Forecasting the Weather
Few professions are criticised as freely as weather forecasting. When a promised sunny afternoon turns wet, the forecaster is blamed; when a predicted storm passes harmlessly out to sea, the warning is dismissed as an overreaction. Yet by almost any objective measure, forecasts have improved enormously over the past half-century, to the point where a forecast for five days ahead is now roughly as reliable as a forecast for the following day was a generation ago. The persistent sense of disappointment, I would suggest, has less to do with the quality of forecasts than with a widespread misunderstanding of what a forecast actually claims to tell us.
People have always tried to anticipate the weather. Farmers and sailors relied on the colour of the evening sky, the behaviour of animals and a large body of rhymes and sayings passed down through generations. What transformed the practice was not a scientific theory but a communications technology. Once the telegraph allowed observations of pressure, wind and temperature to be sent rapidly from distant stations, it became possible to compare readings taken at the same moment across a whole region, and to draw maps showing how weather systems were moving. Forecasting became, in effect, a matter of tracking storms and estimating where they would go next.
The next step was more ambitious. If the atmosphere obeys the laws of physics, then in principle its future state could be calculated from its present one. The first serious attempt to do this, carried out by hand in the early twentieth century, took many weeks of arithmetic to produce a prediction for just six hours ahead, and the result was wildly inaccurate. The idea was sound, but it required two things that did not yet exist: far more observations and machines able to perform millions of calculations quickly. Only with the arrival of electronic computers did numerical forecasting become practical, and it has dominated the field ever since.
Even with powerful computers, however, there is a limit to how far ahead the weather can be predicted. The atmosphere is a chaotic system, meaning that tiny errors in the initial measurements, which are unavoidable, grow steadily as a calculation runs forward in time. Beyond about two weeks, the detail of a forecast becomes essentially worthless. Rather than pretending to a certainty they do not possess, forecasting centres now run their models dozens of times, each with slightly altered starting conditions. If the resulting forecasts closely agree, confidence is high; if they diverge widely, forecasters know that the outlook is genuinely uncertain.
This approach naturally produces probabilities rather than definite statements, and it is here that communication becomes difficult. Surveys repeatedly show that members of the public interpret a phrase such as 'a 40 per cent chance of rain' in a variety of ways: some think it will rain for 40 per cent of the day, others that rain will fall over 40 per cent of the area. In response, a number of weather services have abandoned percentages in favour of words like 'likely' or 'possible'. In my view this is a mistake. Vague terms are understood even less consistently than numbers, and they deny the public information that forecasters already have. The better course is to explain probabilities patiently rather than to hide them.
A further change is now under way. Forecasting systems based on machine learning, which are trained on decades of past observations rather than on the equations of physics, have begun to match and in some cases outperform traditional models, and they can produce a forecast in minutes rather than hours. Some commentators have concluded that human forecasters will soon be redundant. I do not share this view. Machine-learning systems can struggle with rare events precisely because there are few examples of them in the historical record, and it is exactly these events, such as unusually intense storms, that matter most. Human judgement remains essential both to interpret such situations and to issue warnings that people will actually act upon, although the forecaster's role will undoubtedly become more that of an interpreter than a calculator.
Ultimately, the value of a forecast lies not in whether it turns out to be exactly right but in the quality of the decisions it supports. An energy company uses temperature forecasts to estimate demand for electricity days in advance; a farmer may decide whether to harvest early; a city council must judge when to close a bridge. Each of these decisions involves weighing the cost of acting against the cost of not acting. Seen in this light, a forecast is best understood not as a promise about tomorrow but as guidance about risk, and the public would be better served if it were treated that way.
