By Daniele Barchiesi, Senior Data Scientist at Think Big Analytics, A Teradata Company.
How many Granny Smith apples should I deliver to each of my supermarkets? Will there be a run on sun cream this weekend? These are the types of questions retailers are asking when attempting to apply demand forecasting, a way of predicting demand levels for products or services to plan and manage supply chains more effectively.
Historically, statistical demand models have been built using either relatively simple techniques and limited data, or manual processes and aggregated data. However, given today's explosion of data, and the new predictive machine learning based models available to companies, there's more opportunity than ever before to use predictive supply and demand insight to stock the right products and sell them as quickly as possible.
These new models are transforming the way many retailers are achieving astonishing new business results. But how exactly is machine learning helping companies uncover new insight to fuel an on-shelf retail revolution?
Capturing and analysing demand in real-time
Today, instead of having to rely on techniques that predict demand based only on past sales, new machine-learning demand models are using a wide variety of live data sources, including competitive pricing information, and external variables such as weather forecasts.
Beyond this, the models are combing and prioritising the data they analyse: for example, for some products demand might depend on a categorical distinction between promotion versus full price that is not dependent on discount percentage. Additionally, the sales pattern for sun cream is likely to be reliant on the weather forecast in the area it is being sold in. Machine learning based demand models can 'learn' these relationships from data, needing minimal human intervention while using them to effectively predict demand.
Smart operations for supply chain management
Not only can this framework be used for building general and robust demand models, but its many different components can be combined to tackle specific challenges, for example, reporting out of stock items for major grocery retailers. Items become out of stock for a variety of reasons that link to operational processes in stores, as well as supply chain management factors.
It is estimated that between three and six per cent of items in UK supermarkets is out of stock at any given time. This impacts significantly on missed sales opportunities, which can amount to millions in losses, as well as have a negative impact on customer experience. Moreover, retailers pay their staff to manually check shop floor product levels – even if restocks are unnecessary in most cases.
Missed opportunities and operational costs could be heavily reduced if retailers could reliably estimate when and for which products out of stocks occurred. For example, they could order or bake more pastries if these tended to be unavailable at the end of the day, or only check specific products when they are likely to be missing, freeing up time previously required for a repetitive stock count rota.
That's where new machine learning based demand forecasting models involving advanced analytics capabilities can step in. By combining stock data with historical sales patterns, retailers can estimate the expected demand for any given product at any given store location and compare it with sales data and insights, potentially saving tens of millions of pounds annually in operational costs.
The future of retail forecasting
In the future, these demand models will become powerful tools to tackle a wide range of other supply chain challenges – and will ultimately separate the winners from losers. Promotion design can be improved through a data-driven approach that integrates with other functions and operations such as media and in-store support. Also, price reductions can be optimised to reduce waste for highly perishable products while lowering lead times but at the same time increasing availability.
Frameworks will combine flexible and advanced predictive models with robust, business-validated optimisation engines to enable better pricing decisions. For example, these systems will help to inform promotional planning, strategic pricing, markdown pricing and reduced to clear items by analysing and optimising different business functions to maximise revenue, sales volume, and margins.
Leveraging and deploying machine learning at scale is what will enable the most innovative organisations to build the products and services of the future: cutting-edge data science along with the emerging big data technologies are set to continue empowering high impact business outcomes for retailers in 2018 and beyond.