Interpretable Clustering for Demand Forecasting

Identifying similar cohorts to predict product demand more accurately

Demand forecasts are key in the stategic planning of any business

A Fortune 100 company came to us because their current demand forecast model was highly non-interpretable. They were unable to thoroughly explain the final predictions, and hence could not convince the business stakeholders to use their outputs.

Additionally, their data scientists struggled to improve the model because they couldn't identify where the models systematically underperformed or outperformed. This put their team at risk of committing the same mistakes repeatedly.

Interpretable Clustering identifies cohorts with similar characteristics

To address these challenges, we took a holistic approach where we built interpretable clusters of purchase orders based on factors relating to the typical order quantities for different products and customers, as these are factors that have a major impact on supply chain management. Specifically, we trained an Optimal Classification Tree to predict the size of a purchase order: small or large (larger than three times the standard deviation from the mean).

We define a cluster as the collection of observations falling into a leaf of the classification model, as by definition they share similar characteristics (the features used as splits by the tree). Then we train Optimal Regression Trees to generate interpretable demand forecasts for each cluster.

Detect areas of weakness to drive improvement

Compared to a single regression model, this cluster-and-regress approach improved the predictive performance by 5%. In addition, we generated error metrics to benchmark the performance in each cluster against that on the whole dataset to identify areas of weakness or strength.

As an illustrative example, the tree below shows that the predictions were particularly problematic for products with higher than average demand in the past quarter and longer than average lead time. This allows a tailored and direct process of improvement, such as collecting additional data or engineering new features that are better-suited for making predictions in this cluster.

This process of targeted intervention worked well for this Fortune 100 company and enabled rapid and efficient improvements in the model quality.

Unique Advantage

Why is the Interpretable AI solution unique?

  • End-to-End Interpretability

    Full transparency in each step of the prediction pipeline increases data scientist efficiency

  • Better Communication

    Enable business stakeholders to check prediction logic and verify intuition

  • Enhanced Performance

    Training regressions in each cluster allows more flexible and stronger models

  • Useful Insights

    Automatically highlight understandable areas that need improvement

Want to try Interpretable AI software?
We provide free academic licenses and evaluation licenses for commercial use.
We also offer consulting services to develop interpretable solutions to your key problems.

© 2020 Interpretable AI, LLC. All rights reserved.