Supply Chain 101: Portfolio Segmentation for Supply Chain Planning
A big shout-out to Felipe Matheson and Nick Munoz for their feedback on this article.
Prioritizing well is important. As beings with (in)finite time, money and will, it helps us accomplish our tasks and meet our goals- whether we’re dining at a lunch buffet, running errands over the weekend or managing deliverables at the workplace. We may do some tasks with crazy focus and urgency. On the contrary, we may choose to do certain tasks later. Additionally, certain jobs are best done individually while others are best done in teams or completely delegated.
The same concept applies to forecasting customer demand. We look at what lies ahead and what has happened in the past.
While looking at the past, if your historical demand data tells you that the reward of good forecast accuracy is asymmetric across different product-customer groups, would you still allocate the same effort to everything? Using a highly collaborative, touch-intensive (or conversely, highly automated) forecasting process may not yield the best results.
With the demand data available, it is possible to create a starting set of tailored actions for different product-customer groups. This is known as demand segmentation.
Broadly, it classifies demand behavior based on two main factors:
- Monetary Importance; based on Cost, Revenue or Margin Contribution
- Forecastability; based on Intermittency and Demand Volatility
For purposes of illustration, here is a sample prioritization matrix with guiding actions. Some things to consider:
- Certain product-customer groups have stable demand patterns and low monetary importance. Can planning organizations free up some time by automating forecasts with an analytical forecasting engine?
- Conversely, planning organizations will also encounter product-customer groups that have unpredictable demand patterns and high monetary importance. Sales, Marketing, Finance and the customer need to work very hard to provide a good forecast.
- The other two buckets fall somewhere in between the first two buckets, involving varying degrees of reliance on analytical forecast methods and manual forecasting effort.
There are three analyses that can be run on demand data to capture these elements:
- For Monetary Importance: Pareto “ABC” Analysis
- For Forecastability: Intermittency “RIS” Analysis, Variability “CoV” Analysis
Pareto Analysis (ABC)
This analysis sizes up the importance of a product-location in relation to overall costs, revenues or margins. Naturally, planning organizations would want to make sure they are paying special attention to planning combinations that have more monetary importance.
How to do a Pareto Analysis
Consider an example wherein products (P1, P2) are sold in locations (X, Y, Z), and their overall revenue contribution is known over a 36-month period:
Arrange all the product-locations in descending order of revenue contribution. Calculate the % revenue contribution of each product-location using the below formula:
Once the % revenue contribution is calculated for each product-location, the Cumulative % Contribution is calculated, as shown in the table below. There’s no set rule but organizations typically set cut-offs using the “80–20” rule.
In a real-world situation, it is likely you will encounter many more products, and customers. A few product-customer combinations may end up contributing to the top 80% of your metric. Conversely, several other combinations would fall into the bottom 5%. However, the distribution of combinations could depend on the actual long-tail pattern and the cut-offs chosen by the planning organization.
Additional Notes on ABC Analysis
a. Setting Cut-Offs: There is no need to strictly stick to the “80–20” rule. Additionally, one could use more than just three classes (e.g., D, E) as well. But, be mindful of how that might impact your ability to decide and classify, especially if your demand shows a pronounced long-tail
b. Product Criticality Factors: Certain products may fall under the “C” class but may be critical to the customer to use more important products
c. Quantity-based Analysis: When addressing questions on demand forecasting, avoid performing your ABC analysis based on quantity. It will not provide an accurate picture, especially when products within a portfolio may be priced differently and do not have the same use
Intermittency Analysis (RIS)
It is likely that analytical models will perform better with more data points. The intermittency analysis looks at the number of periods of non-zero demand for each product-location combination.
How to do an RIS Analysis
Calculate the number of periods of non-zero demand for each product-location. Then, based on the RIS cut-offs set by the planning team, classify the demand for different product-locations.
For example, a company may decide to set the following ranges when categorizing monthly demand over a 36-month sales history period.
Additional Notes on RIS Analysis
a. Understand your Models: In this analysis, it helps to know how your models are likely to respond as the size of your data increases, e.g., if you’re forecasting demand in monthly buckets, having 24+ points of data would allow models to pick up yearly seasonality
b. Treating New Product Introductions: If a new product-location is introduced and enough history is unavailable, it might a good idea to call that out to define a different course of action, e.g., a product-location is “new” if it has not sold for more than 12 months
Variability Analysis (CoV)
The Co-efficient of Variability (CoV) analysis can be used to get some idea of the volatility of demand for different product-location combinations.
How to do a CoV Analysis
The CoV tells us the approximate variation of demand from the average for a given period. The lower the value, the more stable the demand pattern. The measure is unit-less and useful, especially when comparing different time series data sets. It is calculated using the following formula:
Like the other two analyses, the variability analysis also uses cut-offs. Here’s a set of standard cut-off rules when considering time series data:
Additional Notes on CoV Analysis
a. Setting Cut-Offs: Forecasting can be done in daily, weekly, monthly or quarterly buckets. The cut-offs used need to be tweaked, based on the buckets being used. Generally, as the data gets more granular, it tends to look more volatile. As a result, CoV tends to get higher as well.
b. Handling Highly Seasonal Demand: Highly seasonal demand generally returns slightly higher CoV values. Most analytical models also tend generate good forecasts on highly seasonal data. However, it is also possible to find demand that is legitimately volatile, and not easily forecastable for the same CoV value. As a result, it may be useful to also run tests for trend and seasonality on your data.
c. CoV Inflation and Active Selling Period: Only calculate your CoV within the active selling period. For example, if you’re pulling 36 months of data, it is likely that certain product-customer groups will have demand over all 36 months. You will also find some product-customer groups have started showing demand only in the most recent 5–6 months. Zeroes in your demand data may not always represent the same thing and being aware of the active selling period can prevent the CoV value from being inflated.
Insights
Once this segmentation output is generated, it can be used to drive certain important decisions in demand and inventory planning:
a. Forecasting Granularity: Demand segmentation analysis can be used to assess the quality of data at different levels of data granularity. Be mindful of how your forecast tree is setup, e.g., Does it make sense to forecast at the SKU-Zip Code level if demand signal is volatile and inventory is not positioned in individual zip codes? Would you be better off forecasting closer to an upstream node if manufacturing/storage are centralized if the demand signal looks more stable and planning effort is reduced?
b. Exceptions Management: Exception rules can help planning organizations smartly allocate effort, and single out major variances between different forecasts, e.g., For product-customer groups with moderate forecastability, flag all cases wherein 1-step monthly Sales and statistical forecast differ by more than 20%
c. Stocking Decisions: Demand segmentation can also be used to determine stocking strategy for different products, e.g., important products with moderate forecastability could have cycle and safety stock
Conclusion
Once the ABC, RIS and CoV analysis have been done, product-locations can be segmented into specific groups, as seen in the figure above. As planning organizations do this analysis, it is helpful to keep in mind the following:
a. Strive for simplicity; strike a balance between creating too few and too many categories
b. If useful, consider adding additional flags to capture product complexity, lead times and product life-cycle stage
c. Run this analysis periodically (e.g., quarterly); this would tell planning organizations if they need to reprioritize from time to time
d. Pay special attention to your data; normalize data for product replacements and pay special attention to outlier events (e.g., one-time customer projects, promotion spikes, economic downturns), and how certain flawed demand proxies (e.g., sales, shipments) may affect your output. Use the same data that you would feed into a time series forecasting engine.
e. The past can only tell you so much; do not lose sight of forward market intelligence while forecasting
A big shout-out to Felipe Matheson and Nick Munoz for their feedback on this article.
