Can Forecasting Help Me Staff a Specific Hewlett-Packard Call Center at 10:30 am on a Friday?

Jerry Shan

Tom Reilly

We can do a good job of predicting this, but the question is “Can we do better?”.  The Statistician W. Edwards Deming showed Japan’s auto industry the importance of increasing quality and HP aims to do the same.

How do you staff a calling center optimally so that you balance the phones getting answered and ensure employees are not just sitting around with nothing to do?

We can use reports and get averages of the days and semi-hours from the past few weeks and use that as an estimate.  We can use some body language and bump volume up or down around holidays from what we saw last year.  We can adjust for trends, yearly cycles, holidays and events.  We can adjust the expected volume by the type of call center due to prevailing market conditions.  In the end, it all becomes a complicated mathematical ‘factoring’ chore that needs constant love and attention when time series modeling can identify the relationships within the data and adapt to changes over time automatically.

By using daily data, we can get an overall view of the macro trends in the data.  We can use that data as a causal variable to use in 48 different semi-hourly models and forecasts.  This approach uses a combination of Box-Jenkins with Intervention Detection and Mixed Frequency Modeling.  Sounds complicated?  Isn’t everything?

What would you do?  Your thoughts and comments are appreciated.

Jerry Shan
Principal Scientist
Hewlett Packard

Tom Reilly, VP of Sales
Automatic Forecasting Systems

See JERRY SHAN from HP & TOM REILLY from AFS Speak in San Francisco at IBF’s:

$795 USD for Conference Only Group!

April 28-30, 2010 (3 Days)
San Francisco, California USA

2 Responses to Can Forecasting Help Me Staff a Specific Hewlett-Packard Call Center at 10:30 am on a Friday?

  1. I think to be the most successful, you can look at ways to predict your number of calls and the time each call takes more accurately. I also believe that attacking the problem this way is only half the solution. The truth is that you can only get a prediction so close before the model fails. Unknown and unknowable events have a large impact on your staffing that can make the difference between doing well and poorly. Instead look at ways to manage your business to take up the slack between the accuracy of your model and the actual call volume. Create a base staffing model that can take say 40% of your average call volume with full time 40 hours schedules. Fill in the dips and mountains with part time flex employees who work schedules in the 25-30 hour range. If your volume peaks on a given day beyond your prediction, you can offer to your flex staff up to 10-15 hours each without paying overtime. Have a few of your flex staff call in each day as on call employees. If you need them have them come in for the necessary hours. If your call volume flags, offer voluntary time off to both your full time and flex staff. Set parameters on utilization, occupancy and answer rate (service level) and have your management team work their staffing between the upper and lower specification level for each metric. It is all about working the numbers and having the tools to make changes in the environment to account for the micro changes in the staffing, volume and call duration areas.

  2. David,

    We agree that there is another half of this pie that you describe very well as that is what you do. What we do is the first half of the pie.

    Where you say “Unknown and unknowable events have a large impact on your staffing”, our response is: Detecting when these events have occurred in the past and isolating their effect allows you to model in a more robust manner. Identifying these events leads naturally to heretofore unknown causal variables which might then aid modeling efforts. At a minimum, the distortion introduced by these is alleviated.

    As for models failing, yes they do. Things change no doubt and how to model these “interventions” is key to being able to model these disruptions automatically so that the model will still predictably forecast the week AFTER the interruption is paramount.

    The Monday hourly pattern is different than a Friday. The question is can you model them separately while taking into account the MACRO effects like holidays, etc. and level shifts and interventions. That’s where it gets a little a hairy and where our expertise comes into play.

    Our analytic approach does not assume any prespecified weighting schemes but rather constructs a model based on the data: 1)optimal memory weighting (ie weekly/daily dummies) 2)lead/contemporaneous/lag structures around known holidays/events 3)4 Intervention variables which represent unspecified determinstic series (ie level shift, seasonal pulse, time trends and outliers)

    The best way to evaluate alternative strategies like ours is to do so with a case study. We welcome the opportunity to do this.

Leave a Reply

Your email address will not be published. Required fields are marked *

WordPress Anti Spam by WP-SpamShield