Multiclass Cross-selling Model for Savings and Investments Using Gradient Boosting

Abstract

Danske Bank has for several years modeled customer purchase behavior on category level (e.g. savings or investments). This thesis is a first attempt at predicting (first time) customer purchase behavior on product level. Five products within two categories were chosen and modelling was done with python using gradient boosters (mainly XGBoost, but also Light GBM). Results indicated that predicting product purchase is possible, although not with the target selected for this thesis. Multiclass modelling gives additional insight into customer behavior compared to models on category level, however, additional tuning of the models are required before the accuracy reaches the same level as the category prediction models.

Popular Abstract

How well does your bank know you? As artificial intelligence becomes increasingly sophisticated, the demand for it increases throughout various industries. The financial sector is no exception. Being one of the largest banks in the Nordics, Danske Bank has already employed machine learning for several years. But can these techniques be even more refined?

As a customer oriented bank, Danske Bank aims at being relevant for its customers and their needs. For many years, the banks advisers have received A.I. generated leads on customer purchase behavior. These leads will let an adviser know if a specific customer is interested in a product from one of the banks many product categories (e.g. savings or investments). However, as the categories are not necessarily known the customer, the advisers are often required to make some research on their own before approaching the customer with suggestions on specific products. This brought up the question: Can customer behavior be modeled on product level, rather than category level? If so, it would move leads from internal business definitions to customer oriented definitions (as the customers often are familiar with basic bank products). It would further give advisers a clearer entry point into the conversation with the customer, saving them time spent on research. Using a machine learning technique known as "gradient boosting", it was shown that one could indeed model customer behavior on product level. Gradient boosters use an ensemble of "weak" predictors (predictors that are only slightly better than random guessers) to generate a prediction that is like one from a "strong" predictor (a predictor that has a high accuracy). In recent years, sophisticated free-to-use boosters have emerged which allows anyone to employ the power of machine learning. This project used two boosters known as XGBoost and Light GBM; the former being more accurate and the latter being faster. Throughout the project, the model was tuned using feature engineering and hyperparameter optimizers. In the end, the final model showed clearly that additional information could be gained by modelling on product level. However, a different target and finer tuning is required before the product level result is satisfactory enough for deployment.

Hyperparameter - Parameter whose value is decided manually (e.g. by the user) rather than by the machine learning algorithm.

• author: Sohrabi, Arwin
• supervisor: Alexandros Sopasakis
• organization: Mathematics (Faculty of Engineering)
• course: FMNM01 20181
• year: 2018
• type: H2 - Master's Degree (Two Years)
• subject: Mathematics and Statistics
• keywords: Machine learning, gradient boosting, multiclass target, XGBoost, finance, Danske Bank, customer behavior
• publication/series: Master's Theses in Numerical Analysis
• report number: LUTFNA-3047-2018
• ISSN: 1404-6342
• other publication id: 2018:E76
• language: English
• id: 8969104
• date added to LUP: 2019-07-15 11:41:44
• date last changed: 2019-07-15 11:41:44