Forrester readers sometimes ask me if machine learning can really have any relevance to retail.

We start by discussing their current approach to demand management. How do they determine the assortment and depth and breadth for each store cluster, or even for each store?

Depending On Sales History Alone To Plan Assortment Depth And Breadth Is Risky 

For replenishment items in grocery, convenience, or hard goods, it usually depends on history. Retailers generally use a seasonally adjusted weighted moving average of SKU sales at a regional or distribution-center level. When forecast error increases, the model suggests a change in the smoothing coefficient. But the basic idea is that you can forecast future demand based exclusively on demand history. Retailers also drive their assortment or variety based on sales history, too. Yet the earnings calls for most retailers describe a range of external events (often weather-related) contributing to a shortfall in revenue or margin.

Machine Learning Provides More Resilient Forecasting By Taking Into Account More Variables

The value of machine learning is to exploit elastic storage and computing power to analyze a range of independent variables and incorporate them in more robust predictive models. We recently found that retailers and CPG brands collect data such as clickstream or social data that could help them meet local demand more effectively. But their current applications lack the capability to drive insight from the data and hyperlocalize assortments. This limits their ability to capitalize on store proximity to customers as a tactic to outpace pure-play online competitors.