Good news! The weather app on my phone promises sunny skies and a high of 23C tomorrow. Actually when I click on this cheerful symbol, it breaks the day down into a helpful icon for each 3-hourly chunk. It’s not actually indicating clear blue skies and sunshine throughout. There will be some puffy cloud, but it still looks fine. Even more usefully, each icon has an accompanying probability of rainfall, precise to the nearest 1%. And if I’m willing to pay some more money, I can get a forecast out to two weeks ahead.
A friend shows me a different app that she uses. This gives me an hourly breakdown of the expected weather. Impressive. But hang on a minute – this app expects the day to be chillier, quite grey and indicates light rain for three hours of the day. There’s no clue as to the probability of that rain – so presumably it’s a certainty?
Surely in this day and age we should expect better?
I’m now confused. Which app do I believe? How can I plan my day? The confusion is compounded just a few hours later, when both apps have changed their forecast; but they’re still not the same as each other. In fact the first one has upped the chance of rain to 64%, whilst the other one is completely dry. They both seem quite precise in their own way, but inevitably, at least one of them is bound to be precisely wrong.
Sounds familiar? How can we trust these forecasts when they keep changing? Surely in this day and age, we should expect better?
Well – yes and no.
Weather apps are often very good. This is despite the fact that meteorology is a fiendishly complicated science. The weather we experience at any given hour of the day is the result of mind-blowingly intricate interactions between air molecules all around the world.
Imagine a molecule of air as a (tailless) donkey. We know roughly where the donkey is as it is released into a huge field. But, based on that information alone, good luck trying, blindfold, to pin the tail on the same donkey after it has randomly wandered around the field in the intervening few days.
From the meadow back to meteorology – the same principle applies. Even if we know exactly what each molecule of air is doing right now, predicting where that molecule will go in the future depends on knowing the precise laws of physics which govern its behaviour, and having a powerful enough super-computer to do the billions of calculations required. None of these things can be done perfectly. There are inevitable errors. And with each new run of the model – several times a day – these errors will change, and so the forecast will change too.
Predicting the weather, hard enough for the next few hours, gets increasingly difficult the further ahead we look because the errors get larger. Sure, the computers will still make a prediction. Indeed, they’ll come up with a neat icon for a Saturday afternoon in 6 months’ time if they’re asked to, but the reliability of such predictions becomes vanishingly small beyond the range of a few days.
There’s a difference between precision and accuracy
To muddy the waters further, depending on your app, the forecast will come from one of a number of meteorological centres, each of which has a supercomputer that models the atmosphere in a subtly different way. Each model has its own unique flaws.
To try to take account of these flaws, some centres analyse how consistent their predictions for a particular outcome are. If very diverse, then the weather is deemed to be inherently unpredictable, and vice versa. Using this method, some apps display a probability of a certain weather outcome, for example a 32% chance of rainfall. But to the nearest percentage point? Several days ahead? Really?
“OK”, I hear you say. “But surely the forecast for my postcode should be OK for just the next few hours, right?” Well it depends.
There’s a difference between precision and accuracy. Weather models provide precise forecasts for specific points. But what if there’s a hill or a lake or a coastline between that point and your postcode? Even the most powerful models cannot resolve every crease and fold in the landscape. Very local effects which the models don’t ‘know about’ can make a big difference, even in the next few hours. From clear to cloudy, from -3C to +3C, over just a couple of miles.
Which brings me onto why, perhaps, apps could be a little clearer about their limitations.
Your app has little or no human ‘quality control’. It’s not cost-effective in a world where there’s big money to be made by meeting the public demand for automated forecasts. While forecasters know the situations when apps are going to be more or, conversely, less reliable, they now have little way of manually correcting things the right way. And with the slow decline in the reach of traditional broadcast media, the opportunity to communicate the caveats and nuances in the forecast via TV and radio is shrinking too.
Increasingly then, the public are left to make their decisions from a machine. Like all modern machines, it’s very useful, reasonably reliable but, occasionally, is prone to breaking down.
So – ‘Sunny’ tomorrow? Your app may boldly say “Yes”. But the reality is often more of a “Maybe” – and that’s a fact.
John and Sara give some hints on when (and when not) to trust your apps here