Forecast Variance Calculation
Starting to wonder if I have too much time on my hands. In my defence though this is something that we have had to look at before.
Following on from Dylan's point on accuracy there is a way around this to calculate the accuracy of a forecast in a single period (1/4 hour period, hour, day etc) and to take these periods into account when looking at the accuracy of a forecast over a number of periods (say a month).
I have prepared the below table to demonstrate the calculation.
My calculation is basically the inverse of Dave's for individual periods (days in this example but could just as easily apply to 15-min, half-hour or hour periods). When looking at the overall accuracy as the forecast was overall only 5 calls out then it looks like the forecast was only 4% off. To take the individual periods into account we take the actual from the forecast in each period and square it, sum those squares and take the square root from that figure.
The accuracy of forecast figure is then 1- (square root of the sum of the squares of the variance in forecast versus actual call figures).
This essentially measures the accuracy of forecast across the whole period and takes the wild variations into account.
For Actual Eamon Forecast - actual
Mon 10 8 -20% 2
Tue 11 12 9% -1
Wed 23 14 -39% 9
Thu 34 34 0% 0
Fri 12 12 0% 0
Sat 15 26 73% -11
Sun 7 11 57% -4
Overall 112 117 4% -5
Sum of Squares 223
Square root 14.9
% Accuracy of forecast 87%