AccuWeather offers a 90-day forecast, but here are the problems we found predicting just 10 days ahead

End, begin, all the same. Big change. Sometimes good. Sometimes bad. Picture: Universal Pictures

Ask any farmer and they’ll grumble about weather forecasting, and with good reason.

The government, on one hand, spends millions on forecasting, and on the other, charges farmers millions for the right to store and use the water that falls out of the sky onto their properties.

That – and the increasingly sporadic nature of rain – makes decisions about when to use that water more crucial every passing year.

Yet predicting the weather, like homeopathy and alkaline diets, is one of those arcane arts which, no matter how often is proven to be an exercise in futility, still manages to present as an authority humans need to pay attention to. And we do, and talk about it every day.

So it’s possible that when someone at an AccuWeather all-in proposed a 90-day forecast, everyone in the room simply thought “Why not? It can’t be any less accurate than a weekly forecast.”

No, it can’t, because even weekly forecasts are frustrating and trouble. I know, because I consider myself a small-time farmer, when not being a journalist, and I’ve been watching several weather apps every day for a year now.

There’s no point picking one out for special condemnation, as the fact they all provide different forecasts tells you all you need to know about their accuracy without going any further.

But let’s anyway.

So, you’ve got a 90-day forecast?

AccuWeather rolled out the new service a couple of weeks ago. It sounded revolutionary – insider knowledge no one else has – as if the gospel is whispered from Zeus himself. Here’s the blurb from their Enterprise Solutions department when it launched:

“Detailed weather forecasts, weeks or months in advance, will help you make proactive business decisions and stay ahead of the weather! Many industry leaders consider AccuWeather long-range 90 and 180-day forecasts to be their best-kept strategic secret.

This is a very bold claim, especially when that quote was live 89 days before anyone could even test AccuWeather’s 90-day forecast.

Even meteorologists admit anything beyond 7-10 days is fraught. Gizmodo spoke with several to confirm AccuWeather’s claim is misleading, and also got some carefully worded caveats from AccuWeather’s vice president of Innovation and Development, if you’re interested. His defence is we’re talking long-term trends, rather than actual weather.

“We’re always telling people not to focus on a specific day,” he said, perhaps missing the point behind why most people check weather forecasts.

But just to be certain, here’s how AccuWeather’s 10-day forecast panned out for me last week.

On April 20, the prediction for April 30 was 10°C:

Image: Accuweather

Chilly, but not unheard of in Tasmania, where I live. But over the next 10 days, AccuWeather just couldn’t seem to agree on what will happen on April 30.

On April 21, it added 11°C to the maximum:

Image: AccuWeather

A day later, it sliced 5°C off that:

Image: AccuWeather

Over the next four days, the prediction rose steadily 1°C a day to 19°C:

Image: AccuWeather

And on April 30, the day brought a high of:

Image: AccuWeather

As it turns out, AccuWeather actually nailed it on April 22, but got it wrong the other nine days out of 10.

Don’t get me started about rainfall

Too late.

In the past year, I’ve also learnt a couple of things about rainfall predictions.

If a prediction for rain more than two days out carries anything less than a 70% certainty, I can assume it’s a 0% certainty. I’ve watched a 25-50mmm prediction four days out dwindle to 1-5mm on the morning, and deliver nothing.

Inversely, a 70% chance of no rain is a 100% certainty of no rain. No problems there.

Let’s look at Friday’s rainfall, just in case you still don’t know what’s coming next.

It was a no-brainer because we hadn’t had a serious fall – maybe nothing – since the end of January. But it was obvious something big was coming on April 29, according to every weather site you looked at. Here’s how AccuWeather logged it for the 10 days leading up to the big fall:

10 days out – 2mm
9 days out – 2mm
8 days out – 20mm
7 days out – 2mm
4 days out – 27mm
3 days out – 19mm
2 days out – 18mm
Morning – 17mm

On the day itself, we got 24mm.

Let’s be generous and say the forecast of 2mm, nine and seven days out, was an anomaly, and discount it. That means AccuWeather’s worst rainfall prediction for April 29 came on the morning of April 29. So AccuWeather did a better job predicting rain four days out than on the day of the event.

I’m wasting everyone’s time here, I know. Weather forecasts have been wrong since weather began and I’m about the eleventy-billionth to complain about it. I’m nearly done.

So offering a 90-day forecast is clearly a master troll from AccuWeather, and congratulations, you got me. But at the very least I should point out that AccuWeather sells this as a premium product.

Would you buy it? Or any weather service beyond what the government-owned, paid-for-by-your-taxes Bureau of Meteorology already hands out for free?

Because they all have the same terms and conditions, which are, roughly speaking:

  • We hope this will be useful, but don’t rely on it
  • Especially not in emergency situations
  • Weather forecasting is unpredictable and not an exact science, and
  • Don’t hold us liable for being wrong

Week-out predictions of maximum and minimum temperatures on all of them vary wildly right up the morning of the day in question, and wind speeds are pretty much only correct in real time.

Last year, even the Bureau of Meteorology admitted 95% of forecasts were accurate “to within 3C”, but only for the next day. Apparently, with the help of a new $77 million taxpayer-funded supercomputer, it will become even better.

Until then, and very likely afterwards, the best way of predicting the weather remains the same as it did the day someone first knocked a hole in a wall.