Description
Machine learning (ML) sales forecasting applications occupy a special position in industry and retail, as used in numerous optimization activities, such as price optimization. While optimizing prices, sales forecasting tools require elasticity regarding the optimization objective. When ML methods are used for sales forecasting, there must be sufficient data to learn the proper elasticity requirements. Hence, automated ML-based (price) optimization requires prior (price) optimization sample data. To overcome the problem of low price elasticity of sales forecasts, when historical price-change data are scarce, we present a sales forecasting model marketplace in which trained ML models are shared across companies. Evaluations of a design science project with 43 retail stores show that a model marketplace with neural networks and dropout profiles achieves good forecasting accuracy and high price elasticity. With these results, companies can implement automated optimization algorithms while avoiding extant cumbersome stovepiped methods.
Recommended Citation
Hütsch, Marek, "A Machine Learning Model Marketplace to Overcome Low Price Elasticity of Sales Forecasting" (2022). Wirtschaftsinformatik 2022 Proceedings. 14.
https://aisel.aisnet.org/wi2022/ai/ai/14
Previous Versions
A Machine Learning Model Marketplace to Overcome Low Price Elasticity of Sales Forecasting
Machine learning (ML) sales forecasting applications occupy a special position in industry and retail, as used in numerous optimization activities, such as price optimization. While optimizing prices, sales forecasting tools require elasticity regarding the optimization objective. When ML methods are used for sales forecasting, there must be sufficient data to learn the proper elasticity requirements. Hence, automated ML-based (price) optimization requires prior (price) optimization sample data. To overcome the problem of low price elasticity of sales forecasts, when historical price-change data are scarce, we present a sales forecasting model marketplace in which trained ML models are shared across companies. Evaluations of a design science project with 43 retail stores show that a model marketplace with neural networks and dropout profiles achieves good forecasting accuracy and high price elasticity. With these results, companies can implement automated optimization algorithms while avoiding extant cumbersome stovepiped methods.