Back at the beginning of summer, the Curmudgeon conducted a little experiment to test the accuracy of long range weather forecasts. Our conclusion, after a brief test of the data, was that the five and ten day forecasts typically found on a weather website such as The Weather Channel's Weather.com amount to no more than a toin coss in terms of accuracy.
We also said we'd extend the experiment and report on the results later. Herewith, a report on the more extended data. We'll make another run at some point in the winter months to see if the long term forecast fares better in colder weather.
(Today, the area's weather forecasters predicted intermittent steady to heavy rain throughout the day. They were wrong. By noon today, the rain will have moved out of the region, leaving us with a cloudy and damp--but not rainy--afternoon. Certainly not the washout that was forecast.)
To recap the experiment, we asked the question: What if I want to plan an outdoor activity in Arlington between 1-6 pm on the day in question? If the forecast was for a 50% or greater chance of rain, we would plan to stay indoors. If less than 50%, we'd plan for outdoors. We also gave Weather.com a generous +/- 5 degrees on the temperature forecast before we'd say the forecast was wrong. (I.e., if the actual temperature was within 5 degrees of the forecast temp., we said the forecast was accurate.)
In our initial experiment, the 10-day forecast was correct only 43% of the time--worse than a coin toss. The five-day forecast was better, being right 61% of the time, which is still not much comfort if you really needed to rely on it.
We extended our observations to roughly six weeks--from June 9 to July 20--and here's how we came out.
Ten Day Forecast
The ten day forecast improved a bit over the longer observation period, but still was just barely better than a coin toss. Out of 35 observation days (some days we had no data), the 10-day forecast was correct 18 times and incorrect 17 times, for a 51.4% accuracy rate. In short, there's really no point publishing a 10-day forecast with the current state of the art.
Five Day Forecast
The five day forecast stayed the same in terms of accuracy. It was correct on 23 out of 37 observation days, for a 62% accuracy rate.
What may be the most interesting observation here is that the five day forecast really wasn't much more accurate than the 10-day forecast. We would've expected that the five additional days would result in a vast improvement in the forecast, but it didn't. In other words, there's really not much point in having a 9-day, or an 8-day, or a 7-day or 6-day forecast either, because the extra days don't help the accuracy meaningfully.
In contrast, our observation was that the one-day forecast--i.e., the forecast for the following day--was generally quite accurate, exceeding 90% (the missed days were almost always due to a short-lived thundershower). Thus, the accuracy rate does improve considerably between the five day and the one day forecast.
We'll see how the experiment goes in the winter months, when weather patterns can be more stable. For now, however, our conclusion is this: weather forecasters are little better than a monkey with a coin when it comes to predicting the weather more than 5 days out.