Demand forecasting is a complex balancing act between a business having the right amount of product in stock to ensure it does not run out at inopportune times, causing customers to buy elsewhere, and having too much, resulting in unnecessarily high costs of inventory. Over the last few months, we have all been challenged with access to daily groceries and supplies that we had previously taken for granted due to the COVID-19 pandemic. With disruptions across the food supply chain, there’s concern about the devastating impact this could have on the world’s poorest citizens. According to an estimate made by the U.N. World Food Programme (WFP) in April 2020, at least 265 million people are at risk of going hungry this year, partly due to the fragile food supply chain, which may be exacerbated by resulting lags in supply chain response times as demand fluctuates in an unexpected way causing retail stock-outs.

Under normal circumstances, retailers often stock a large range of products, which results in a long tail effect, as it’s referred to in statistics and business. This effect is linked to the 80-20 rule, where 80% of daily revenue comes from 20% of the product portfolio, making the demand for them more predictable and forecastable; in contrast, a significant number of other products (a long tail of 80%) are typically low volume with a less predictable profile.

Critical or more popular products, referred to as the head, are usually easier to forecast, while the long tail of many unique items which are less popular need to be highlighted in order to form a replenishment strategy when forecasts are unreliable. In other applications of sales, forecasting can fold extra layers of complexity. For instance, the freshness of hot food is highly important to retailers in the food-on-the-go market, where they carefully manage when new sandwiches and hot products are made. If demand is higher than expected, stocks can run out, and customers can be disappointed; conversely, if demand is low, stock can go cold, and the customer experience is poor.

Richard Benson

Head of Data Science

Balancing Supply & Demand Efficiently

Keytree, a Deloitte business, is an international award-winning design and technology consultancy and product developer, helping to transform digital potential into business reality. We help companies become leaner, faster and more agile with our team of SAP experts and user-centric applications. With this experience, we were able to address the problems our customers were facing by developing Keytree TimeSeer demand forecasting platform.

Built around cutting edge machine learning techniques, TimeSeer incorporates available operational data such as point of sale systems and shipment data, as well as demand influencing factors (DIFs) such as time-of-year, weather, promotions and more to give a single consolidated unit forecast with a special “expectation” score, indicating if the forecast is within normal bounds or is falling outside expected ranges, allowing increased time to enable the supply chain to adapt. The TimeSeer algorithm shows how predictable each product line is and how much a forecast can be relied on. For items that are forecastable, TimeSeer produces 7, 14, and 28-day forecasts of up to 93% accuracy to enable not only replenishment planning but also ensuring the business has current and future access to the human capital it needs to perform effectively.

With regards to food-on-the-go retail, TimeSeer analyses rates of product sales at each store, and from there derives live intra-day optimal bake times to avoid stock-outs and maximise freshness. In case of unexpected demand surges, recommended make and bake times are adapted, and stock-out warnings are given so staff have time to react.

Accelerated Performance with Optimised Libraries on Intel Architecture

As a member of Intel’s AI Builders Program, Keytree has access to the latest generation Intel Xeon Scalable processors; this access combined with technical support from Intel engineers, meant we were able to optimise TimeSeer to provide our customers with enhanced performance at a potentially lower cost.

One of the optimisations used to gain a major performance boost on Intel’s latest hardware was daal4py, which was created to give data scientists the easiest way to utilise Intel DAAL’s (Intel® Data Analytics Acceleration Library) powerful machine learning building blocks directly in a highly productive manner.

A simplified API gives high-level abstractions to the user with minimal boilerplate, allowing for quick-to-write and easy-to-maintain code.

For framework designers, daal4py has been fashioned to be built under other frameworks from both an API and feature perspective. The machine learning models split the training and inference classes, allowing the model to be exported and serialised if desired.

This design also gives the flexibility to work directly with the model and associated primitives, allowing one to customise the behaviour of the model itself.

Looking Forward to Future Collaboration

The Intel AI Builders team were essential in enabling us to achieve crucial optimisations for our forecasting algorithm solution and provided deep support to ensure we got maximum performance utilising Intel technologies. Having access to cutting edge hardware architectures through the program enables us to match and spec our algorithms to the optimal server configurations and appliances quickly and efficiently, saving time and money.

Keytree stands at the forefront of change, development and implementation experience in the retail sector, working closely with leading international retail brands.  We look forward to building on our successful partnership with Intel and the continued support to optimise and promote our solutions on Intel Architecture.