Machine Learning Powered Intelligent Replenishment in Retail

by Kumar Singh, Research Director, Automation & Analytics, Supply Chain Management, SAPinsider

 

The criticality of establishing an efficient store replenishment process

The process efficiency of replenishing store inventory is critical to a retailer’s overall operating efficiency and even profitability. It is not news to anyone that store replenishment impacts the on-the-shelf availability. In today’s competititive retail landscape, having your shelves running out of products can have wider consequences than the immediate loss of sales. The one sentence summary is – every time a customer walks in to buy a specific product and can not find it on the shelf, your brand as a retailer takes a hit. This has become much more critical in the ecommerce era, where a customer may eventually decide to stop visiting your location and shop entirely online, where a product is always available. As we have seen recently, even fresh produce and groceries are now under the ecommerce net so a stock out will be akin to providing your customers one aditional excuse to not visit your stores.

Key challenges of retail replenishment

While many challenges can be listed, the most critical challenge comes from the granularity of forecasting and planning for tens and thousands of SKUs (for a large retailer), and almost each of these are stocked at hundreds or thousands of stores. These numbers make item level and store level forecasting and replenishment planning extremely challenging. The challenges with having all the key data points required, the robustness of the model, the internal skill set required to design planning process and to manage it, among many others, are secondary.

Leveraging Intelligent replenishment

Considering the complexities of the process, there is no doubt that an AI based intelligent replenishment planning solution is the answer. The solution of these challenges can not be achieved by vanilla automation or conventional mathematicl heuristics. A “learning” aspect has to be involved and hence AI and ML based solutions are the only answer. In my opinion, there are multiple ways to design an effective solutions. Such intelligent replenishment strategies not only reduce cost and waste , they also help enhance customer experience in tandem. If properly designed and developed, these highly automated solution can help take the guesswork out of retail replenishment. The illustration below shows how many data points are actually involved in store replenishment process.

What can be the high level flow of such a solution

The flow described here is at a high level, in order to simplify the architecture, but here is the most important aspect:

The USPs of this solution are two Machine Learning algorithm that drives the process by generating the replenishment quantities. A predictive algorithm and a reinforcement algorithm. You may know all the input data points but your selection, design, training, testing and validation of these algorithms will eventually determine how effectively the solution works.

I. Starts with a tactical and dynamic forecast algorithm

A Machine Learning algorithm that generates detailed forecast tapping multiple data points and attributes like the ones mentioned below. Remember that replenishment forecasts, unlike some other demand planning forecasts are more dynamic and tactical. As you can imagine, many retailers replenish their stores multiple times a week and hence you need to select and design an algorithm that aligns with this specific nuance.

  • Point of Sales (POS) data
  • Seasonality
  • Holidays
  • Price
  • Promotions
  • Weather
  • Shelf life & spoilage

II. Reinforcement Learning then comes into play

The solution will then leverage the forecast to create detailed replenishment and order quantities. This can be drilled down at SKU level, filtered by business exceptions and allows you to make adjustments when necessary. For calculating these quantities, the solution looks at multiple aspects to find the lowest expected costs. It also allows you to define your own priorities in the configuration and the tool then not only calculates optimal store orders for you but also shows you how the numbers were calculated (which is critical to build confidence in the tool). This is where you leverage a customized reinforcement learning algorithm. If you really have the technical expertise, you can leverage a deep learning algorithm as well to replace reinforcement learning, but the training requirements will be exponentially bigger. In experimental data, both these approaches performed similar but the reason behind that was that the experimental data set was small and hence true benefits of deep learning could not be exploited.

What does this mean for SAPinsiders ?

  • SAP’s ecosystem has the capability to create innovative solutions to address conventional challenges retailers face. For an example of such a solution, please refer this link: https://news.sap.com/2021/01/sap-add-on-replenishment-planning-retailers/
  • To design innovative solutions like these, a best-in-class combination of people, processes and technology is required. And the key first step is to understand every minute detail of your existing replenishment process.
  • There are external partners that can help you design these solutions
  • Emerging technologies (like IoT) need to be leveraged to make these solutions more robust.

 

Kumar Singh, Research Director, Automation & Analytics, SAPinsider, can be reached at kumar.singh@wispubs.com