GOALS AND OBJECTIVES
-
BUSINESS OBJECTIVE
To forecast the buyout of perishable goods more accurately, so as to increase the efficiency of procurement decision making.
-
IT OBJECTIVE
To build a mathematical model and train it using historical data.
IMPLEMENTATION
Utkonos ONLINE is actively developing its machine learning initiatives. The analytical system currently used by the retailer can predict buyout of goods for the week with a high degree of accuracy. Where a shorter planning horizon is needed, however, the accuracy of the current model is noticeably lower, entailing potential losses. Unlike classical BI systems, machine learning solutions take account of a greater number of factors and, therefore, have greater analytical potential.
In this case, Jet Infosystems specialists built a mathematical model and trained the model using two years of historical purchase data from Utkonos ONLINE. In building the model, not only inventory balance data, but also data relating to the production calendar (including weekends and holidays) and weather conditions was taken into account. Over a two-month interval, the model is able to forecast as accurately as 80%; over a six-month interval, accuracy levels are approximately 75%.
The project covered a range of products which have short shelf-life and challenging supply conditions, yet for which it is extremely important for the company to accurately determine required purchase volumes. In addition, some categories with non-standard (for example, seasonal) demand were analyzed, given that Machine Learning techniques can reveal hidden correlations.
PROJECT RESULTS
Despite difficulties with the accuracy of historical data over the past two years (the turnover of Utkonos ONLINE over this period has grown significantly, making it difficult to apply historical data to present conditions), Jet Infosystems specialists were able to harness additional data so as to implement a ML solution with sufficiently high accuracy, resulting in significant procurement economies and real increases in revenue due to an increase in the efficiency of warehouse supply.
-
2 years
Historical database for training the ML-model
-
2 days
Planning horizon enabled by the new ML model
-
75-80%
Forecast Accuracy for various categories