Please don't expect any large difference. It's an exercise in accuracy and consistancy and details.

Yes, It will be a lot of windows. It might be better to do in in C or C++ considering the size of the integers. I have no idea how long it would take to run, but of course that will depend on the window size of the prediction, and the size of that catalog (timewise) we are comparing it to.

There should be no partial windows anywhere, beginning or end. Let me give you an example. Lets say we have a catalog that is 60 seconds long. Lets say the prediction is for 5 seconds. The sample windows would look like:

Window 1 0-5 seconds
Window 2 1-6 seconds
Window 3 2-7 seconds
Window 4 3-8 seconds

and so on

Window last-2 53-58 seconds
Window last-1 54-59 seconds
Window last 55-60 seconds

We should have a total of 56 windows at the end.

No partial windows. No end of catalog to be tossed out.

I would only expect to calculate this for each new hit, not each new prediction. I think misses are deemed to be a big zero in the value department.

Michael