts. Does that mean we can’t make good forecasts for years characterized by average values?

Exactly. Either the models aren’t yet good enough, or the climate itself hasn’t made its mind up as to how it will develop on the basis of these moderate conditions.

How accurate are the forecasts you can currently make?

For instance, a typical, reliable forecast would be: the risk of a hurricane hitting Northern Europe within the next 8 to 10 years has risen considerably. But I can’t make a forecast like: a hurricane will hit Lower Saxony in April 2026.

What will now happen with your findings?

Hopefully they’ll encourage the community to reconsider current scientific practices. After all, it’s just a first example for the North Atlantic. We could also investigate heat transport in other regions, like for the recurring climate phenomenon El Niño. Here, more accurate forecasts would be of global interest.

Link to the article: https://journals.ametsoc.org/doi/10.1175/JCLI-D-17-0734.1