As Accenture notes, the retail industry is quickly reaching the peak of the AI hype curve, as 86% of retailers are experimenting with it to forge new paths to growth. We are now seeing the boldest of retailers tackle a new frontier when it comes to AI adoption: data-scarce demand forecasting.
Forecasting in situations with rich historical data continues to be a top-performing use case for AI in retail, as we’ve seen stronger adoption in the grocery and fast-moving consumer goods spaces. But while fashion retailers also want to take advantage of advanced forecasting techniques, they have mostly focused their AI investments to date on automation and machine learning at a process level. The ability to forecast demand when retailers lack extensive and granular data is something retailers are only just now starting to explore.
The ability to make predictions without sufficient historical data is considered possibly one of the most challenging applications of AI. But the obstacles retailers face in properly stocking individual stores, especially with products with short lifecycles, not only remains, but it has also been greatly intensified by the pandemic, resulting in erratic demand signals and extensive shifts in consumer behavior.
When it comes to merchandise planning and forecasting, AI-driven demand forecasting is possible, but it’s not easy. Here are three considerations for retailers that would like to infuse greater levels of science and certainty into their forecasting processes for products with short lifecycles.
Advertisement
Data-Scarce vs. Data-Rich Environments
Retail historically has been considered an industry with an overabundance of data. Between millions of shopping trips a year and millions of store/SKU combinations’ worth of inventory, the amount of data to track in retail rapidly becomes overwhelming.
However, when using the data for forecasting, there isn’t as much rich data as would appear on the surface. Part of this is because, ultimately, retailers want to be able to forecast at a very granular level, understanding each SKU at a store or warehouse location by day or even by hour.
The grocery industry has seen early success in this regard due to the large volume and long shelf life of each product. Milk, bananas and cereal are all products that consumers buy regularly, which means there’s plentiful data to collect. This data-rich environment makes it highly advantageous to use AI for replenishment forecasting of the grocery shelf.
On the other hand, the fashion industry is typically only able to collect 12 weeks’ worth of data (one selling season’s worth) about a particular product before a new season of clothes is released. There’s also less transaction activity that happens on a daily basis.
Not to mention the variety of different colors and sizes of each apparel item, which makes the data even more granular. When looking at how often one item — such as a black jacket, size small — sells in one store in one day or even one week, the data becomes very sparse indeed. This data scarcity makes AI-driven forecasting much more challenging, but still able to potentially drive a tremendous amount of value.
AI Making Sense of Unprecedented Events
Since the pandemic started, the future is less predictable than it’s ever been, and the past has become less and less helpful as a base for predicting the future. Retailers looking at comparable store sales are seeing a massive increase in March 2021 store comps versus March 2020’s, when stores started to close.
This is because there’s no longer any relevancy in last year’s data to what’s happening right now, especially as vaccines roll out and consumers feel comfortable returning to stores. Furthermore, the data from 2019 is no longer valid either, because consumer behavior has shifted so much since the pandemic due to the rise in omnichannel and online shopping.
Since historical demand is not very helpful in unprecedented events like the pandemic, retailers are relying more on AI to put weight on more recent data, and using AI to intelligently apply algorithms around cyclicality, seasonality and consumer behavior from the past few weeks or months to inform the near future.
AI can identify patterns that are difficult for people to see and understand, and then apply those patterns to predict the future. Machine learning automates this process and allows the AI to take in new data sources that the retailer may never have considered before and analyze their relevancy or impact on a forecast.
Approaches to Data-Scarce, AI-Driven Forecasting
To overcome data-scarce situations and utilize AI to make the best and most accurate predictions possible, retailers can take two different approaches.
One approach is to continually aggregate demand to a higher level until you get enough data to actually use it, and then simply spread the data across the lower levels. Retailers often do this for store forecasting today — they may aggregate a number of stores together in order to get a more complete data set of historical sales, and then use that to create the AI-driven forecast at that aggregated level before then deciding how to take that forecast and apply it to each individual store. Retailers historically used fairly traditional forecasting techniques to allocate forecasted demand across stores in this situation, like using the size of the store or the average sales volume of each store as the basis.
The second approach is similar to a Sudoku puzzle, where actual sales represent the numbers that the puzzle gives you to start with, and where the x-axis of the puzzle is every store and the y-axis of the puzzle is every item the retailer carries. The AI algorithm must use those numbers as clues to puzzle out what could have sold if every store carried every item, which then becomes the basis for a traditional demand forecast.
This approach helps address the problem of color and size. Not every store carries every color in every size, nor sells every color/size option it carries every day or every week. One of the most critical pieces for forecasting demand is to make sure that your sales history also includes a view of lost sales, or sales of items that should have sold but didn’t because the retailer was out of stock.
Each approach on its own can leverage AI techniques to provide a forecast. But combining both of these approaches poses a next-level opportunity for managing data scarcity. Retailers will be able to operate at the most granular level that the data supports, moving up or down the level of aggregation as needed, and also provide a view into whether the most granular combination — store/SKU level — results in a forecast that makes sense for that individual store.
Retailers Must Become Data-Driven in all Instances
While AI continues to shine in its ability to solve data-rich forecasting problems, retailers are on a quest to become data-driven in all instances, including those situations when they do not have extensive and granular data to inform their actions or algorithms.
With the right tools to better understand location-level sales patterns, retailers — empowered by AI, advanced analytics and automation — can have greater confidence in their forecasts, even when the data itself does not show a complete picture.
Nikki Baird is the VP of Retail Innovation at Aptos, a retail enterprise solution provider. She is charged with accelerating retailers’ ability to innovate. Baird has been a top global retail industry influencer for several years, with a background in retail and technology.
