TESCO IS THE UK’S LARGEST GROCER, operating over 2700 stores. As a business, we are committed to serving our customers, communities, and planet a little better every day, by offering the products that customers need while also reducing their impact on the planet. A key step in Tesco’s value chain is what happens at the end of a product’s lifecycle, when it is no longer displayed to customers. This is the last opportunity to sell an item to a customer or donate it to the community so that it doesn’t go to waste.
WHAT DO YOU DO WITH EXPIRING STOCK?
Tesco, like most large retailers, discounts items that are close to being removed from shelves. This process is applied across Tesco’s product range, from general merchandise and clothing to fresh food. In particular, food items are reduced in price as they get closer to expiry to sell them before they go to waste. The question that every retailer must answer is: By how much should the price be reduced? Reduce too little and the item simply won’t sell but reduce too much and the sale might become a net loss for the retailer.
Finding the optimal pricing strategy was the task given to Tesco’s Data Science Team. To tackle this problem, we developed a novel multi-stage Clearance Pricing Optimisation system and deployed it across all Tesco stores in the UK where it is applied to 100,000s of unique products annually. The objectives of this solution are to: (1) clear excess stock by a specific date (either expiry or new product roll-out date), (2) increase revenue by finding the optimal discounts, and (3) reduce operational costs and provide further insights of in-store processes. Our solution reduced the number of fresh food items going to waste by 5%, and increased the revenue generated by 1.5-13% across multiple food and non-food product lines.
Our solution reduced the number of fresh food items going to waste by 5%, and increased the revenue generated by 1.5-13% across multiple food and non-food product lines
THE PROBLEM AND ITS CHALLENGES
The challenge of expiring stock of both food and non-food items is something Tesco faces daily. The question we need to answer is: What clearance strategy should we use to increase revenue and, most importantly, decrease waste?
Tesco employs a product line specific, multi-phase reductions strategy to reduce waste and recover revenues from soon-to-expire stock. The time horizons involved vary depending on the business domain, for example:
Fresh food items enter a pre-defined three-stage reduction process 24-48 hours before they reach their sell-by-date. Tens of thousands of unique products are reduced each year.
Non-food items covering electronics, home & entertainment, and clothing enter an up to 4 stage reduction process, where each stage could last weeks. Tens of thousands of unique products are reduced each year.
Packaged food items with a long shelf life but highly seasonal demand (e.g. Christmas, Easter, Halloween etc. themed items) enter a pre-defined two-stage reduction process after their peak demand. Every seasonal event accounts for hundreds of unique product reductions across most stores.
Despite the differing time horizons, the high-level problem statement is the same: Calculate an optimal multi-phase pricing strategy, that at each phase further reduces the price of items as they approach their expiry date.
The main challenges associated with this problem are:
It likely requires solving two sub-problems: prediction and then optimisation, which scientifically makes it hard and requires different types of expertise.
Historic data is not always available to build robust prediction models.
Given the scale of the problem, driven by number of products and stores, the optimisation problem is computationally intensive.
The next section describes our approach for tackling this problem and the associated challenges.
THE SOLUTION
The solution that we built leveraged the same three component technique to find an optimal pricing strategy in all business domains: (1) first we model how we expect demand to react if we were to set a specific reduction (prediction), (2) we then choose optimal reductions accordingly to optimise for revenue maximisation and waste reduction (optimisation), and (3) we pass the results to the business (deployment).
Prediction
The first problem we face is building appropriate elasticity demand models, which predict how many items will sell if we set a specific price point. The biggest problem posed here is the lack of diversity of past reduction data. Tesco has typically applied static reduction strategies in the past, meaning similar reductions are applied in similar circumstances. It is therefore difficult to model exactly how customers will react to new prices that have never been used in stores before.
Model selection was crucial to getting this right – our solution has focused on the use of machine learning models that consider domain expertise combined with strong assumptions about the relationship between features and targets to deliver predictions that are useful in the context of price optimisation. This allows pricing outside the regime of reductions that were previously observed. shows an example of predicted waste and revenue curves, where the red dotted line represents waste, and the blue line – sales.
