Forecasting Bank Net Interest Income: A Comparative Study of Statistical and Machine Learning Approaches

Sammanfattning

This thesis investigates forecasting of Net Interest Income for banks using publicly available macroeconomic and financial data, focusing specifically on the Nordic bank Swedbank as a case study. A comparative analysis of traditional time series models (ARIMA, ARIMAX, VAR) and machine learning methods (Neural Networks, Random Forest) is conducted. The study introduces two different strategies for input variables: traditional dimensionality reduction through Principal Component Analysis and a qualitative domain-informed variable selection strategy. The combinations of model and variable selection are assessed through rolling window backtesting to understand performance under realistic forecasting conditions. The findings indicate that models based on qualitative variable selection consistently outperform the PCA-based models, both in terms of interpretability and forecast accuracy. Neural Networks and Random Forest produce the most accurate forecasts using a selected set of economic indicators and bank figures, achieving the lowest overall mean absolute errors. These results highlight the importance of informed feature selection and model transparency in financial forecasting, suggesting that compact, economically grounded variable sets can enhance predictive performance without sacrificing interpretability. The study provides practical insights for financial institutions aiming to improve risk management and strategic planning in dynamic economicenvironments.