Forecasting New Product Demand Using Domain Knowledge and Machine Learning

Abstract

Overview: Forecasting demand for new products is a challenging task, as it involves capturing relations of complex variables in markets where little or no historical data exist. Managers usually rely on surveys, intuition, and heuristics to forecast new products. Linear statistical tools used to predict demand for existing products are not suitable because there are not enough data to capture complex nonlinear relations in yet-to-be launched products. Other tools are appropriate for aggregate new categories but not for incremental company-specific products. Machine learning can capture complex nonlinear relations, but it usually requires significant amounts of data. Using an expert’s domain knowledge can circumvent the need for vast training datasets. To support product development activities, we propose a method that combines domain knowledge and machine learning to forecast market share of complex incremental new products. An experiment from the automobile industry shows the approach yields expressive results (82 percent forecast accuracy).