What they wanted For our client, a large non-food retailer, promotional folders are essential: a significant revenue share originates from folder products. But sales forecasts were on average off by 100% (volume weighted MAPE) resulting in excess stock, tied up working capital, sharply increased operational complexity and added discount pressure. Our client wanted to move away from an opinion and anecdotal-evidence based forecasting system toward a rigorous historical data driven estimate.
What we did Using our proprietary data engineering algorithms we started by unifying data from 35+ disparate data sources into an automated and centralized database with a complete end-to-end product level view incorporating 100+ product characteristics, e.g. size, color, and type.
Using our database structure we deployed a collection of gradient boosted regression trees and additional statistical techniques to make product-level predictions of folder sales. We improved forecast accuracy by 60 percentage points. Key predictive variables were seasonality, markdown amount, price, article category and location of the product within the folder.
What we achieved Given the significant improvement in forecast accuracy, our algorithm was implemented in core business processes. Today, roughly 80% of our model forecasts are used as a final sales estimate. Because of this, the supply chain management team can focus its attention on the remaining forecasts where the math just doesn’t cut it. In these predictions the value is provided by retail sector insidership and other unique human insights.
Our applied combination of human expertise and mathematical precision results in a reduction of 22% in handling complexity, a 20% reduction in excess stock, and – most importantly – the start of a data driven culture.
What they said
“This is pure gold.”