Boosted Varying-Coefficient Regression Models for Product Demand Prediction

Abstract

Estimating the aggregated market demand for a product in a dynamic market is critical to manufacturers and retailers. Motivated by the need for a statistical demand prediction model for laptop pricing at Hewlett-Packard, we have developed a novel boosting-based varying-coefficient regression model. The developed model uses regression trees as the base learner, and is generally applicable to varying-coefficient models with a large number of mixed-type varying-coefficient variables, which proves to be challenging for conventional nonparametric smoothing methods. The proposed method works well in both predicting the response and estimating the coefficient surface, based on a simulation study. Finally, we have applied this methodology to real-world mobile computer sales data, and demonstrated its superiority by comparing with elastic net- and kernel regression-based varying-coefficient model. Computer codes for boosted varying-coefficient regression are available online as supplementary materials.

Key Words:

• Boosting
• Gradient descent
• Tree-based regression
• Varying-coefficient model

SUPPLEMENTARY MATERIALS

Code: The R source code for boosted varying-coefficient linear model and simulation study. A readme file is included in the archive (boost.zip).

ACKNOWLEDGMENTS

The first author gratefully acknowledges the constructive comments from Kay-Yut Chen, Enis Kayis, Jose Luis Beltran, Ruxian Wang, and Shailendra Jain in Hewlett-Packard Labs, and Guillermo Gallego at Columbia University. The authors also thank the associate editor and two anonymous referees for providing comments that helped substantially improve this article.