Towards a Deeper Understanding of Yield Curve Movements

Abstract

From a modeling perspective, daily movements in risk-free yield curves deserve further exploration. In this paper, I study daily yield curve changes for three major liquid sovereign yield curves, namely US, UK and Germany. I focus on two main themes:

(1) characterizing the statistical distribution of the multivariate vector of daily yield curve changes, and

(2) exploring efficiency of deep learning based dimensionality reduction approaches for the daily yield curve changes.

Specifically, I am able to reject the null hypothesis that daily yield curve changes are normally distributed. I also show that a multivariate student-t distribution with 2 degrees of freedom or a gaussian mixture model are more accurate at capturing the leptokurtic distribution of daily yield curve changes. In addition, on the dimensionality reduction front, I show that a deep neural network auto-encoder outperforms a PCA based linear decomposition approach.

Keywords: Deep Learning, Auto-encoders, Dimensionality Reduction, PCA, Yield Curve