Seasonal Forecast

In my day job I spend a good deal of time working with seasonal weather forecasts, and recently I have focused on developing a new scheme to calibrate seasonal forecasts from climate models such as NOAA's CFSv2 model.  The idea here is to transform the model ensemble forecasts into reliable probabilities so that users can make confident decisions.  The word "reliable" has a technical meaning: if a probabilistic forecast is reliable, then the predicted probabilities correspond to the observed frequency of occurrence over a long period of time.  For example, if the forecast says there is a 70% chance of above-normal temperatures, then over the long haul above-normal temperatures will be observed 70% of the time.  It's important for users to have confidence that forecast probabilities are reliable in this sense; if they're not, then the forecast isn't suitable for use in quantitative decision systems.

There is a lot more that could be said on this topic, but I'll just illustrate the probabilistic forecasts by showing the latest CFSv2 seasonal forecast for Alaska.  The maps below show forecasts for temperature, liquid-equivalent precipitation, and 10m wind speed, for March.  It's traditional with seasonal forecasting to divide the historical data into three equally-likely categories of below-normal, near-normal, and above-normal, and then predict the probabilities of each category.  In the forecasts shown here, all three probabilities are predicted at each point, but the map shading indicates the highest of the 3 tercile probabilities.  According to the CFSv2 forecast, significantly warmer than normal conditions are quite likely over all of interior and northern Alaska, with the probability exceeding 60% in the eastern half.  The forecast also favors drier than normal and less windy than normal conditions, but with much lower probability.

The rather high probability of very warm conditions in March reflects two facts: the raw model output is showing a large departure from normal, and the model has some skill at predicting March temperatures at this lead time.  The first point is illustrated by showing a comparison of this year's forecast for Fairbanks with forecasts from previous years (which are run in retrospective mode) - see below.  The blue dots represent the individual ensemble member forecasts made in February, and the red markers show each year's ensemble mean.  The bold horizontal lines indicate the tercile boundaries for the raw model forecasts.

Remarkably, the current forecast for March shows warmer conditions than in any of the retrospective forecasts from 1982-2010.  In the forecast history, the highest temperature predicted for March was -5.4°C, but 4 of the 40 latest members are showing a March temperature higher than that. (One caveat: the temperatures are not bias-corrected here; they are simply taken from the nearest model grid point.)

So the model has a strong warm signal this year.  But does that mean a warm March is actually more likely?  Obviously if the model has no skill, then it doesn't mean anything.  This is where the historical calibration comes in: the retrospective forecasts tell us how good the model is, and the calibration scheme transforms the model signal into a reliable probability.  Here's a chart showing the performance of all forecasts at 1-month lead time from January through March, i.e. January forecasts for February, February for March, and March for April.  The performance isn't great, but it's enough to do something with.

In a few days NOAA's Climate Prediction Center will release an official forecast for March, so it will be interesting to make a comparison.  In the meantime, here are the tercile probabilities for Fairbanks and Anchorage (closest grid points):

Fairbanks March temperature: 12% below-normal, 27% near-normal, 61% above-normal
Fairbanks March precipitation: 35% below-normal, 43% near-normal, 22% above-normal

Anchorage March temperature: 6% below-normal, 23% near-normal, 71% above-normal
Anchorage March precipitation: 28% below-normal, 31% near-normal, 41% above-normal

Finally, here's the forecast for the 3-month period March through May.  The probabilities of warmth are even higher, exceeding 70% in parts of the interior.  It probably goes without saying that this may bode ill for the early part of the fire season.

Update Feb 18: here's the performance chart using actual temperature observations from Fairbanks airport rather than the CFS Reanalysis for verification.

And here's the CPC forecast, hot off the press today.  The CPC has similar high probabilities of warmth for March, but somewhat lower probabilities for the March-May season.  Also, they expect relatively wet conditions in southeast Alaska during March, whereas the CFSv2 is showing dry.

Update April 20: here's the verification for March.  The CFSv2 did better than the CPC for precipitation in southeast Alaska, as it turned out to be relatively dry.