Truth About Accuracy, AccuPOP, New Model Tools
I got this question from a blog reader yesterday afternoon: "I am planning an outdoor event in Bethlehem, Pennsylvania on Saturday and would like to know how accurate the forecast is that far in advance?"
It's hard to say how accurate weather forecasts really are -- everybody's got different numbers and it varies with how well the Forecast Models [JessePedia] agree with each other, the proximity to geographic effects such as mountains and bodies of water, and even population. The truth is, forecasts tend to be more accurate where more people live (or in the case of private weather companies, where their clients are) because more attention is paid to verification there. And, it depends on what you mean by "accurate": Are we talking about +/- 1 degree? 5 degrees? What about rain and snow?
That said, I believe that weather forecasts in general are fairly accurate, probably above 90%, up to four days out, where "accuracy" is defined as what the average person can discern (i.e. people can't tell the difference between 70 and 75 degrees but usually aren't bothered by less than 0.10" of rain). This is mainly because when you look at the dozens of computer models out there, the most common length of forecast is 4 days out, so beyond that you get less and less "opinions" on the forecast.
When you're trying to plan outdoor activities, you want to look at probabilities, not specific forecasts. AccuWeather.com has a great tool to do this, which I've mentioned before in regards to predicting ice and snow probabilities, but it works for rain too. It's called AccuPOP, and you can see the next 24 hours of it on AccuWeather.com, with 96 hours (4 days) on AccuWeather.com Premium (for which we have a 30-day free trial).
For more examples of AccuPOP and tutorials on how to interpret it, read my blog entry from last winter.
Here's how Bob's AccuPOP looks as of this writing:
BOB'S ACCUPOP FOR BETHLEHEM, PA FROM AccuWeather.com PREMIUM
Looking at the current AccuPOP as of this writing, it looks like Bethlehem has a decent (50/50 shot) of rain on Saturday. If he's planning an outdoor event that absolutely can't have rain, it's not looking good. The most likely time for the rain is between 6 and 9 am but it's possible (30%) almost all day. Not good news for Blog Reader Bob. On the other hand, the amount of precipitation is only 0.06", which is not a downpour, though this time of year you'd also have to worry about an all-day drizzle.
This is a good place to start for planning, but Bob should keep an eye on the AccuPOP forecast as we move towards Saturday because, like all forecasts, it may change. But if he has to make the decision today, it doesn't sound like a good date if his event is not rain-tolerable. If moving the event to Sunday is possible, I would highly recommend it.
AccuWeather.com (24 Hours) | PREMIUM (96 Hours) | PRO (96 Hours)
While we're talking about forecast probabilities, let me tell you about two new tools recently added to AccuWeather.com Professional.
The first is probability maps from the "SREF Ensemble" model. We've talked before about how running a forecast model with slightly different initial conditions can produce wildly different forecasts. When you do this a few times in a row, then merge the results, you get a pretty accurate probabilities map. If we look at the probability of more than 0.0001" of rain (in other words, a drop), the map looks pretty depressing for eastern PA with an 80% chance at 2 PM on Saturday:
But if we increase that amount to 0.25", which I would call a "significant" rainfall likely to adversely affect any outside event, it doesn't look so bad, with only a 20-40% chance.
The other recent addition to the Pro site was the extra-high-resolution WRF model, which we debuted last month. The model has been turned off in favor of hurricane models this week by the government, so I can't use it to illustrate Bob's forecast. But let me just say, this model has incredible resolution, meaning that it can be extremely precise, showing small differences in temperature over minor elevations and bodies of water, and pick out individual thunderstorms.
At first blush, this sounds great, it's like a 48-hour PredictiveRadar! But I decided to keep tabs on it about this time last week, for the big Penn State game on Saturday (I Don't Do Sports [JessePedia], but this game affected a lot of people). On the left below is the WRF forecast for Game Time Saturday night, on the right is the actual radar image.
Not so great. Of course, this forecast was at the end of the 4-day period, but this is proof that, just because the model can predict the weather better in general, even the best models won't be able to put those high-res storms in the right place when it comes down to it.
Still, this extremely-high resolution is intriguing and I look forward to the future in forecast modelling, now that we can get a realistic-looking picture from them. (For example, check out the courser GFS and 4-KM WRF comparison below - even better than the high-res (12KM?) NMM-WRF that I reviewed last year).
