The Ensemble Method for Censored Demand Prediction

Many economic applications, including optimal pricing and inventory management, require predictions of demand based on sales data and the estimation of the reaction of sales to price change. There is a wide range of econometric approaches used to correct biases in the estimates of demand parameters on censored sales data. These approaches can also be applied to various classes of machine learning (ML) models to reduce the prediction error of sales volumes. In this study we construct two ensemble models for demand prediction with and without accounting for demand censorship. Accounting for sales censorship is based on a censored quantile regression where the model estimation was split into two separate parts: a) a prediction of zero sales by the classification model; and b) a prediction of non-zero sales by the regression model. Models with and without censorship are based on the prediction aggregations of least squares, Ridge and Lasso regressions and the Random Forest model. Having estimated the predictive properties of both models, we empirically test the best predictive power of the model taking into account the censored nature of demand. We also show that ML with censorship provides bias corrected estimates of demand sensitivity to price change similar to econometric models