Review of Winter Forecasts
By: Phil Arkin
Since winter (meteorological winter, which is December – February) has just ended, we can look back at the various forecasts that were made to see how well they did. I took a look at 4 relatively serious winter forecasts for the College Park area that can be validated (this isn’t meant to be a serious scientific evaluation, obviously – more of an introduction to what forecasts are available and some of their characteristics). These are climate forecasts, and therefore contain only the projected difference from the long-term average of mean temperature and total precipitation (one of these forecast total snowfall rather than precipitation).
To begin with, let’s consider why we believe that such climate forecasts are possible at all. Weather forecasts are made by building mathematical models of the atmosphere that include as much detail as is practical, specifying the actual conditions at a specific time, and allowing the model to apply the governing physical laws to project future conditions. This has been going on since shortly after World War II, is done regularly at national and international centers around the world, and has shown fairly steady improvement in predicting all sorts of weather phenomena for the near future – the process is referred to as numerical weather prediction, and the models are often called NWP models. The main reason that we can’t simply expect these models to predict specific storms and such indefinitely into the future is the chaotic nature of the atmosphere, in the sense that future conditions depend very sensitively on extremely fine details in current conditions, and since we can’t measure current conditions with such fine detail we can’t expect to make useful predictions indefinitely. Experiments have indicated that the realistic limit on such weather forecasts is around 2 weeks, although it seems possible that some aspects of the weather/climate interface, such as the location of storm tracks and blocking systems, might be predictable farther into the future.
So what leads us to think that we can predict anything about weather, even the statistics, farther into the future? Well, obviously the climate is predictable in a certain sense – every January is more similar to other Januaries than it is to, say, any July, so if we predict that next January will have temperature and precipitation equal to the average of the last 20 or 30 Januaries we will probably be reasonably close. (This is a good example of a forecast that is so straightforward that most people don’t think of it as forecasting at all!) To do better than that, we rely on the fact that climate anomalies (the difference between any specific month or season and the long term average) do exhibit some dependence on what we call boundary conditions, such as anomalies in sea surface temperatures or continental snow cover. Speculation along these lines began a long time ago – Henry Francis Blanford, founder of the India Meteorological Department, was discussing the relationship between Himalayan snowfall and the Indian Monsoon more than 125 years ago – and the role of at least ocean temperature anomalies is quite well accepted now.
Let’s look at the actual forecasts for the recent winter. I’ll try to specify the forecast of departure from average of temperature and precipitation for College Park, MD, although none of the forecasters is so optimistic as to try to predict on such a fine spatial scale. For the area as a whole it was the 3rd warmest and 3rd least snowy winter on record, and total precipitation was a bit above average, with December quite wet and January-February dryer than normal.
Climate Prediction Center: the CPC provides the US government’s official climate forecast. Their forecasters use a variety of statistical tools along with a set of mathematical models that represent the ocean, land and atmosphere and simulate the effect of surface anomalies on the atmosphere on average. (These models are often referred to as coupled climate models – CCM.) CPC issues its forecasts at least 2 weeks before the beginning of the forecast period (so by November 15 for DJF) so as to avoid any skill that might be provided by predicting the short-term weather variations at the beginning of the period. CPC’s forecast was not exciting for the local area – equal chances of above and below normal for both temperature and precipitation. You can see how CPC’s forecast verified here.
As mentioned above, CPC makes use of at least one CCM in their forecast process. Two organizations go a bit further and publish forecasts based on multiple forecasts from one or a number of models. The European Center for Medium-Range Weather Forecasts (ECMWF) uses their own model to make 51 seasonal forecasts from which they derive the forecasts on their web site. They publish maps of several predicted statistics for temperature, precipitation and a few other quantities for several areas. They avoid displaying North America on their site, but their tropical forecast ensemble mean seems to give the DC-area conditions of 1-2° above average temperature and 50-100 mm below average precipitation.
The International Research Institute for Climate and Society uses CCM from several sources (a multi-model ensemble) and certain other tools in creating their forecasts. For the DC area, the IRI forecasts (this links to the current forecast – I didn’t find an archive), like those from CPC, predicted equal chances of above or below normal temperatures and precipitation. They do provide quite a bit of validation information for global scales.
The Capitol Weather Gang made their own forecast, which included an excellent discussion of the background and also linked to a few others. They predicted near normal temperatures and slightly below normal snowfall (and no prediction for precipitation). I won’t go through the others, but you can see them by using the links in the CWG’s forecast article.
My overall impression is that each forecast had some idea of the relative warmth that we experienced, but no one projected the extreme values that occurred. The precipitation and snowfall forecasts were less successful, not surprising since it’s known to be a harder problem. The ECMWF forecast is interesting to me because it’s the only one that is a pure model product and I’ve seen studies where it appears to deliver some useful skill (although it’s hard to tell for this winter). My impression is that weather forecasts began to improve dramatically once the models were clearly contributing something useful (this is based on personal observations in pretty ancient history, but one can see the model forecast skill improvement and lots of other statistics here), and I hope that something similar might happen with climate forecasts.