FIGURE 1 SOLUTION LIFE CYCLE
Optimisation
The second part of the solution involved building an appropriate optimisation routine to select the optimal set of reductions. The optimisation problem was to decide an appropriate set of reductions to apply across multiple phases to (a) reduce waste and (b) maximise revenue.
The optimisation problem was to decide an appropriate set of reductions to apply across multiple phases to (a) reduce waste and (b) maximise revenue
From our demand elasticity models, we will be able to predict how many items we’d sell, and thus what the appropriate pricing strategy might be. Setting the price high leads to fewer sales, and setting the price at close to zero leads to high sales but next to no revenue – the optimal price is usually between these two extremes (). Setting a lower price than the one which optimises revenue price leads to more items sold and less waste, but does this by sacrificing some revenue. The reduced to clear price optimisation is therefore always a trade-off between waste and revenue.
FIGURE 2 WASTE AND REVENUE CURVE FOR VARIOUS REDUCTION LEVELS
We use both exact and heuristic approaches for solving the optimisation part of the markdown problems arising in various business domains. The choice of the appropriate technique is driven by the business constraints, scale and complexity.
Deployment
The final problem to tackle was the deployment of the solution to the business. As was the case with the optimiser, the deployment technique has been crafted with each business use case in mind. For general merchandise, clothing, and seasonal packaged food items the problems are automatically triggered overnight. The optimiser then selects the best reduction strategies and suggests them to the merchandisers as a decision support tool. Fresh food reduction recommendations must be automatically suggested instantaneously: to deploy this solution, we pre-compute a wide range of product and trading scenarios we expect to see in store, which are then looked up by Tesco colleagues as a specific scenario occurs.
SUCCESS STORIES
The solution described in the previous section has been implemented and deployed in three business domains. Overall, we achieved improvement in both main KPIs, revenue and waste, with three observations below expanding more on the obtained results:
The largest impact of our algorithmic solution is present across fresh food items. We have reduced the number of expiring fresh food items going to waste by 5%, which prevents millions of fresh products from going to waste each year, yet also increases the revenue generated from reduced-to-clear fresh food items. By tuning the optimiser’s constraints, we were able to explicitly favour lower prices and waste reduction over revenue in this trade-off.
The longest lasting solution delivered by our team is a human-in-the-loop decision support system for clothing and general merchandise, that in addition to optimising the reduction strategy for each line, provides human merchandisers with accurate forecasts of future revenue and stock. By investing effort to continuous scientific development and leveraging the feedback from live tests, we demonstrated the results that consistently outperform manual decisions making in terms of revenue.
Last but not least, we increased the revenue generated from reduced-to-clear seasonal food (e.g. Easter eggs, chocolate Santa) without impacting waste.
We have reduced the number of expiring fresh food items going to waste by 5%, which prevents millions of fresh products from going to waste each year
SUMMARY
We were delighted to be awarded the OR Society’s President’s Medal in 2022 for this work.
Finding an optimal reduction strategy is a problem faced by every retail business. There are two conflicting objectives to not only increase revenue but also reduce waste and finding a solution that achieves both is a non-trivial task. At Tesco we have built and deployed a predict-optimise solution across multiple product areas that successfully reduced waste and our impact on the planet, and also increased revenue for the business.
Additional information
Notes on contributors
Aleksandar Kolev
Aleksandar Kolev is a Senior Data Scientist at Tesco. He has been working on the two projects related to food price reductions. Aleksandar, as a keen statistician and researcher, helped the team improve not only the scientific solution but also identify key research areas that enabled us build even better algorithms.
Ross Hart
Ross Hart is a Data Science Manager at Tesco. He has a PhD in astrophysics, and now uses machine learning and optimisation techniques to solve both online and in-store Data Science problems.
Ekaterina Arafailova
Ekaterina Arafailova is a Lead Data Scientist at Tesco. Her background is in operational research, and at Tesco she focusses on optimisation problems arising in various business domains. Using operational research approaches, Ekaterina and her team have already delivered a significant value across the business, with more to come.