Artificial Intelligence capabilities have transformed the promotional forecasting process

The quality of a promotional sales forecast affects the service level, write-offs, customer relations, storage and shipping costs, and much more: the price of improving accuracy by every percentage point can turn into meaningful efficiency gains. Given the fact that the majority of company’s business is related to perishable products, it takes extra effort to build and improve forecasting quality.

Following the company’s global strategy to harness AI technologies in order to improve business process efficiency, the project team, including client’s employees from Supply Chain Management, Marketing Review, IT, and the Data team, sought a next-generation IT solution based on machine learning algorithms. After analyzing the technology market, the decision was to develop a customized ML prediction model, taking into account all the specifics of the business process and the data available to manufacturer.

The team of consultants developed and launched an ML model for predicting sales during promotions with key national retailer clients:

• A multi-parameter model was developed using a wide range of data: historical shipments, off the shelf sales, promotional plans, price levels, orders from customers, stock levels, calendar events, assortment changes.
• The model is based on a combination of several ML algorithms such as Ridge, Lasso, KNN, XGBoost as well as data normalization and automatic feature selection.
• The forecasting horizon embedded in the model consists of current and next quarter by month, broken down by product lines, promotions (“slots”) and customers.
• Automatic integration between incoming data and the model was set up, and the functionality to run automatic recalculation of the forecasts by the business team was defined.
• The existing process was audited and a redesign of the promotional forecasting process was realized, taking into account the advantages of the new AI technology.
• The team of consultants keeps the technical post-project support of the model in “business as usual” format through continued support and scaling.

The accuracy of ML forecast increased up to 74.3% – up almost 4% in comparison with the former IT solution (70.6%).