Deep Learning Approaches in Interest Rate Forecasting

• Main Authors: Bata, Halle Megan L, Victoria, Mark Jayson A, Alvarez, Wynnona Chezska B, De Lara-Tuprio, Elvira, Allado, Armin Paul D
• Format: text
• Published: Archīum Ateneo 2024
• Subjects: Pitch; Data science; Financial economics; Deep learning; Artificial neural networks; Machine learning; Learning models; Applied Mathematics; Data Science; Economics; Macroeconomics; Physical Sciences and Mathematics; Social and Behavioral Sciences
• Online Access: https://archium.ateneo.edu/mathematics-faculty-pubs/288; https://doi.org/10.1063/5.0231027
• Institution: Ateneo De Manila University

Interest rates are among the most important macroeconomic factors considered by financial market participants, both government and private entities, when making investment and policy decisions. Pricing and valuation of financial securities rely heavily on both current and future interest rates, among others. A reliable forecast of interest rates is a requisite to sound management of exposure to different types of risk including market risk, liquidity risk, and credit risk. Determining an optimal debt strategy for the government also requires reliable forecast of interest rates. These are among the reasons why the area of forecasting the term structure of interest rates has created a huge literature. Several theoretical models, such as factor models and stochastic models, have long been the forecasting tools used by big financial institutions. The advent of machine learning has recently provided good alternatives to theoretical models. This paper aims to add to the literature on the use of machine learning to forecast short-term interest rates by applying Artificial Neural Networks (ANNs). In particular, we propose two deep learning models – Multilayer Perceptrons (MLP) and Vanilla Generative Adversarial Network (VGAN) – to forecast the 1-month, 3-month, 6-month, and 1-year Philippine (PH) BVAL rates. As a base case, 16 economic variables were considered as inputs. Considering the principle of parsimony, we used the statistical methods of Backward Stepwise Regression and Variance Inflation Factor (VIF) Testing to also create models with smaller number of input variables. Thus, we applied MLP and VGAN to each of these three sets of input variables. We determined the best model for each interest rate tenor based on the graphical simulation and comparison of the actual and predicted values alongside the validation scores of the model. In general, both MLP and VGAN produced reliable forecasts of the interest rates.