Can Demand Segmentation Improve Your Statistical Forecast?

Imagine a demand planner working with 10,000 unique combinations. One of the not so envious tasks for this person would be to generate a statistical forecast for all these combinations. These days, the statistical forecasting tools available on the market can forecast these combinations using a list of forecasting methods and figure out which method works best for a particular combination.

Demand planners often ask the following question:

Can demand segmentation be used to improve the quality of this statistical forecast even in the presence of a very powerful engine?

The answer is a resounding yes.

[Read More: 6-Step Guide to Improving Your Statistical Forecast Technique]

Understanding Statistical Forecasting Engine Methods

First, let’s look at how statistical forecast engines pick the best method. A few statistical forecasting software packages still pick the best method based on overall fit over the historical data. Most, however, use a technique of holding out some data from the statistical engine and evaluating the accuracy of the forecast over those periods. Increasingly, the second approach is the preferred approach.

While this is a great approach and enables a large degree of automation, it is not risk-free. For example, if you throw bad methods into the mix, even occasionally, it will get picked because of the underlying data conditions. Furthermore, because these methods usually put a higher weight on the more recent observations, a swing one way or the other can impact which method gets selected. And last, but not the least, there is a problem of the consistency. For the same series, the forecast method can change over two successive runs with minor data changes.

Combining Demand Segmentation with Statistical Forecasting Engine Results

Combining the automatic selection power of the forecasting engine with an approach to demand segmentation is a better approach.

Examples of how to Use Demand Data Segmentation to Improve Statistical Forecasting

Example 1: Using existing data segments
An easy thing to do is to look for segmentation already available in your data. For example, one might want to forecast one family with an input list of five methods, and another with an input list of 10 methods. Perhaps one of the product family is very seasonal and should be run predominantly with methods that account for seasonality. A certain product family might be affected by external factors such as housing starts. Similar thought process can be extended to different geographical regions.

Example 2: Generate segments based on data analysis
More interesting is the idea of generating segments based on analysis of your data. For example:

Analyze upfront the data series that simply do not have enough data points to generate any kind of forecast. It is best to channel these towards no forecast.

Example 3: Product Lifecycle Analysis
An analysis to figure out the stage of your product lifecycle can be used as part of your data segmentation analysis. This can lead to segments such as New, Active, and end of life. New combinations would get a different set of methods, compared to the ones marked end of life. Actively selling combinations might get their own set.

[Read More: 8 Common Issues to Avoid When Creating a Customer Segmentation Strategy]
 

Example 4: Seasonality Analysis
Another analysis might be to look for combinations clearly showing seasonal traits. Once identified, this segment of the data might be forecasted with a mix of methods that look at seasonality.

Example 5: Intermittence Analysis
One of my favorite analysis is to look for intermittence, which essentially measures a ratio of non-zero observations and zero observations. This then allows one to use specialized methods designed for sporadic or sparse data for this segment. In Arkieva, we deploy several ways of measuring this very important quality.

Example 6: Volume and Variability Analysis
Last, but not the least, one might do a 9-cell analysis on volume and variability. Once done, one should take a very pragmatic approach in assigning methods to different segments. For example, High Volume, Low variability combinations should be forecasted using all kinds of methods. By contrast, Low volume, high variability methods should probably be forecasted using Croston’s or average method.

At Arkieva, we routinely apply these techniques with our customers. Combining data segmentation analysis with statistical forecasting methods often leads to forecast accuracy as high as 11%.

[Read More: 5 Customer Product Segmentation Pitfalls to Avoid]
 

Enjoyed this post? Subscribe or follow Arkieva on Linkedin, Twitter, and Facebook for blog updates.