Recurrent Neural Networks in Diverse Market Conditions

Abstract

This thesis investigates the predictive performance of Recurrent Neural Networks (RNNs) in forecasting excess return in zero-coupon bonds. We evaluate their performance using data from the U.S. and German bond markets. The study assesses predictive accuracy and the economic value in different market conditions.

Technically, we implement various models, including linear regressions, Random Forest Regressors, Principal Component Regression (PCR), Partial Least Squares (PLS), and Recurrent Neural Networks (RNNs). Forward rates and macroeconomic variables are integrated to enhance predictive accuracy, and their impact is analyzed across different market conditions, including the COVID-19 pandemic.

Our analysis shows that RNNs achieved statistically significant improvements R-Squared out-of-sample over the benchmark. For longer maturities, we found improvements of up to 35%, much of it as a result of out-performance in 2020 and 2021. These forecasting accuracy gains translated into significant economic value for the U.S. market. Although promising prediction results were also observed for German bonds(bunds), they did not yield the same economic utility.

This study highlights the promise of RNNs for financial forecasting, but also emphasizes the challenge of making models that generalize to the characteristics of multiple markets.

This thesis investigates the predictive performance of Recurrent Neural Networks (RNNs) in forecasting excess return in zero-coupon bonds. We evaluate their performance using data from the U.S. and German bond markets. The study assesses predictive accuracy and the economic value in different market conditions.

Technically, we implement various models, including linear regressions, Random Forest Regressors, Principal Component Regression (PCR), Partial Least Squares (PLS), and Recurrent Neural Networks (RNNs). Forward rates and macroeconomic variables are integrated to enhance predictive accuracy, and their impact is analyzed across different market conditions, including the COVID-19 pandemic.

Our analysis shows that RNNs achieved statistically significant improvements R-Squared out-of-sample over the benchmark. For longer maturities, we found improvements of up to 35%, much of it as a result of out-performance in 2020 and 2021. These forecasting accuracy gains translated into significant economic value for the U.S. market. Although promising prediction results were also observed for German bonds(bunds), they did not yield the same economic utility.

This study highlights the promise of RNNs for financial forecasting, but also emphasizes the challenge of making models that generalize to the characteristics of multiple markets.