Discovering customer behavioural patterns from financial transactions

Abstract

In an era where there is a significant need of converting huge amounts of data into useful and valuable information, the task of data mining saw a huge increase in importance, popularity and applicability. Data mining refers to the process of examining large volumes of data so as to discover hidden patterns, relationships and other insights conveyed in the data. It allows companies to process their data, while aiming to produce new growth opportunities in order to outperform their competition. Nowadays, many companies and businesses are resorting to data mining techniques, to save costs and time, as well as to understand customers’ needs and market conditions. Analysing such data is beneficial for any company, leading to better informed business decisions, higher profits and more contented clients. In this dissertation, we present a data mining study that is applied on millions of transactional records that were collected for a number of years, by a leading virtual credit card company based in Malta. In this study, 2 machine learning techniques, namely Artificial Neural Networks (ANNs) and Gradient Boosting (GBM), are analysed to identify the best modelling framework that predicts the churning behaviour of this company’s customers. Apart from helping the marketing department of this firm by providing a model that predicts customers that are willing to churn, in this study we contribute to literature by analysing and identifying the minimum amount of customer activity needed, to predict churn. In addition, we also analyse the “cold start” problem by performing a time-series experiment based on the hardly any data available at the beginning of the customer purchase history. When evaluating our system, we have seen that the Gradient Boosting Model (GBM) predictive model is more suitable than an Artificial Neural Network (ANN) in our Customer Relationship Management (CRM) problem, since it is capable of identifying the majority (70%) of churners correctly whilst the constructed ANN is more capable of classifying the non-churners. Furthermore, we have shown that reducing the observation window size does not reflect in huge performance loss, thus giving the ability of leveraging prediction performance with the amount of data observed. Finally, we have seen that the demographic information present in our dataset is not effective to predict churn behaviour of customers.