Can we trust statistical predictions?

This is the overarching question raised in a new paper by Erik Kolstad and James Screen. The backdrop is that many people have studied the relationship between the extent of the sea ice in the Barents and Kara Seas in autumn and the wintertime weather farther south several months later (see for example this paper by Hall and co-authors). In recent decades, lower-than-normal sea ice extent has typically been followed by lower-than-average temperatures in Europe.

On the face of it, these kinds of relationships are great, because they can be used for prediction. Other phenomena like El Niño often lead to quite predictable changes in the weather patterns in many places. The potential value of a similar relationship between sea ice and weather is huge. Yet, a crucial question is: Even if there has been a statistical relationship between two variables in the past, does that guarantee that the same relationship will be valid in the future?

In our paper, we found that the lagged relationship between October sea ice extent in the Barents/Kara Seas and the North Atlantic Oscillation in winter (December to February) has changed a lot over time. To illustrate this, we made this figure:

Fig2
The colours show the correlation between October sea ice cover in the Barents/Kara and sea level pressure in the North Atlantic region in December–February during the periods indicated above each panel.

During the first period starting in 1916, the correlation between sea ice and pressure is positive over the Northeast Atlantic and negative farther south. This means that below-normal sea ice cover in autumn was generally followed by westerly winds and mild and wet winter weather in Europe. But in the period between 1980 and 2010, the more familiar pattern with cold winters following low sea ice was evident. This shows how much the correlation can change over time.

The paper shows that climate models have the same fluctuations, and this led us to conclude that—as the title indicates—the relationship is non-stationary. Which again led us to “caution against indiscriminately using Barents‐Kara sea ice to predict the NAO”. We believe this has important implications for seasonal prediction.

Our first 2019 summer forecast

We are now ready with our first ever in-house forecast. It’s based mainly on dynamical forecast models from the Copernicus Climate Change Service (CS3), but it also includes the Bjerknes Centre’s own Norwegian Climate Prediction Model (NorCPM).

Without further ado, here are our predicted temperature anomalies for June:

fc_t2m_6

It requires a bit of explaining. The colours indicate deviations from the normal temperature, which means the average over the past 20-30 years, depending on the model. The first five panels show the predictions from the individual models, the sixth panel shows (more or less) the average of these, and the bottom panel is our own combination of the model forecasts (Multimodel+). For more details on how we do this, please go to our Methodology page.

In short, the forecast is that it will be ever so slightly warmer than normal, but no large anomalies are forecast. The reason for the not-so-exciting forecast is that the models disagree. For instance, the ECMWF model predicts a warm June, while the UK Met Office predicts a cold June.

A pertinent question is: how skilful are these models? To answer this, we have prepared so-called skill maps. Here’s the one for June:

skill_maps_6new

The short story here is that blue means high skill, whereas red means low skill. Deep blue means a 20% improvement on ‘climatology’ (details). The maps are based on the models’ re-forecast skill. Our Multimodel+ forecast is moderately skilful in June in coastal regions.