Is this a standard measure for determining accuracy?
EDIT: ah nevermind the questions about the formula, it's an average to smooth out bias towards under and over predicting.
I've been wondering a little about it myself. I'd heard that stock markets use a system where a 2% increase on day one and a 2% decrease on day two brings you on exactly the same value. While when you would normally use percentages it would be 100*1,02*0,98 = 99,96.
I searched for it and a site told me it is: (new - old) / ((new plus old)/2)
So I compared its effects to our own equation and it seems they are pretty much the same for the range: 50% underpredicting to 100% overpredicting.
More extreme under and overpredicting does show significant differences
Take real sales at 100
if you had predicted 0 or infinity our method results in infinity points while the other method results in 200 points max.
This other method penalties bigger mistakes in predicting less hard.