Deep Learning and Statistical Approaches in Financial Modeling of Foreign Assets and Liabilities of Nepal’s Banking System

In an environment marked by financial volatility and rapid economic shifts, reliable forecasts are critical for informed policy-making and strategic financial planning. This study investigates the detailed mathematical exploration followed by its computational performance of time series and deep learning models namely Auto Regressive Integrated Moving Average (ARIMA), Recurrent Neural Network (RNN), and Temporal Convolutional Network (TCN) applied to foreign assets and liabilities of banking system of Nepal consisting monetary authorities and various depository corporations. Using available primary data, we analyze trends, seasonal patterns, trajectories and their descriptive statistics to capture underlying behaviors. Also, in our work, we have identified the optimal ARIMA order that most effectively captures the linear trend. Empirical study shows that the RNN is able to handle the non-linear patterns which is determined by performance metrices on the training and testing split. TCN being computationally extensive model is not able to capture robust relation of the data due to lack of long-range dependencies and large time windowed dataset for training. Based on our results, the RNN could be used as the most suitable time series forecasting model for the foreign assets and liabilities of Nepal as it enhances the accuracy, minimizes error, and improves effectiveness in contributing to decision-making in banking system. JEL Classification: C22, C45, E44, E60