Slashdot Mirror
Statistical Accuracy of Internet Weather Forecasts
markmcb writes "Brandon Hansen considers the statistical accuracy of popular on-line weather forecast sources and shows who's on target, and on who you probably shouldn't rely. Motivated by a trip to a water park that was spoiled with hail despite a 'clear sky' forecast, he does a nice job of depicting deviations, averages, and overall accuracy in a manner that stats junkies are sure to love."
What a nice piece of work.
I can't help but smile a bit that MSN weather in this test turns out to be the absolutely worst when it comes to accuracy in almost all categories.
I would think there is a lot of difference on how the forecasts are made in the different channels, some of them probably do get a lot of their information from meteorologist working on their own stations. I wouldn't wonder if MSN doesn't have a meteorologist (or maybe only one) working to provide their forecasts.
Computers and simulations play a big role in predicting the weather today, but human eyes are worth a lot still.
I don't myself live in the USA, so my primary use of these are to check on when there is severe weather in areas where I know someone.
I have gotten used to check on weather underground for this information, I haven't checked on many other weather channel, but I feel quite well capable of following what is going on in the USA with tornadoes and such here from Denmark.
For a long time we only had one weather forecast service here in Denmark, a national institute. Since a primarily private TV station (TV2) a few years ago started their own weather forecast service, I really feel the national institute have been pulling themselves together and have provided many services that they didn't provide until now. So even though some of the services provides terrible accuracy they might still serve a good purpose in giving the other services competition and thereby forcing them to improve also.
When I am really dependent on the forecast I tend to study the information behind the simple prediction of the given weather, that way I am also much better prepared for possible scenarios, knowing which front move where and can better "read the skies".
I use the NWS website, mostly because I hate all the annoying flash ads on most of the other sites. I was also under the impression that most of these sites get their information from the NWS and pass it along to you (along with a bucket of ads). There was a lot of complaining amongst the popular weather sites when the NWS opened its own web site.
"Can't you see that everyone is buying station wagons?"
I remember some years ago a radio presenter saying that you could achieve greater accuracy than supposed weather forcasters simply by using the assertion: today's weather will be the same as yesterday. Have we moved on from this position?
Any sufficiently advanced bug is indistinguishable from a feature.
You'd need a lot of data for that one, if you want to establish the accuracy of the probabilities. Unless I'm missing a more mathematically clever way to do it, I'd assume you'd require lots of 20% days to determine whether rain happened on anything near 20% of them. Similarly, 10, 30, 40, 50, etc would require their own groups of lots of days.
It'd be even trickier in, say, the SF Bay Area, where it only rains for two or three months a year, and then almost every day. Your 0% and 80-100% groups would be well-stocked, but not so much the other ones.
I am not a meteorologist, but I have worked with them a few times.
Generally the competing weather models will show a range of possible outcomes with various probabilities. You can average across all scenarios and come up with a 60% probability of rain, but the more days out you go more the scenarios diverge, so the less useful a single average will be.
Most people would not find it useful to hear that "there will be probably be thunder on Wednesday if it remains hot enough, but if it cools down on Tuesday then the thunderstorm will be off to the north somewhere"
Additionally, a lot of weather conditions are influenced by thin layers of cloud high up, so thin that precise measurements are critical so precise forecasts in one location more than 3 days out are difficult.
for my area is that they are usually accurate down to a period of about 3 hours. As an example.
1 79 predicted that it would be snowing yesterday morning by 0600. Sure enough, I woke at six, and it was snowing. I awoke earlier in the night (about 0400) and it hadn't yet started.
http://www.bbc.co.uk/weather/24hr.shtml?world=4
Similarly, that site predicted that the snow would drop off by noon, and turn to sleet or rain by 1600. Again, this prediction came true, within an hour of the predicted time.
Generally speaking I find the BBC weather site to be accurate significantly more often than not (guesstimate 80% accuracy) with the 24 hour forecast being almost universally correct, and the 5 day forecast being the least reliable. (as expected)
This is a FAR cry from the weather predictions when I was a lad. Then the weather forecast on TV was simply a way to poke fun at the meteorologist, who clearly was doing the best he could, but invariably got it wrong.
I was quite curious about weather forecast accuracy as well. So three years ago I started collecting weather forecasts from the primary providers (Accuweather, Weather Channel, NWS, CustomWeather, Intellicast, etc.) and comparing them to actual observations. It's tougher than you might imagine, and there are a lot of factors that need to go into creating usable verification statistics.
I have a public site with some statistics for about 800 locations in the US available at ForecastAdvisor.com. There is also a blog with more in depth analysis (like how do temperature forecasts fare relative to how deviant the actual temperature is...in other words how well do forecasts do the further away from normal the actual is, and how to forecasts fare the further out they forecast for, and how does forecast accuracy compare over time.).
ForecastWatch.com is used by meteorologists and professionals. Accuweather, The Weather Channel, and several private meteorological companies use this system to help them understand and improve their weather forecasts.
And a geek note: ForecastWatch.com runs on Quixote (a Python web framework), while ForecastAdvisor.com runs on Ruby on Rails. The back-end forecast and actual collection, and calculations are Python with a MySQL database. Both sites are close to migrating to Django, a new Python web framework and ORM.
-Ace