Using Data Science to Optimize Inventory in Retail

Give customers what they want, keep storage costs under control

Published: Tuesday, May 28, 2019 - 11:03

Do you know what a retailer and a tightrope walker have in common? They both have to balance. For the tightrope walker, the logic is clear. But what’s the balance that a retailer is looking for?

A typical dilemma of shortages vs. storage costs

Although the dilemma of shortages vs. storage costs is applicable to any product category, it’s much more painful with perishables. If their quantity can’t meet the demand, retailers should be ready to see a frown from an unhappy customer who didn’t find her favorite dairy, fruit, or vegetable on the shelves.

However, staying on the safe side by ordering more perishables is hardly a cost-effective solution. Perishable products require special storage conditions, and their shelf life seldom exceeds a couple of days, which means retailers must address disposal issues. So, it’s easy to understand why retailers, by all means possible, try to find the optimal balance between storing too much and too little.

A way of handling this dilemma with data science

There is a way to handle the storage/shortage dilemma efficiently: It’s via a deep neural network (a DNN), the most advanced data science approach.

Here, I’ll get straight to the point and describe how a DNN works. If you are completely new to the concept, I recommend that you check out DNN fundamental terms and architecture first.

When just created, a DNN is ignorant. It gets its intelligence in the course of learning, just like a student. If we draw the parallel further, the teachers and course books are also there. Only, in DNNs’ language, they are called data scientists and training data sets. For calculating inventory, the latter could be your sales data—for example, daily sales per SKU for the last two years, including promotion support, day of the week, and other details.

At first, the neural network looks at historical sales data, applies random weights to them, and produces a test prediction, which is supposed to be the optimal stock level. Given that it is based on historical data, we already know what the output has to be. And, naturally, we are interested to know how much the network’s test prediction differs from reality. There’s a special way to do this called a loss function. It calculates the network’s error and “penalizes” it for mistakes. Based on these penalties, a DNN adjusts the weights and tries to minimize the error for the next prediction. This happens repeatedly until the network finds the ideal set of weights and manages to make accurate predictions.

A loss function has another important job to do: It follows the above-mentioned principle to weigh storage costs against shortage costs for a certain SKU, and lets a DNN produce inventory-level predictions with the optimal balance factored in.

What do you need to make a DNN work?

Data
If course books were full of errors or provided students with insufficient information, we wouldn’t expect them to become pros in any subject, right? The same is applicable to DNNs that are dependent on the amount and quality of data in the training data set.

To produce precise inventory forecasts, the neural network needs at least detailed sales history. However, in the case of calculating an optimum supply of perishables, you can also consider the shelf life for a certain SKU, disposals history, and supply-chain constraints such as lead times as well as minimum and maximum order quantities.

Factors
A DNN is an obedient student but it lacks initiative. If you instruct it to consider promotion influences and seasonality as factors that influence your inventory, it will do exactly that. But if weather conditions have a dramatic impact on your inventory level, a DNN won’t look at them unless you tell it to do so. Thus, you need to come up with an extensive list of factors that a DNN must take in. Here are several examples of possible factors: perishable marker, promotion support, holidays, day of the week, weather conditions, and supplier category.

Data scientists
Data scientists are the “teachers” of your DNN “students.” However, their role is not limited to applying a loss function to give marks; they also create the whole learning course for a DNN. Apart from that, data scientists are the main contributors to the list of inventory-influencing factors. As discussed above, the process of defining key factors should ideally go off without a hitch. For that to happen, data scientists usually perform a statistical analysis of available data and clarify all the meaningful details of inventory prediction during the meetings with subject-matter experts.

Examples of using data science in practice

We picked three inspiring use cases of how retailers optimized their inventory as well as achieved stunning results in closely related initiatives, such as ordering and replenishment.

As companies are mostly laconic when it comes to describing their data science initiatives, we’ll have to settle for the “we applied machine learning” degree of detail. However, even without knowing what stands behind this wording—a DNN or some other data science technique—it’s still useful to take a look at the results of these smart endeavors because they are real-life proof of data science capabilities.

Example No. 1: Kaufland, a German grocer, applied machine learning and automated the replenishment process for its fresh meat division, improved product availability, optimized stock levels, and reduced write-offs.

Example No. 2: The second-largest e-commerce retailer, OTTO, has also benefited from machine learning algorithms that enabled it to reduce its inventory levels by 20 percent without compromising on service levels.

Example No. 3: Morrisons, the fourth-largest supermarket chain in the United Kingdom, also used machine learning to optimize its replenishment and automated the ordering process for 26,000 SKUs in all its stores.

The article ends here, but hopefully, your interest doesn’t

We’ve considered here deep neural networks, the most recent data science advancement. Their design allows for accurately predicting optimal inventory levels. In addition to being able to find and factor in a much-needed balance between shortages and storage costs, DNNs can consider other factors that influence stock levels.

As we see from real-life examples, retailers already use data science to optimize their inventory, both perishable and nonperishable. And although we focused on the retail industry, manufacturers and distributors that strive to find optimal stock levels will find data science of interest, too.

About The Author

Alex Bekker

The author is Alex Bekker. The Head of Data Analytics Department at ScienceSoft, an IT consulting and software development company headquartered in McKinney, Texas. Combining 20+ years of expertise in delivering data analytics solutions with 10+ years in project management, Alex has been leading both business intelligence and big data projects, as well as helping companies embrace the advantages that data science and machine learning can bring. Among his largest projects are: big data analytics revealing media consumption patterns in 10+ countries, private labels product analysis for 18,500+ manufacturers, BI for 200 healthcare centers.