Gordon wrote: ↑Sun Dec 09, 2018 11:11 am
Speaking of ensembles HB, how do rate WATL?
Very good. It’s an optimal combination of most available weather models. But, you need to look at the range and the probability of exceedance products.
It’s pretty easy to show mathematically that if your ensemble is reliable (captures the true spread) then you can’t beat its mid point (or more precisely the mode or highest probability outcome) and your forecasts error will equal the spread of the ensemble. If you look at the current guidance the mid point rainfall for Ferny Creek and the spread of the ensembles for the week ahead is about 50mm and about 20mm (this is the standard deviation). That tells me there’s about a 95% chance the final number will be between 10mm and 90mm or 2 standard deviations and about a 68% probability that the rain will land between 30 and 70mm. These are not precise, but should be pretty close. The WATL has a probability of exceedance forecast which basically does this for you. I’d be happy with something in the 30 to 70mm range, and anything above that is a bonus
One thing this tells you this that you can’t on average beat the spread unless you know how it is wrong. That’s a very tough mathematical problem. You can do this a in your head if there’s a consistent pattern of model failing - for example in the models don’t capture drizzle as the mountains in the models are smoothed and will look like 100-200m size hills, rather than and sharp 400-600m wall.
When I see large spread I shrug my shoulders and accept that’s quite simply that the situation is so chaotic and that is the limit with current computers and observations, but you know that each model run you get closer the narrower the spread. That wiggle watching and never quite knowing is what makes the weather so interesting