AnalyticBridge
Data Intelligence, Business Analytics
A great statistical model for weather forecasts
For places such as the Pacific Northwest, UK or Germany, where weather is highly volatile, I believe that my mentor's model (when I did my PhD) might actually be the best one. In places where weather is easy to predict by looking at satellite images, you might be able to do well without any model.
The model described below is based on data from the Puget Sound (Seattle) area. For many years, weather forecasts for the Seattle area have been notoriously erroneous. They've been more often wrong than right, resulting in schools not trusting weather forecasts to decide on whether closing or not. This in turn has caused many problems.
This week forecasts is a culmination on erroneous forecasts. Mon-Tue were supposed to be light snow days, and Wed to be the "snowstorm of the century". Indeed, nothing happened Wed and we got a fair amount of snow Mon-Tue.
To make things worse, the meteorologists from University of Washington claimed they were absolutely certain about their forecasts, and had a perfect understanding of what was going to happen. Even on weather.com, you could read "severe storm warning", "100% chance of precipitation", "up to 14 inches of snow", "heavy snow", "100% chance of snow", despetite the temperature being very close (just a bit above) 32 degrees. No confidence intervals of any kind were provided.
So let's now discuss my mentor's model, who also holds a PhD in statistics and has studied in Cambridge and Australia. The model is as follows: tomorrow will be the same as today.
Since this model is better than what meteorologists can do, are they only there for entertainment?
Views: 679
Comment
-
Comment by Vincent Granville on January 20, 2012 at 3:58pm -
The culmination of "worst forecasts" actually took place the next day: by making Wed looks like "the storm of the century" and Thursday the day where supposedly it would get much better, it forced many people to wait until Thu to do their grocery shopping. Then came terrible Thursday with its ice storm: it trapped people in their home with no light, no Internet and no heat. I guess after this week nobody will ever believe them again.
Note that the basic model defined as "tomorrow will be the same as today", one more time, proved superior. Prof Cliff Mass, meteorologist at University of Washington, profusely apologized, saying it's a bad day for weather forecasting.
I also believe that their models (at least those used to warn the public) are not real time, but are updated no more than 3-4 times a day. They haven't figured out yet a system that measures temperatures delta, wind velocity and direction delta in real time, and that generate warnings updated in real time.
-
Comment by Vincent Granville on January 18, 2012 at 10:06pm
-
All models should be compared with this baseline model (tomorrow = today), to assess meteorologists performance. If they can't do better, their "return" is negative, their models provide no lift.
Also note that even Tue night at midnight they were still predicting that the "storm of the century" would arrive within the next few hours.
Top Content
Follow us
© 2013 AnalyticBridge.com is a subsidiary and dedicated channel of Data Science Central LLC
You need to be a member of AnalyticBridge to add comments!
Join AnalyticBridge