Application of Machine Learning for Predicting U.S. Bank Deposit Growth: A Univariate and Multivariate Analysis of Temporal Dependencies and Macroeconomic Interrelationships

Abstract

This research applies machine learning algorithms to predict the growth rates of bank deposits in the United States using data from 1973 to 2019. The dataset includes weekly deposit records from U.S. commercial banks and key macroeconomic indicators, including GDP, inflation, money supply (M2), recession periods, and interest rates, obtained from the Federal Reserve Economic Data (FRED). The study involved preprocessing steps including date conversion, stationarity testing with the Augmented Dickey-Fuller (ADF) test, and differencing to achieve stationarity. Various models were tested for univariate time series analysis, including SARIMA, Prophet, ETS, LSTM, and Transformer models. LSTM demonstrated the highest predictive accuracy, with the lowest error metrics and the highest R² value, proving effective in capturing complex temporal dependencies in deposit data. The study conducted a multivariable analysis incorporating several macroeconomic indicators to explore their relationship with bank deposits. This process included feature scaling, creating lag features, and preserving temporal order during data splitting. Recurrent Neural Networks (RNNs) were evaluated with different lagged periods to assess their impact on model performance. The results indicated that while increasing the number of lags improved the model’s fit to the training data, it did not consistently enhance performance on unseen data, highlighting the trade-off between model complexity and generalization. Cointegration analysis confirmed long-term equilibrium relationships between bank deposits and macroeconomic indicators. Further analysis using FMOLS and DOLS revealed that inflation and recessions negatively impacted deposits, while M2 and GDP had positive effects. This study demonstrates the effectiveness of machine learning models, with LSTM proving particularly successful in forecasting bank deposit growth rates. Incorporating multiple macroeconomic variables significantly enhanced predictive accuracy, providing valuable insights into the factors influencing deposit levels. This research contributes to financial forecasting by showcasing the ability of machine learning techniques to integrate economic dynamics into predictive models. This research contributes to the field of financial forecasting by demonstrating the efficacy of machine learning techniques in economic analysis.