Skip to main content
SAGE
Search form
Auto-Scroll
  • 00:00

    [MUSIC PLAYING]

  • 00:11

    ANDREW ZELIN: My name's Andrew Zelinand I'm a statistician with over 20 yearsexperience in a number of sectors-- market research,telecoms and health.Much of my work has involved gaining actionable insightsfrom the data to keep costs down and help businesses run better.And the reason why I'm talking about call centers today

  • 00:34

    ANDREW ZELIN [continued]: is that call centers are a big part of the world economyand ever increasing worldwide.So it's important to make sure these businesses runeffectively and efficiently.I'll be discussing how it's possible to predict, quiteaccurately, the number of calls that a call center can expecton any given day, week or month, and help

  • 00:58

    ANDREW ZELIN [continued]: match resources to that.Let's say you're offering a servicebut you're finding it difficult to knowwhen demand is at its quietest and its busiest.So how can you get it right?Well a model, generally, is a hypothesisabout how things work in one's mind to relate to reality.

  • 01:21

    ANDREW ZELIN [continued]: It may be right.It may not be.And the model, in this case, is our hypothesisabout the relationship between the number of calls thatcome in on a particular day and whether that day isa Monday or a Tuesday or in the middleof November or on a school holiday or on a bank holiday.

  • 01:43

    ANDREW ZELIN [continued]: So we have this conceptualized situation in our mindthat Mondays are busier by a certain number of calls,Tuesdays are quieter by a certain number of calls,and so on.To put flesh on the bones on the model,one carries out a multiple regression analysis.And the multiple regression analysis

  • 02:05

    ANDREW ZELIN [continued]: would take all the data you've got over all the three years,or however many it is.It would say whether that day is a Monday or a Tuesdayor on a bank holiday, and how manycalls you've got on that day, and link the two together.It would then select which driversare statistically significantly related to calls

  • 02:28

    ANDREW ZELIN [continued]: on a given day as opposed to those which aren't, or areonly indirectly related.So you'd come up with a subset, a smaller number, of drivers,but the ones that are significant.And we're going to use our model from the ones thatare statistically significant.So the coefficients of these modelscould then be used in the forecast.

  • 02:52

    ANDREW ZELIN [continued]: Once the data's been analyzed, it'sgood to understand why you're getting what your getting.Does it make sense that Mondays are busier than other days?Does it make sense that claims go up in the winter?Also, if there's a spike in demand,do we know why this is the case?

  • 03:12

    ANDREW ZELIN [continued]: Was there additional press activity or somethinggoing on with the weather?Is it something that's happening transiently or goingto continue?And you'd need to build into your forecast accordingly.So let's say you're currently in February, 2015,

  • 03:32

    ANDREW ZELIN [continued]: and you want to know how many calls thereare likely to be on Friday the 14th of August, six months'time.We already know quite a lot.You know it's going to be a Friday,you know it's going to be in August,and you know it's going to be in the middle of the schoolholidays.And your model, your multiple regression model,would probably have told you what the Friday effect is,

  • 03:53

    ANDREW ZELIN [continued]: what the school holiday effect is,and what the August effect is.How far back do we need to go?How much data should we have collected in orderto have a reasonable level of confidence on our forecast?Well, obviously the more data we've got,

  • 04:15

    ANDREW ZELIN [continued]: the further back, the greater the level of confidence.If we're taking into account seasonal patterns,then I would say at least three years-- enoughto know whether Januaries are consistently busieror quieter than the rest of the year.If you don't have anywhere near that length of data,

  • 04:38

    ANDREW ZELIN [continued]: you can still do simpler models basedon day of the week and increasing or decreasingtrends.If we look at a couple of examples,the lower chart shows the further forwardyou try and predict, the less confidence or precision

  • 04:59

    ANDREW ZELIN [continued]: there is about your forecast.Now if you look at the upper forecast, the upper chart,what's happening?Is there an increasing trend of volume over time?It's tempting to put a line of best fit through that increasesdiagonally.

  • 05:20

    ANDREW ZELIN [continued]: But then if you look a bit more closely,you find it's only the first half of the time serieswhere this increasing occurs.The second half, it tends to level off.So the question is, what do you thinkis going to happen in the future?Is it going to continue to level off,or is it going to start going down or go up again?

  • 05:43

    ANDREW ZELIN [continued]: And this comes back to the fact that the maths doesn't tell youthe whole story.You need to understand or speak to people in the businessabout why this increase has come about, why it's leveled off,to make an informed understanding of what'sgoing to happen in the future.

Video Info

Publisher: SAGE Publications Ltd.

Publication Year: 2017

Video Type:Video Case

Methods: Statistical modelling, Multiple regression

Keywords: call centers; demand characteristics; drivers

Segment Info

Segment Num.: 1

Persons Discussed:

Events Discussed:

Keywords:

Abstract

Statistician Andrew Zelin explains how a model can be used to forecast demand at a call center. He cautions viewers that math alone can only do so much; it's also important to have a human understanding of the field.

Looks like you do not have access to this content.

Call Center Forecasting: Linear Regression Models

Statistician Andrew Zelin explains how a model can be used to forecast demand at a call center. He cautions viewers that math alone can only do so much; it's also important to have a human understanding of the field.