The availability of large amount of data on consumer behaviour has resulted in multiple opportunities for marketers to design effective marketing strategies. The transaction level data can generate useful insights on consumption patterns in a wide variety of settings. In this research, we focus on analysing customer behaviour data to estimate the size of wallet (SOW) of the customer using machine learning techniques. SOW is the measure of spend capacity of a customer on a specific product category across different firms. Hence there is a direct financial interest of the firm in increasing the SOW of a customer. We highlight the benefits of estimation of SOW using machine learning techniques over the traditional time-series approach and benchmarking approach. We determine the performance of the two quantile modelling approaches namely K-nearest neighbour (KNN) and quantile regression (QR) in estimating SOW and compare their results. The dataset used for analysis is a publicly available dataset at UC Irvine (UCI) repository which includes information on yearly spending in monetary units on different product categories for clients of a wholesale distributor. We also use machine learning based clustering techniques to segment the customers by size and industry to generate better estimates for SOW. Finally, we compare the two quantile modelling techniques by comparing their error and discriminatory power. We provide managerial implications of our study.
