Using ML to Optimize Slow Movers in the Retail Space

Yes, the title of this post is a bit of a mouthful, but inventory management optimization is something our machine learning team has been working on. We think it is an underdeveloped area in the world of high-end retail analytics.

The Problem

The aim of our latest foray into this space was to predict slow movers, these being products that remain in inventory for a long time. Of course, the definition of how long this actually is will change depending on the particular retailer. Some measure slow movers in terms of inventory turnover, which tracks how often inventory is sold and replaced in a given period. Slow movers have a low inventory turnover ratio, which can be due to a number of factors including overstocking, a change in customer behavior, unanticipated changes in demand due to seasonality, or poor marketing performance.

Accurately forecasting demand for slow-moving products is crucial to optimizing inventory levels and minimizing wastage.

Our Approach

We began with short-term forecasting, believing we could build on what we learnt when developing this model to then apply it to longer-term forecasts. However, we found that it is not so simple, as the forecasts don’t appear to scale out particularly well from, say, 1 to 3 to 6 months. We found that this appeared to be partially due to the lower granularity of the data we used for the shorter-range forecasts, as well as the focus on more recent trends not being directly applicable to the longer-range time horizons.

Not being able to base our model on the best quality historical data was also somewhat responsible for this. Whilst we did have some historical data, the COVID-19 pandemic diminished the usability of the data from that period. Applying appropriate weightings can go some way to fix this, though nothing beats clean, untouched primary data.

A single model is therefore not the catch-all solution for the slow movers problem and a series of similar models that each use slightly different features is required to cater for each time horizon forecast.

Generalizing the Approach

In fact, forecasting slow movers is pretty similar to other inventory optimization tasks, where the goal is to maximize profits by ensuring that the right products are available in the right quantities at the right time. The main difference is that slow-moving products require a more nuanced approach, as demand for these products may be unpredictable or irregular, and excess inventory can result in significant costs.

Just as we need to apply more than one model to the slow mover scenario, it is quite common to apply multiple machine learning models to solve other optimization problems. This would usually be done to predict particular metrics for particular time ranges.

Examples include stockout rate prediction which predicts the rate that a retailer will run out of a particular product or SKU, gross margin return on investment (GMROI) prediction, which is a measure of inventory profitability, and lead time forecasting, which is a prediction of the time taken to replenish inventory once it is sold out.

It is important that retailers can get a handle on their constantly changing stock levels and manage them to effectively maximize customer satisfaction and profitability. Machine learning is an important piece of a complex puzzle that can assist in optimizing inventory management when done correctly.

If you would like to improve your stock management processes, or you have another machine learning challenge, please feel free to reach out to continue the conversation.