ForecastWatch.com January 2006 Customer Newsletter
Welcome to the January 2006 Newsletter of ForecastWatch.com. You are receiving this email as a subscriber to a ForecastWatch.com product or by signing up for ForecastWatch news. Please see the bottom of this email for instructions on how to unsubscribe from this newsletter.
Normally, weather forecast accuracy numbers are calculated by taking a forecast and then looking at the actual and seeing how close the forecast was from the actual. But what if you looked at actuals, and grouped all the actuals together that shared some characteristic, and then looked at how well forecasts did?
We did just that. We grouped actual temperatures by how different they were from climate normals. For example, all days where the temperature was 10 degrees above the climate normal were grouped together. We then looked at all forecasts for those days, to see how well forecasters did on average for each of those groups. By doing that we were able to answer questions like "How accurate are weather forecasts when the temperature ended up being ten degrees above normal?" We can also look at other things like bias. Are forecasts equally biased, or does bias change over different temperatures above and below the climate normal?
You can find the article at:
http://www.forecastadvisor.com/blog/2006/01/weather-forecasting-extreme.html
ForecastWatch helps businesses and individuals understand and place value on weather forecasts so that they can be more accurately used quantitatively in modeling and prediction. This is just one small example how.
All of 2005 data has been re-audited and re-aggregated and the final 2005 numbers are in, including December 2005 data. Over 19 million forecasts (19,032,834 to be exact) and 310,253 actuals have been combined to create the world's largest repository of weather forecast accuracy information anywhere! We are working on the 2005 annual report which will be available soon.
In 2005, 68%% of one day out high temperatures and 61%% of one day out low temperatures were within three degrees. 12%% of high temperatures and 11%% of low temperatures were exact. The difference in accuracy is most likely due to timing. Since forecasts are collected in the evening, the one day out high temperature is about 24 hours away, while the one day out low temperature (since it occurs overnight "tomorrow") is about 36 hours away. In 2006 we will begin collecting the "zero day" forecasts, which would be the overnight low for the day the forecasts are collected.
In 2005, average one day out high temperature error was 3 degrees for high temperatures and 3.4 degrees for lows. For five days out, high temperature absolute error was 5.16 degrees, while it was 5.26 degrees for lows. Average one to five day out high temperature bias was 0.21 degrees, while it was -0.21 degrees for lows. That means, on average, high temperature forecasts were 0.21 degrees too high, and lows were 0.21 degrees too low.
For climate watchers, since we use 1971-2000 climate averages in our unskilled forecast calculations, we can compare climate normals to actuals. For that, actual highs were 0.03 degrees lower than climate normals, on average, but actual lows were 2.38 degrees warmer than climate normals. If you assume that average temperature is the average of highs and lows, the average temperature would be about 1.18 degrees above 1971-2000 climate normals. That number is similar to the 1.2 degree anomaly reported by the NWS against the 1895-2004 climate normals.
In 2004, high temperatures for those 800 stations were 0.63 degrees below climate normals, and low temperatures were 2.41 degrees above climate normals. This averages to 0.89 degrees above the average 1971-2000 climate normal temperatures. This is similar to the 0.7 degree anomaly reported by the NWS for 2004 against 1895-2003 climate normals.
Now, this is for just 800 stations, with heavy location bias around population centers, so it may or may not be representative of the climate as a whole. Also, its against 1971-2000 climate normals, which may or may not be a representative period for climate. It is interesting to note that, at least for these stations, the increase in average temperature came solely from an increase in average low temperature, not an increase in high temperature.
Please let us know what you think at ideas@forecastwatch.com . We welcome your ideas and suggestions to make our services better and more useful.
