The retailer solution makes forecasts for 5,000+ products. Using three years of accumulated data on daily sales, purchased items, average checks, loyalty card types, etc., the application identifies seasonality in demand by month and day of the week, determines trends for each product group, and evaluates the impact of discounts and promotions on product sales. Based on this analysis, the program builds sales forecasts for 1, 3, and 12 months, as well as identifies changes in demand for individual groups of customers.
As a result, the solution helped increase the accuracy of forecasting for the next 3 months from 77% to 93%, which helped optimize logistics and storage of goods in the warehouse.
What Was Done
A dataset with a time series of sales by product subcategory was prepared from the raw data.
For modeling, various approaches were tested for predicting time series: ARIMA, ARCH, recurrent neural networks, etc.
At each stage, the symmetric mean absolute percentage error (SMAPE) was calculated on the model predictions and actual sales for the last 3 months. Thus, for each subgroup, the seasonality of demand for goods was identified: Annual Monthly Weekly.
As a result, the solution helped increase the accuracy of forecasting for the next 3 months from 77% to 93%.
Technologies and tools: Python, ARIMA, ARCH, Recurrent neural networks, Prophet.