15 Machine learning

Tell us the things that are to come, so that we may know that you are gods.

—Isaiah 41:22

In this chapter, you will learn what machine learning is, how it works, and what you can expect when using it to forecast demand. We will also discuss pitfalls and best practices when launching an ML initiative.

If you want to learn how to create your own machine learning models, feel free to check my book, Data Science for Supply Chain Forecasting.

15.1 What is machine learning?

So far, we have discussed statistical models using predefined mathematical relationships to populate demand forecasts. The issue was that these models cannot adapt to demand patterns. For example, if you use a statistical model that doesn’t include any seasonal factor to forecast demand for a seasonal product, it will fail to interpret the cyclical patterns. And it will likely interpret them falsely as changing trends. On the other hand, if you use a seasonal model to predict demand for a non-seasonal product, it will overfit historical random variations by interpreting them as a recurring seasonality.

Machine learning is different.

With the technological advancements of machine learning algorithms, we have new tools at our disposal that can achieve outstanding performance on typical supply chain demand datasets. These new models can learn complex relationships using historical demand and demand drivers to predict future demand.

15.1.1 How does the machine learn?

15.1.2 Black boxes versus white boxes

15.2 Main types of learning algorithms

15.2.1 Short history of machine-learning models

15.2.2 Tree-based models

15.2.3 Neural networks

15.3 What should you expect from ML-driven demand forecasting?

15.3.1 Forecasting competitions

15.3.2 Improving the baseline

15.4 How to launch a machine-learning initiative

Summary