AI-driven efficiency: How Target’s data science models optimise demand and inventory for 2,000 stores

In today’s competitive retail landscape, precision in demand forecasting and inventory management is a cornerstone of success. At Target, data science drives this precision by leveraging AI and machine learning to optimise operations across its vast retail network. Through fully automated, integrated systems for forecasting, purchasing, and product positioning, Target has significantly reduced manual intervention, enhancing operational efficiency and ensuring seamless product availability for consumers, says Sharad Limaye, Senior Director, Data Sciences, Target. In this editorial interaction, Sharad shares details into how these systems have been implemented, the collaborative strategies that drive their success, and the emerging innovations in AI and data science that are poised to redefine inventory management and demand forecasting for the retail sector

Some edited excerpts from the interview:

How are AI and machine learning models improving the accuracy of demand forecasts?
Given the size and scale of our operations AI and ML models are absolutely critical in improving the accuracy of demand forecasts. Target sells thousands of products in its stores and online via target.com. Generating accurate demand forecasts for millions of item-location combinations requires algorithms that are fast, scalable, explainable and flexible. Target uses Generalised Additive Mixed Models (GAMM) for generating highly reliable demand forecasts every single day across various selling channels and across the entire network for short-term execution as well as long-range planning.

ML models have helped significantly reduce the forecasting error and improve accuracy compared to our legacy systems and processes, across all our selling channels

What is the role of data science in optimising inventory management, from warehouse stock levels to shelf placement in stores, ensuring a smooth flow of products to meet consumer demand?
Data sciences is at the front and centre of decision making when it comes to designing and developing solutions for all areas of retail which includes inventory management. Some of the ways in which Data Science is being used:
Data science models are fully embedded in decision making for optimising inventory flow across the entire supply chain lifecycle. These models generate optimal policies that help decide how much inventory to buy to be placed in the warehouses and how much to be replenished in stores. Having the right balance here is very critical so that you don’t buy too much (as storage space is limited) or too little inventory (as it could go out of stock).

Within the stores,  space and planning models help better plan and position items in the stores  considering the size of the stores, category affinities, expected sales and other operational constraints.  Allocating optimal shelf space to all categories is very important to drive sales for Target.

Data science models help make smarter decisions both upstream and downstream ensuring the seamless movement of inventory across the supply chain and in our stores, thereby improving operational efficiency

How has Target implemented fully automated systems for forecasting, purchasing, and positioning, reducing manual intervention and enhancing operational efficiency across the supply chain?

We have built algorithmic solutions and data science models to power a fully automated and integrated forecasting, purchasing and positioning system, that are setting the foundations for modern inventory management at Target.
It all starts with understanding customer demand considering multiple factors around promotions, events, holidays etc. Combining all these factors, ML and deep learning models help us get an accurate estimate of unconstrained demand across our network of 2,000 stores. This feeds into replenishment and purchasing systems wherein different algorithms look at making the best possible decisions under uncertainty. These are complex multi-echelon inventory optimisation problems where decisions have to be made for millions of item-location combinations every single day.
The integrations that we have built across our production systems, give us end-to-end visibility on various decisions and events and the flexibility of adding more capabilities and enhancements as we modernise our supply chain.  As this has significantly improved explainability of models, it has also helped reduce manual interventions over time which otherwise was very difficult with our old legacy systems.

Can you touch upon a little bit regarding the need for cross functional collaborative efforts in developing and scaling these AI-driven solutions?
Forecasting and inventory management systems are the bread and butter of any retail organisation and it requires integration with complex systems that have to work with clockwork precision and are completely synchronised. To make this happen, data sciences works very closely with our partner teams in engineering, product and business to build and scale these models. Bringing in the right process knowledge and experience as well as knowing what capabilities to build with the right level of engineering and science is very critical. The teams work very closely and collaboratively in an iterative way to build these systems. Building large-scale production grade systems driven by AI requires such joint planning and execution,

What are the upcoming innovations in AI and data science that could further enhance inventory management and demand forecasting in the retail sector?
As supply chains become more complex than ever, the speed and accuracy of decision making needs to evolve as well. Simulation-driven optimisation,  Markov models, bandit algorithms and reinforcement learning are all areas of research that will enhance complex and sequential decision making on problems which are stochastic in nature and algorithms that have to deal with lot of variability.

Leveraging additional signals from operations to enhance intelligence, making algorithms more explainable, faster experimentation coupled with right platforms that operate at speed and scale will all be areas to look forward to, to enhance inventory management and demand forecasting.