In an exchange on the iSixSigma Discussion Forums, a representative from a telecommunications firm was interested in learning about what methods are best to improve forecast accuracy at the company’s call center. After some discussion about defining the goals of the project, the responding Six Sigma experts then offered their views about using such tools as general linear models and multiple regression analysis, and also some advice about improving call wait times.

Message: 186701
Posted by: Leopard
Posted on: April 26, 2010

Hello. I’m currently working for a telecommunication company, and we’re forecasting call volume and staff requirements…at least we try to. However, when measuring forecast accuracy results don’t exceed 80 to 85 percent.

How could I use Six Sigma in order to improve this kind of situation? I was thinking about DMAIC approach, what is your opinion to that?

Currently, we use calculations in Excel. We measure daily and weekly distribution of call volume, using historical data, in some cases apply seasonality factor, exclude special days (bank holidays), and finally use monthly and annual growth rate in order to apply relevant increase or decrease in next month’s traffic.

Kindly let me know if you have any ideas on this matter.

And how could I possibly use Six Sigma method to improve this situation? Any suggestions will more than welcome! Thank you in advance for any reply!

Message: 186711
Posted by: Jsev607
Posted on: April 27, 2010

Why don’t you start by defining what you would consider to be acceptable forecasting accuracy? Keep in mind also that rather than improving the accuracy of your forecasting you would most likely be better served in looking at your entire value stream and eliminating non-value added activities so that you can handle variation in call volume without the need to adjust staffing levels.

Message: 186712
Posted by: Leopard
Posted on: April 27, 2010

If I understand correct, you mean starting first with presenting our goals, such as forecast accuracy >92 percent, in monthly base. ?

However, I cannot overcome staffing requirements since they’re fully connected with call volumes, correct? ?

Should I use maybe another forecasting method? In my company we use currently a WFM [workforce management] tool, but we cannot evaluate forecast figures well at the end. ??I’m a little bit confused with this. What’s your opinion to that?

Message: 186714
Posted by: Sensei
Posted on: April 27, 2010

Is “forecast accuracy” the goal, or isn’t the “customer wait time” and/or “dropped call rates” really the goal??? Once you determine your Big-Y, set goals, normality, stability, capability, etc., you will probably eventually get to either general linear models (which can handle both discrete or continuous Xs), or multiple regression (for continuous Xs) to determine which measured Xs (if any) impact your Big-Y…??Good Luck!

Message: 186716
Posted by: Leopard
Posted on: April 27, 2010

Well, the goal is to achieve forecast accuracy, since, if we can achieve being as accurate as possible, we will achieve at the same time schedule adherence, reduction ACR and improvement of FCR. ?

Finally, I want to compare forecast vs. actual call volume; service level and abandoned call rate achievement; and scheduled FTEs [full-time equivalents]. Then, I will show the impact to the whole business. Please correct me if I’m wrong.

Message: 186717
Posted by: Sensei
Posted on: April 27, 2010

I suggest that your goal – CTQ, Big-Y, whatever you want to call it – is reducing customer wait time. That wait time, in turn, is a function of by-call volumes, FTEs to answer calls and call lengths when answered. There is also going to be variation inside each of those components (inaccurate forecasting, FTE-to-FTE differences, sick-rate of FTEs and call-to-call resolution time). ??So don’t get paralyzed worrying about your accuracy rate. Start some regression analysis or general linear models (GLMs) to determine which of those Xs is the biggest impact on the Y and why. Maybe it turns out your effort would be better spent on FTE training and tools to reduce resolution times and your 80 percent volume accuracy isn’t painful anymore since your customer wait times will go down… Just sayin’…

Message: 187209
Posted by: Union of Conjoined Scientists
Posted on: May 19, 2010

You’re going to run into a couple problems if you’re trying to use regression analysis; regressions are only as good as the factors you have as inputs to them. If you’re not capturing all your input variables (you mentioned a few above), you’re not going to get an r-sq large enough to build a model of any value. And even if you did, it wouldnt be beneficial for daily forecasting. You could try Monte Carlo simulation, but you’d have to create several different models for day of the week, week of the month, month, etc.

In summary, you can definitely use statistical tools to improve your forecasting, but you’d be way better off looking at improving your call centre processes instead – make them robust enough to deal with reasonable variation in call volume changes. Because no matter how tight your forecasting is, you’re still going to have common cause volume variation that would be essentially impossible to account for. ?I’d focus on improving your average handle time and first-call resolution via DMAIC.

About the Author