Individual revenue forecasting in the banking sector

Abstract: This paper analyses data from a Swedish bank combined with macroeconomic indicators to forecast revenues for individual customers over the course of four years. Separate models are created for recurring customers and customers who have just joined the bank. XGBoost is shown to outperform linear regression, random forest, neural network and support vector regression when comparing both mean absolute error and mean squared error. Macroeconomic variables reveal little to no significance for such forecasts. Finally, a cluster-based method is proposed where customers are first assigned a cluster and then different models are trained for each cluster. We conclude that such a method is only effective if the classification of customers into their respective cluster is sufficiently accurate.