Russian convenience store, Pyatyorochka, is working with Yandex Data Factory to accurately forecast demand for products during its regular sales events to optimise delivery volume and meet consumer demand.
Pyatyorochka – owned by Russia's X5 Retail Group – needed to ensure it was not over ordering stock which would lie wasted on the store shelves. The retailer found it was important to invest in accurate machine learning technology during sale periods, because when a cost of a grocery item does not change, it already understands consumer demand is relatively constant.
"In our business, it’s very important to find the balancing point between delivery volume and product sales. That’s why using the modern forecasting methods developed by Yandex Data Factory is such a promising venture for us," said Polina Kiseleva, director of marketing and business support, Pyatyorochka.
During a pilot of the technology, Pyatyorochka and Yandex created a specialised machine learning model based on analysis of historical promotional data across the store estate between April and July 2014. Firstly Yandex Data Factory was able to predict sales for September 2014, and because Pyatyorochka already had the real data on this month's sales, it was able to compare accuracy of the model and was pleased with the similarities.
The next step for Yandex was to predict the consumer demand for new products which Pyatyorochka had no historical data for. Using detailed classifications of all products in the model, the technology was able to learn and predict new products based on data for items from the same product categories.
The retailer had already been using its own tools to predict demand during discount periods, so Pyatyorochka decided to compare forecasts by calculating the mean absolute error for each item.
In order to measure accuracy, Yandex rounded figures to a multiple of the number of items per wholesale package, reflecting retail conditions. Yandex Data Factory's model was able to predict demand for every product on promotion in each of the 91 pilot stores and when used in real-time resulted in an 87% accuracy.
"Demand for discounted products differs greatly from normal buying patterns, that’s why forecasting it is a challenging task," explained Yandex Data Factory’s COO, Alexander Khaytin.
"We successfully solved this task using machine learning and big data analytics technologies. Predictive models can be useful for other retail tasks as well. For example, they can forecast sales figures in stores that haven’t opened yet, or personalise recommendations."