Demand Segmentation - Simplified

For supply chain professionals demand segmentation is a very familiar word . If you have worked in any organization as demand planner then you might have either implemented or adopted some variant of this standard statistical approach of segmentation. At the core it lies our need to improve demand planning accuracy ( say metrics ) by focusing limited resources where it matters most . Let us pretend that you have thousands of SKUs in your portfolio and you have been tasked by your manager to find a better strategy to do demand planning . One approach ( Like 6 Months moving average *) may not be best way to model your data . May be there exists some SKUs which are better left to stat models releasing lots of time for planning team to focus on other activities . In this article, objective is to explore the idea of doing demand segmentation using common sense . You don't need to remember the jargons and you can yourself do it if required. Is it really possible ?

Demand is a function of two variables . Quantity which is observed . Depending on the business process applicable in your case , you may measure it in different units like Boxes , Dozens , Each , Meter, Liter etc. It cannot be negative however potentially take any value between zero and a positive number . The second variable is Time . Many organizations measure demand for SKUs every day ( daily sales ) , some does it every month ( Monthly ) and so on. There does exist plethora of flavors of this variable . The caveat here is that interval between measures remains consistent . It never happens in operations that today i am using daily sales to do demand planning and next day we switch to alternate day . This variable can have fixed intervals or dynamic measures but we will limit ourself to fixed interval time series . Within this scenarios we may not report demand for every instance . If we are measuring monthly demand then it may be possible that we witnessed no demand during any number months for a given year . There is no guarantee that every month will have demand as a positive number .

Demand

Variation in demand can be measured statistically . However let us see some patterns keeping other variable time fixed (demand will be observed across the time span ) before we go there .

We can statistically interpret above mentioned TS using two common measures . Mean and the coefficient of variation (CV). Mean does not require introduction however let us recap CV .

The coefficient of variation (CV) is defined as the ratio of the standard deviation σ to the mean μ

CV = STD /MEAN

It shows the extent of variability wrt the mean of the population. For example let us look at mean and CV of above plotted time series .

SKU ABC_1 which has inconsistent demand pattern has higher CV vs ABC_2 . Hence CV can be used as a measure of variability . In some literatures you will notice CV square being utilized . The most important question here is what should be the value of CV which filters low variability and high variability demand patterns? If STD is half of Mean then we get CV as 0.5 . 0.5 would be our boundary line . Yes, you may adopt lower values like 0.4 or even 0.3 but then remember higher values drive us towards lower population of SKUs where we can use stat forecast ( We will come back to it soon ) .

Hence any SKU with CV higher than 0.5 , we will put it in high ( H) variability Zone where as rest should be parked under Low ( L) Variability zone .

Time

It is frequency of observation of demand ( we keep demand value fixed ) . See below some of patterns -

We will introduce a term called OFD ( Observed frequency of Demand ) . Feel free to give it any other name as per your convenience . What it is-

= 12 ( Assuming that you have last 12 month of Demand data ) / Number of months for which non-zero demand was observed

Since intermittent TS ( see above chart ) has 6 demand points hence OFD = (12/6) =2 however for Inconsistent TS OFD comes out to be 1.7 . A demand planner would always like to have lower value of OFD!

Like Demand , we will treat OFD value 1.34 as boundary line ( 12/9) . You can define your cut off value .

Combining it all together

Real life TS is function of both D & T hence both varies over observation horizon making it quite complex . For example -

D is changing over time as we don't observe same value and T is inconsistent as Demand does not exist for all periods (Like no demand in April and May ) . Fortunately not all TS has this pattern.

Segmentation Zones -

When we apply scatter plot to TS data ( CV and OFD) , we get matrix like above ( Note that it depends . You may find that entire portfolio is sitting in Zone-3 or  your TS data is distributed between ZONE-1 and 2  ETC . You should not get disappointed if we don't see SKUs mapped to all four zones ! )        

Segregating data into different zones is the easiest part . Expertise and experience drive the action plan and it comes with premium . We briefly talk about ZONE-1

For TS , which is part of Zone-1 ,there is very high probability that stat forecast algorithm will generate better forecast ( vs FVA) provided planner selected right model . For a SKU falling under Zone-1 if FVA MAPE is better than Stat MAPE then as a planner you should review the model that you opted to generate forecast . It is an indication that DP process are not completely working . In this Zone Algorithm will almost always prevail over Human! Zone-1 is better left to statistical models ( conditions apply ) . For example see the below TS -

You observed year on year linearly increasing trend and you opted for linear regression as it is expected to do a great job in such cases . How many times we take out time to plot errors ( Forecast - Demand ) and validate if it is a white noise ? If residuals are not white noise then you should not use Linear regression on that TS rather look at other options .

What should be strategy for other Zones ?