Demand prediction for e-commerce advertisements: A comparative study using state-of-the-art machine learning methods

Abstract

Demand prediction for sales has mostly been studied with respect to historical demand and pricing models as dominating predicting factors. While these factors effectively predict demand for physical retail, they do not address the volatility of demand in Consumer-to-Consumer (C2C) e-commerce which exhibits a perfect competition market structure. This study establishes that advertisements are an effective measure in determining the expected demand of products in C2C e-commerce. It highlights the importance of product description, images and context of advertisement in estimating the deal probability for products. Previous demand prediction studies have delineated the performance boost obtained from using Artificial Neural Networks (ANNs) over linear regression models. This study does a comparative analysis of the performances of the state-of-the-art (SOTA) linear models, Decision Tree ensembles, and deep learning methods. We present benchmark results for this task on a dataset containing approximately 1.4 million training examples of C2C e-commerce advertisements. The results reveal that deep learning is by far the most effective method for demand prediction studies.