Yes, we start off looking at a very small amount of error data at the start because there is only 6 hours worth of error data at the start. That really tells us very little because all the models will have around the same error. We want to sample a longer period of error data as soon as possible and do that over time until we finally always start seeing how a model's five day forecast does. The difference between how it was and how it is now is that rather than averaging up to two model runs for around the first five days, I average four model runs shortly after model data is first available. Also, we drop the first runs once we get a little bit more error data since there were less models released at the start and they are more likely to get better after a few runs anyway as the storm gets better organized, more models start picking up on it and you might have better data going into the models as well.

After 7.5 days, there is enough data to determine how well a model's five day forecast is performing over the last twenty four hours, which would include up to four runs of a particular model. (models that come out every 12 hours would only have up to two runs averaged) While we do have a five day forecast sooner than that, we wait an extra few days because every 24 hours we dropped a previous model run from the calculation. By doing it this way, we do not put a lot of weight on the first few model runs, which may not be that great anyway since a storm just started.

The old way got to using the longer period sooner, but at the expense of using only two model runs for the first five days and putting the weight on the first few model runs. I think the new method should be better. Although, even so, any model can sometimes be right.

I'll have to try to work on coming up with how to have an explanation on this page eventually on approximately how it works:
http://www.hurricanecity.com/models/data.cgi?page=bestperformingmodels

Something like:

After about a week, the best performing models map calculates how well a model's five day forecast is performing over the last twenty four hours. Prior to that, there is less than five day's worth of forecast data using the method we use. The BAMM and LBAR models are displayed for the first twelve hours after models has been released. Starting at twelve hours since model data was first released, more and more data can be used to determine how the model's forecast is doing. After twenty four hours since model data was first released, there will always be up to four models runs over a twenty four hour period that are averaged for the calculation. We can continue to increase, by six hours, the forecast we are checking the accuracy of for about the first week, expect that every twenty four hours the oldest model run used for the calculation is dropped to try to put less weight on the earlier few model runs. When we do that we do not increase, by six hours, the forecast we are checking the accuracy of. After about a week, we always start dropping the oldest model run every time new model data comes out since we no longer increase the forecast we are checking the accuracy of beyond five days. Over the latest short term period of twenty four hours, we can then always determine which model has performed best in its five day forecast.