Whilst it is relatively easy to achieve a weekly forecast accuracy of 5% this does not really mean a forecast was particularly good. Depending on how you calculate the forecast variance a weekly variance can mask many issues, not least the range of the variance by day and variance by interval.

My question is what forecast accuracy do you targetin your centres by Week, Day and interval and what is your calculation for this, do you report varaince, weighted or absolute variance?

Thanks in anticipation.

Depending on your call volume, it can be very difficult to be accurate once you get down to the time interval level. Most forecasting models work best with call volumes in the thousands. So when you analyse at a calls/week level it is really easy to be within 5%. Once you analyse this breakdown at the daily or interval level then the variance will creep up. One way to manage this is to run a real-time feed into a forecast model that will shift the projected line accordingly. This is best married with a back-office structure that will allow you to log extra agents on as required or move agents off into other tasks.
As for targeting accuracy, I prefer to target accurate prediction of the main peaks and troughs and bulking staff accordingly. As you've intimated, targeting weekly accuracy is pretty pointless as you can achieve near perfect accuracy in figures but the actual impact at the front-end could be quite different. Unless you are getting hundreds of calls per hour, I do not think you will achieve great interval level accuracy.
A good way to measure variance is to plot your variance on a "Statistical Process Control Chart". This will give a good indication of where you are particularly good/bad at forecasting
Hope this helps formulate some thoughts without having gone off on too much of a tangent. I have tried to be brief and so there is a lot more that could be said on this topic.

Agree with Ian.
Forecasting the peaks and troughs is more important.

I would consider a 10% variance as the limit & anything below 5% is good. More important is when the forecast is done. Forecast on which your roster is done is an important factor because that tells you as an analyst if you need to ask for additional marked time or if you need to rope in staff from other teams expected to have a lull period.

Once the roster is done, a weekly forecast done after that (close to actual period) may be more accurate in terms of Variance % but how much use is it going to really be? You may be able to adjust the shift & break timings a bit here and there (which you should) but with respect to having more hands on team (or less hands on team) it is too late to really benefit the activity.

Forecasting-Scheduling-Evaluating is a cyclical process and although we as planners hate to evaluate our forecast in terms of Variance, it is imperative for us to be honest with ourselves.

Cheers!

Kex

Steve:
Your desire to be more granular about evaluating forecast accuracy is admirable. You concern that forecast accuracy may ignore the true customer experience is insightful.

One of the most significant problems with forecast accuracy as a KPI is that forecasts easily become self-fulfilling prophecies even when the forecast is quite distant from reality. In general, if we under forecast, our forecasts will come true. This is a side effect of something we call capacity based forecasting.

Any time demand outstrips our capacity to handle calls, the wait time and abandon rates go up but our ability to handle calls is limited by the number of agents that we put in place. If we are under staffed, the handled call statistics that get generated will lie to us about how many agents we should forecast for in the future. If we staff according to this distorted forecast then our forecast will seem to come true according to handled call statistics.

Lets say our forecast is not bad but we under-forecast only between 12:00 and 12:30. Our forecast accuracy will be close to 100% between 12:00 and 12:30 and much less accurate during other periods of the day for which we planned correctly. Also notice that the true workload from 12:00 to 12:30 gets shifted into later periods causing future forecasts to give us an inflated impression of the resources needed in the later intervals.

If you would like to learn more about the science of overcoming forecast distortions and measuring true forecast accuracy, please contact me and I’ll set you up with a web based video tutorial on the subject.