Seasonal Forecast

Regular readers will know that I have strong personal and professional interests in long-range forecasting, including both sub-seasonal (weeks ahead) and seasonal (months ahead) outlooks.  It's a challenging discipline, of course, but seasonal forecasting in particular is a mature field with well-established methods and tools that (can) lead to a modest degree of success.

Among the tools that are regularly consulted by seasonal forecasters are both dynamical model forecasts produced by well-funded supercomputing centers and statistical forecasts from any and all sources.  An attractive aspect of statistical techniques is that they can be simple enough to run on a desktop computer in minutes; there's no barrier to entry in this field.

In the past week I've been digging into a new approach reported by Matti Kämäräinen of the Finnish Meteorological Institute:

https://journals.ametsoc.org/doi/abs/10.1175/JCLI-D-18-0765.1

I won't reproduce details, but the essence of the scheme is that it develops a large ensemble of regression models over a long 70-year history to predict surface temperature in 3-month periods.  The predictors are simple: global sea surface temperature and 150mb (lower stratosphere) geopotential height patterns in previous 3-month periods.

Kämäräinen and his co-authors report significant skill in predicting area-average temperature over large domains such as Scandinavia and western Europe.  I managed to reproduce most of the results with my own code and then apply the scheme to a domain over Alaska and western Canada.  The model produces forecasts for the area-average temperature in terms of departure from trend, and the chart below shows results for the entire verification period, 1986-present (click to enlarge).

Here's the forecast domain: note that we're predicting area-average temperature for the whole region, not for any particular point.

The overall correlation is modest but unquestionably positive.  Here are the correlation coefficients for the last 20 years of 3-month forecasts, i.e. 80 consecutive seasonal forecasts for 2000-2019:

Dec-Feb  +0.66
March-May  +0.35
June-Aug   +0.58
Sep-Nov   +0.36

Bearing in mind that this is strictly out-of-sample verification (i.e. we're not cheating), and that the long-term trend was removed (i.e. the trend does not boost the skill), this is quite impressive in my view.

So what does the scheme predict for this summer?  The last data point on the chart shows the forecast for June-August 2020, and it's a cool look; in fact, the regression predicts the lowest trend-relative temperature for summer since 2009.  The cool forecast doesn't look like much on the chart, because variance is low in summer, but it's a big change from last year's forecast, which was (correctly) quite warm.

Given that the latest dynamical model forecasts are pointing warm, and NOAA's Climate Prediction Center also shows warm, it will be interesting to see how the summer turns out.  Of course, a true comparison of these forecasts would require us to reconcile the different climate reference points, because "below trend" could easily be "above normal" if we're talking about the 1981-2010 baseline that CPC currently uses.  I'll leave that detail for another day.