Exploring Trajectories of Government Bonds for Debt Planning Using Machine Learning Models

Abstract

Reliable projection of government bond markets is crucial for effective public debt management, development financing, and reducing rollover risks. In Nepal, bond markets serve as a key instrument for mobilizing resources, yet their future trajectories remain under explored despite their growing role in fiscal planning. This study investigates the mathematical exploration and computational performance of time series models ARIMA, RNN, and LSTM are applied to the government bonds of Nepal. The analysis examines trends, seasonal patterns, and trajectories using descriptive statistics to capture underlying market behaviors. An optimal ARIMA order was identified to effectively capture the linear growth path, while the RNN demonstrated strong capability in learning nonlinear patterns and outperformed the other models in predictive accuracy on un seen data. In contrast, the LSTM model, constrained by the limited size of the dataset, showed weaker generalization despite achieving comparable or lower training errors. The results highlight that Nepal’s bond market is characterized by a steady trajectory in Development Bonds, uncertainty in Citizen Saving Bonds, and weak participation in Foreign Employment Bonds, with total borrowing projected to rise. These findings suggest that while ARIMA emphasizes stability, deep learning approaches reveal momentum-driven growth potential, offering complementary perspectives. The paper aims to inform policymakers by presenting insights into how bond market forecasting may strengthen long-term development financing, mitigate refinancing risks, and foster wider participation in underutilized bonds, ultimately enhancing the effectiveness of debt management in Nepal.