The solution based on AI of Jet Infosystems predicted the redemption of goods in Duck-bill with an accuracy of 80%

On February 12, 2019 the Jet Infosystems company announced implementation for Utkonos hypermarket of the pilot ML project (Machine Learning) on prediction of volume of the redemption of perishable goods. Forecast accuracy on different categories averaged 75–80% at planning horizon in 2 days (such is regularity of deliveries to a warehouse).

The analytical system used by Utkonos company with a high accuracy predicts the redemption of goods for a week. However where shorter planning horizon is necessary, its accuracy considerably decreases that attracts potential losses. Unlike the classical BI systems, solutions based on machine learning allow to consider bigger number of factors, so, give more opportunities for analytics.

It is extremely important to us not to disappoint on the one hand the buyers with lack of goods necessary to them, and with another — not to buy surplus from suppliers, especially if the speech about perishable products. This pilot project showed us possibilities of machine learning in the field of demand forecasting and also revealed a number of the interesting patterns affecting behavior of buyers. Dmitry Sukhorukov, chief of the department of planning and merchandising of Utkonos company

The project mentioned several trade names in category "The Meat Cooled" as it is goods of short term of storage and difficult conditions of deliveries by which it is especially important to determine required purchasing amount precisely. Also the category "egg" on which within a year the specific seasonal demand is traced was considered, and Machine Learning just allows to reveal the hidden interrelations.

Experts Jet Infosystems constructed a mathematical model and trained it at historical data on purchases in Duck-bill for 2 years. At the same time information not only on a commodity remaining balance, but also a production calendar with days off and holidays and also data on weather conditions was considered. On a historical interval in 2 months the predictive accuracy of the created model appeared at the level of 80%, and on an interval of half a year — about 75%.

For the last year turnover of Duck-bill considerably grew, including thanks to saturated marketing activity of the retailer. It complicated our task since did not allow to apply correctly historical data of last years to the corresponding periods of the current year. Nevertheless due to attraction of optional external data we managed to implement the ML solution of rather high accuracy which can give noticeable economy of means on purchases and real increase in revenue at competent supply of a warehouse. Vladimir Molodykh, director of development and deployment of software Jet Infosystems