Beltel Datanomics developed a forecasting model for products with intermittent demand
Task
FOREST is a leader in advanced fastening technologies. It is necessary to test the applicability of automatic forecasting for goods with intermittent and irregular demand.
Solution
During the execution of the project, a research analysis of sales data was conducted, it showed:
- Strong variability in the data, caused by a large number of days with zero demand
- The absence of a distinct annual seasonality in sales
- Anomalies in the data in the form of a sharp increase in sales in one day, which is explained by a successful marketing campaign
All of these points indicate a high factor of randomness in sales for most distribution centre (DC)/product pairs and the difficulty of applying classical approaches to time series forecasting. Therefore, a sales volume forecasting model was developed and implemented for two months ahead and two quarters ahead, which takes into account not only changes in sales volume, but also works with the frequency of demand for a particular item and cleans historical data from abnormally large sales.
Results
The models were validated on 12 periods of two months each. Quality metrics (MAE, SMAE, RMSE, SRMSE, average deficit and surplus) were calculated for each test period. Cross-validation on 12 test cases showed better quality metrics for the model developed by Datanomics than for simple statistical models such as Naive and SES. For example, for the SMAE metric, the improvement was 15%, and for the average surplus, it was 20%.
The ultimate goal of the project is to generate an order to replenish certain positions in all distribution centers for two months/quarters ahead.
«As a leader in our field, we strive to apply advanced technologies and automate processes to increase the efficiency of the enterprise,» says Alexandra Sinitsyna, project manager at FOREST. «This project was an experiment for us. We wanted to test the applicability of automatic forecasting for products with our demand characteristics. The project was not smooth — efforts were required at the data extraction stage, but we confirmed the hypothesis and obtained a high-quality forecast result in the context of DCs/products and detailed further steps of development. I would like to separately note the high level of expertise of Beltel Datanomics specialists.» (Quote translated from Russian)