Tesco’s Intelligent Trading Platform (ITP) provides a mechanism for automatically recommending the prices of online general merchandise sold on the Tesco Direct website. Using the ITP, Tesco’s Commercial Science Team has developed a framework for ensuring customers get competitive prices for products they buy on Tesco Direct. This framework is comprised of the following parts:
• a simple linear forecasting model to predict the total expected week-ahead demand of online merchandise, and
• an optimisation algorithm that calculates the best price for each product within Tesco’s customer focused pricing framework.
The Smith Institute was asked by Tesco to develop novel, and more robust, forecasting algorithms capable of generating accurate predictions of week-ahead demand. The intention was to use the resulting forecasts within the pricing framework. Furthermore, the Smith Institute was entrusted with the task of proposing a suitable set of metrics to review and compare the performance of the proposed models at forecasting week-ahead demand.
The Smith Institute designed a suite of algorithms to forecast the demand of general merchandise for use within Tesco's pricing framework. In particular, we developed and tested two independent classes of demand forecasting models.
The first class of models comprised of a selection of statistical models for count regression. More specifically, three independent models for count regression were adapted to model the demand for a specific product as a function of various features (inputs). Initially a standard Poisson regression model was trialled where the parameter of the underlying Poisson distribution, λ, was the mean number of sales and was modelled as a linear combination of underlying features. Next, a negative binomial model was proposed as an extension of the first model. This was able to accommodate excess variability in expected demand, which the standard Poisson model could not effectively take account of. Finally, a hurdle mixture model was proposed, which in addition to excess variability, could accommodate excess zero demand values. The parameters of all three models were learnt using actual historical data.
The second class of models were based on supervised machine learning (ML) algorithms. In this instance we chose regression models. Each ML algorithm was adapted to predict the week-ahead expected demand of each product and the algorithms were trained on historical data. Amongst the possible approaches, we selected Random Forests and Gradient Boosted Regression, both of which are a class of ensemble methods for machine learning, because of their known ability to perform well on a wide range of supervised learning tasks.
An important aspect of the work consisted of selecting the right set of features for both classes of models in order to improve their predictive performance. The feature selection strategy was informed both by domain knowledge provided by Tesco staff and empirical observations, with a view to improving the predictive ability of the individual models.
The Smith Institute also proposed a set of metrics based on Scoring Rules to provide a comparison between the different demand forecasting models. These metrics give an absolute measure of how well a particular model describes the data.
The Smith Institute was able to provide Tesco’s Commercial Science Team with an overall framework for forecasting and optimisation. The advantage of including two approaches was to give more robust foundations for further work and higher confidence for potential large-scale deployment.
“As a business, serving customers is at the heart of everything we do. Improved forecasting and optimisation means we are able to offer the best value to our online customers. This framework enables Tesco to offer its online customers products at the best possible price, which in turn helps us to meet our core purpose of serving Britain’s shoppers a little better every day.
Collaborating with the Smith Institute allowed us to explore potential modelling approaches in a very short time-frame before our live trials and rollout. This gave us more confidence in our final approach and led to very successful early trials.”Ben DiasLead Data Scientist at Tesco