Forecasting temperatures and the costs of simplification

Or why there will always be differences between forecasted temperatures. Between forecasters and between the forecast and the actual value. If there is such a thing as a single value for a whole city or region. 

Stephen Davenport, Senior Meteorologist at MeteoGroup, reflects on some recent events. 

There were some reports in the news that Friday 27th July would likely see the UK’s July record temperature of 36.7˚ C broken, and that there was a 20 to 30 per cent risk that the all-time temperature record of 38.5 ˚ C would be surpassed. Given the extremely hot airmass persisting across the country and other factors such as feedback from the arid ground, then, if all the ingredients had come together perfectly, the temperature might have been nudging those records somewhere in southeast or eastern England. Might have.

At MeteoGroup we took a slightly different view, with an eye on thunderstorm development and a risk that associated medium-level cloud would begin to spread across the potentially hottest areas: one of the ingredients that can easily suppress the maximum temperatures by a degree or two.

Most important is to bear in mind that a given forecast temperature is always the most probable figure, with a sharp bell-curve of probabilities either side. And for any given area, forecasts as presented to the public are quite often an aggregation and/or a simplification. 

The take-away from last week is that whatever the highest temperature ended up being, Friday was yet another extremely hot day in this furnace of a summer – on this all forecasting agencies agreed. A couple of degrees then makes little difference at this level of heat but, it is not unfair question to ask why there can be these slight differences from one forecasting body to another or from one meteorologist to the next? A discussion that's flaring up at regular intervals. 

We all, of course, use numerical weather prediction (NWP) models as a basis but they are just that – mathematically-derived models of approximated conditions throughout the atmosphere and across the globe. They provide, especially these days, an exceedingly good view of expectations, especially the local area models which concentrate on smaller regions at higher resolutions, but the atmosphere is so vast and deep that even these are unable to bring fine enough resolution to pick up on every slight but impactful variation. There is one quote that is always worth bearing in mind, from the statistician George Box: “All models are wrong; some are useful”.

As such there can never be a perfectly precise prediction, and this is seen every day when models often show differences, even sometimes at quite short range, in predicted outcomes. So although numerical models are the basis, what comes next is important. For example, different weightings can be applied to each model depending on historic performance or current accuracy. This is an approach taken by MeteoGroup, where we use not just this “multi-model” approach but also statistical methods which take into account the historical behavior of forecast locations to produce our Model Output Statistics (“MOS”, MeteoGroup Multi-Model MOS). This combination of high-quality data and sophisticated statistical application adds value to raw model output and has proven as successful as any throughout MeteoGroup’s history.

There are two final things to note: a 20-30% chance of breaking the all-time record naturally means a 70-80 per cent chance of not doing so. And while concentrating on temperatures we must not ignore the fact that the expected showers and thunderstorms will have brought some regions a very welcome bout of rainfall.

Stephen Davenport 
Senior Meteorologist / Energy Meteorologist

Picture: Carbon Brief

For more reading on forecasts and predictability of the weather:
See: Long term weather forecasts predicted European dry heat over six weeks in advance​
and: Translating accurate forecasts into unambiguous icons - and how to read them
and: Predicting Rain. Harder than you think.

All models are wrong; some are useful