• Black Facebook Icon
  • Black Twitter Icon
  • Black LinkedIn Icon

© 2019 Vemo, Inc. All Rights Reserved - Privacy Statement

Please reload

Recent Posts

BPI Announces Vemo CEO Peter Louch Best Practice Award Winner!

December 7, 2016

1/1
Please reload

Featured Posts

Demand Forecasting and the Value of a Straight Line

June 22, 2018

When you have a large workforce divided into tens or even hundreds of granular groups, forecasting headcount needs can be difficult. A good baseline forecast usually relies on recent trends, but where a workforce is highly differentiated, growth trends tend to be as well. Some employee groups may be growing steadily; others may show sudden surges in growth that quickly dissipate; still others may be declining as shifting business priorities decrease their necessity. A good baseline forecast must take this variation into account.

Enter Vemo’s demand model. Each time a client refreshes their data, our solution combs through the recent history to identify growth trends at the granular workgroup level, then project these trends into the future. This baseline forecast essentially says ‘based on our past behavior, these are how our hiring patterns will look in the future.’ The client can then edit this baseline forecast using business insight, or explore how it might change under a variety of what-if scenarios.

One ingredient that makes our forecast more accurate is our appreciation for simplicity. Faced with fashionable new machine learning tools in R and other packages, an overly-enthusiastic data scientist may be tempted to fit trends using the most complex non-linear methods available. But in our experience, in the domain of demand forecasting, it’s generally best to KISS.

 

Here’s an example adapted from one of our clients. In this example, we are using data from months 1-12 to predict the upcoming year of headcount growth (months 13-24). Again, out solution typically combs through tens or hundreds of groupings to identify trends, but in this example we examine just three example groups. A few things pop out when we look at the raw data the model has to work with:

  • Unit 1 has a fairly clean linear growth pattern

  • Unit 2 is much more volatile but trending downward

  • Unit 3 exhibits fairly stable linear growth until month 10, when growth increases dramatically

 

Next, we fit trend lines to the three units. We fit linear trends to units 1 and 2, but for Unit 3, given the departure from linearity at month 10, we first try a smoothing method from R’s Generalized Additive Modeling package in order to capture the non-linear shape of the data:

 

 

These approaches appear to capture the shape of each trend, but let’s see how each trend-line performs when projected forward to forecast the next year:

  • For Units 1 and 2, the linear projection performs well – projected headcount in month 24 (represented by the line) is very close to actual headcount in month 24

  • For Unit 3, the recent uptick in growth has been projected into the future, but this appears to have been in error. Projected growth skyrockets beyone the scope of the graph, while actual growth regresses to approximately the same rate exhibited in months 1-10, prior to the uptick.

The Unit 3 case embodies a common pattern in headcount growth – sudden surges in growth are often followed by a regression to the mean, since a sudden staff-up may be the result of a re-org or a one-time boost in headcount to accommodate a new initiative.  We therefore find it best to embed this regression-to-the-mean assumption into our baseline trend model, with the ability to tailor accordingly depending on the client. Here is what that assumption looks like in practice (Units 1 and 2 are the same, but note the change for Unit 3):

  • For Unit 3, projected headcount now follows the linear trend prior to the month 10 surge in growth, resulting in a projection that is much closer to reality

 

Two key lessons emerge from this case study:

  1. In the domain of workforce demand forecast, linear models can outperform more flexible fitting methods, since headcount growth often regresses to the mean. Projecting a linear trend based on several months of history ensures that temporary blips are not mistaken for long term patterns

  2. No matter how sophisticated one’s statistical methods, there is no substitute for subject-matter expertise. Black box forecasting methods can’t be applied without regard for the business context, and one should always consider supplementing one’s model with common sense business rules.

 

 

 

If you are interested to learn more about Vemo technology please email info@vemoworkforce.com. 

 

 

 

 

 

 

 

 

Share on Facebook
Share on Twitter
Please reload

Follow Us
Please reload

Search By Tags