The biggest shopping season in the U.S. is here. Dillard’s, a major department store chain, turned to IBM’s Snap ML AI tool to better forecast demand at a time when trends change faster than the weather.
With holidays fast approaching, retailers are anticipating a spike in demand. To help estimate what to prepare in its warehouses and on store shelves, the US department store chain Dillard’s has turned to IBM’s AI tool, Snap ML.
Machine learning and AI have been increasingly penetrating enterprise. They offer flexibility and adaptability to changing business conditions that existing rule-based decision systems can’t address. The retail sector in particular has seen a rapid proliferation of AI systems in the past few years, and demand forecasting is one important business process that retailers have looked to improve with the help of AI.
Demand forecasting is how companies predict future customer demand for a given product and period, using historical sales and other data sources – including days of the week, seasons, and customer preferences. Accurate demand forecasting gives stores crucial detail about how their products will likely sell, and enables them to make informed, data-driven decisions about pricing, investments, and growth strategies. Demand forecasting is the most widely used AI application in supply-chain planning, according to Gartner.
In the past, forecasting has traditionally been carried out using statistical methods that mainly come down to regressing the time-series trend from historical sales data. This method can be useful for forecasting basic products with stable demand. But with the way the world moves today, where customers are affected by fashion trends and social media, product demands can be irregular and volatile. And predicting demand for new products is especially tricky, as there is no or very little historical sales data for it.
AI can provide help, by not only incorporating historical sales data but also near-real-time data from advertising campaigns, prices, local weather forecasts, and other sources. Accounting for such factors can result in accurate predictions of demand, which retailers can use to perform better inventory allocation, assortment, and replenishment planning. Gartner predicts that half of supply chain organizations will have invested in artificial intelligence and advanced analytics by 2024.
This is where Snap ML comes in: The AI library developed by IBM Research can be used to help retailers, among other businesses. Now publicly available, Snap ML implements some of the most popular ML models, including generalized linear models, random forest, and gradient boosting machine. It also provides a familiar scikit-learn compatible API.
What truly differentiates Snap ML is its ability to accelerate model training and model inference runtime while not sacrificing accuracy. Although Snap ML can exploit GPUs, much of the acceleration comes from algorithmic optimizations and can also be achieved on CPU-only systems. Acceleration is important when training complex models, such as boosting machines, or when the training data is very large. Acceleration is also crucial when optimizing the parameters of a given model, known as hyper-parameter tuning (HPT), involving multiple training rounds of the same model structure but with different parameters. It can also be useful when a user needs to train multiple models to achieve different objectives, such as to forecast demand for different products.
Dillard’s is one of the largest retail companies in the US, with 276 stores and an online shop. The company has been an early adopter of AI techniques for demand forecasting, which it has since applied to optimizing inventory allocation, assortment, and replenishment planning. For the fine-grained product demand forecasting it’s after, Dillard’s needs to train hundreds of ML models and re-train these models periodically, once every few days, to account for fresh data. Each model involves extensive HPT to reach the best possible forecasting accuracy. With this in mind, Dillard’s needs an AI framework that provides high accuracy and the shortest possible runtime.
Dillard’s uses AI models built on tools like IBM’s Snap ML to predict the average rate of sales during various timeframes, called forecast horizons. The retailer’s specialists also build seasonality profiles with time series models and rely on them to adjust the average sales rate accordingly – meaning that the demand forecasts peak at the right times of a year depending on the season.
After a series of trials, Dillard’s has seen results, such as quick turn-around time for retraining Snap ML models for inference. The retailer is now able to retrain Snap ML models every week, making it easier to adapt to ever-changing demand trends. Dillard’s then tests the models and deploys them in production, effectively reducing the production time by an hour. Crucially, the chain has also trained a set of Snap ML models for new products, addressing retailers’ “cold-start problem,” where it’s difficult to predict demand for products with little or no sales data. The models learn well from similar products with more historical data – giving a decent forecast. Snap ML enables them to quickly and accurately estimate future demand for numerous products, based on which they can order the right amount of stock at the right time and for the right store.
Going forward, Dillard’s is planning on using Snap ML for seasonally adjusted forecasts. This approach would not need any seasonality profiles and has the potential to pick up recent trends in the demand as well as minimize the cold-start problem by more accurate forecasts when compared to the average-rate-of-sales models.
Snap ML is maintained by IBM Research and evolves continuously. One main source of inspiration for us for creating future library capabilities is the interaction with customers, partners, and software vendors. Beyond retail, there are multiple use cases in other sectors, such as in financial services and insurance, where we successfully engage with customers and users. If you have a machine learning training and/or inference use case and would benefit from runtime acceleration, feel free to try Snap ML or reach out to our team.