Weather forecasts suck

And yet, it is one of the most common things we’re doing in prediction. Here are a few things I learned about weather forecasts that matter for all forecasting & prediction tasks:

1. Weather forecasts are so inaccurate, mostly because we can’t replicate complex systems with many moving parts. Like… say, businesses and industries, and about anything in our world, other than a few selected isolated local tasks…

2. Weather forecasts can be made more accurate locally, by incorporating lots of local data. If we invest in a single point of prediction, we can make it better by using lots of specialized data. ⇒ Good predictions will always be expensive.

3. A 10-day forecast is a coin flip; a 1-day forecast is pretty good! (96%+ accuracy) ⇒ Forecast more, and you’ll be at completely random predictions pretty fast. Don’t think that just because your prediction works on a short horizon will work for a longer time.

What I took away is that forecasting is hard! And expensive! But it can work. But it looks like there’s no easy way out of this one.