Bond return predictability: Macro factors and machine learning methods

• Published date: 01 November 2024
• Author: Ying Jiang,Xiaoquan Liu,Yirong Liu,Fumin Zhu
• Date: 01 November 2024
• DOI: http://doi.org/10.1111/eufm.12483

DOI: 10.1111/eufm.12483

ORIGINAL ARTICLE

Bond return predictability: Macro factors and

machine learning methods

Ying Jiang

1

|Xiaoquan Liu

1

|Yirong Liu

1

|Fumin Zhu

2

1

Nottingham University Business School

China, University of Nottingham Ningbo,

Ningbo, China

2

College of Economics, Shenzhen

University, Shenzhen, China

Correspondence

Yirong Liu, Nottingham University

Business School China, University of

Nottingham Ningbo China, 199 Taikang

East Rd, 315100 Ningbo, China.

Email: yirong.liu@nottingham.edu.cn

Funding information

National Natural Science Foundation of

China, Grant/Award Number: 72071132;

Humanities and Social Science Fund of

Ministry of Education of China,

Grant/Award Number: 23YJA790034

Abstract

We investigate the impact of macroeconomic variables

on bond risk premia prediction via machine learning

techniques. On the basis of Chinese treasury bonds

from March 2006 to December 2022, we show that

adding macroeconomic factors improves bond return

forecasts and generates higher economic benefits to

investors. This is achieved when the nonlinear

relationship between macroeconomic variables and

bond returns is modelled via machine learning

methods. Furthermore, the importance of macro-

economic determinants changes along the yield curve.

Our study sheds new light on the information

contained in macroeconomic variables for treasury

bond valuation and highlights the importance of

utilizing appropriate machine learning methods.

KEYWORDS

Chinese bond market, government bond returns forecasting,

machine learning, unspanned macroeconomic information, yield

term structure

JEL CLASSIFICATION

G12, G17

Eur Financ Manag. 2024;30:2596–2627.wileyonlinelibrary.com/journal/eufm2596

|

© 2024 John Wiley & Sons Ltd.

EUROPEAN

FINANCIAL MANAGEMENT

We are grateful to John A. Doukas (the editor) and two anonymous referees for their insightful comments and

suggestions which have significantly improved the quality of this paper. Financial support from the National Natural

Science Foundation of China (Grant No. 72071132) and Humanities and Social Science Fund of the Ministry of

Education of China (Grant No. 23YJA790034) is gratefully acknowledged.

1|INTRODUCTION

The failure of the expectations hypothesis (EH) (Fisher, 1896), which postulates constant term

premia, has been well established in the literature. Since then, the literature has developed in

two directions. Cochrane and Piazzesi (2005) show that a linear combination of forward rates

can explain a substantial portion of treasury bond return variation; because forward rates

represent the interest rate for a future bond commitment, they are informative for bond return

prediction (Thornton & Valente, 2012). The second, and more recent, strand of the literature

highlights the information content of macroeconomic variables because they are important

determinants of the yield curve (Gargano et al., 2019; Joslin et al., 2014; Ludvigson & Ng, 2009,

among others).

Motivated by these studies, we aim to examine the predictability of treasury bond risk

premia on the Chinese market. Specifically, we investigate whether macroeconomic factors

carry incremental predictive information for bond risk premia beyond the yield curve.

Furthermore, we explore whether machine learning methods improve forecasting accuracy for

bond excess returns.

From a theoretical perspective, macroeconomic variables impact bond risk premia through

several channels. According to Brandt and Wang (2003), bond excess returns can be explained

by the unexpected inflation rate, which increases investors' risk aversion and raises bond risk

premia. Wachter (2006) highlights the impact of the difference between current and past

consumption levels. This difference affects investors' borrowing behaviour and interest rates,

which in turn influences bond risk premia. Jermann (2013) demonstrates that investment in

equipment and structures is the key driver of bond risk premia based on a production‐based

model. The investment‐to‐capital ratio impacts the volatility of short‐term interest rates,

thereby affecting bond excess returns. These theories offer potential explanations for the

relationship between macroeconomic information and bond risk premia observed in data.

However, empirical evidence on the predictive power of macroeconomic factors unspanned

by the yield curve remains inconclusive. On the one hand, economic growth, for instance,

influences bond risk premia by shaping investors' expectations about future economic

conditions, thereby impacting their expected returns for holding bonds (Cooper &

Priestley, 2009). Monetary policy affects the term structure of interest rates, potentially

influencing treasury bond risk premia (Fan et al., 2012). Additionally, the literature shows that

other macroeconomic variables contain incremental predictive information for bond excess

returns that is not captured by the yield curve, such as fiscal policy, industrial production ratio

(IP), consumption–wealth ratio, and global macroeconomic factors (Bianchi et al., 2021;

Bretscher et al., 2020; Costantini & Sousa, 2020; Gargano et al., 2019; Joslin et al., 2014;

Wright, 2012; Zhao et al., 2021). On the other hand, several studies suggest that macroeconomic

indicators have limited explanatory power when using different statistical methods (Bauer &

Hamilton, 2018), real‐time data (Ghysels et al., 2018), or in an emerging market (Zhang

et al., 2022). Motivated by these findings, our study aims to predict excess returns on Chinese

treasury bonds by incorporating a comprehensive set of macroeconomic factors, thus

contributing to this ongoing debate in the literature.

In terms of methodology, prior studies largely rely on standard linear predictive models,

such as ordinary least squares (OLS) regression (Cochrane & Piazzesi, 2005; Fama &

French, 2012, among others). In recent years, machine learning techniques have gained

increased popularity as a viable alternative to linear forecasting models in the field of asset

pricing (Akyildirim et al., 2023; Kelly et al., 2019; Leippold et al., 2022), and they outperform

JIANG ET AL.EUROPEAN

FINANCIAL MANAGEMENT

|

2597

the traditional linear methods for two reasons. First, when many factors are utilized in the

prediction process, machine learning algorithms perform much better than traditional OLS

regression in variable selection and dimensionality reduction (Gu et al., 2020; Hoang &

Wiegratz, 2023; Rapach & Zhou, 2021). Second, prior studies reveal that the association

between explanatory variables and bond returns is nonlinear (Bianchi et al., 2021; Feldhütter

et al., 2018), and machine learning methods are well suited to capturing complicated, nonlinear

relationships due to their flexibility in accommodating rich functional forms. Given these

considerations, we employ machine learning approaches in our empirical analysis.

Our empirical tests focus on the Chinese bond market mainly because it has several unique

features. First, before 2015, the interest rates on deposits and loans of all maturities were set by

the Peoples Bank of China and known as the official rates (Hu & Wang, 2022), which may

render bond excess returns less predictable in China. However, interest rate liberalization since

then has been shown to impact the monetary policy transmission, suggesting an effective pass‐

through from the monetary policy rate to the bond market rate (Li & Liu, 2019; Li et al., 2021).

With the progress of interest rate marketization, it is interesting to explore the predictability of

treasury bond returns in China using macroeconomic variables.

Second, in terms of market structure, in addition to the exchange and over‐the‐counter (OTC)

bond markets in the US, China has an interbank market, which is larger than its exchange‐based

and OTC markets in terms of market size and primarily driven by institutional investors, with

commercial banks being the dominant participants (Livingston et al., 2018; Zhang et al., 2022).

However, most commercial banks are state‐owned, making them prone to government

intervention and less sensitive to real macroeconomic activities (Hu & Wang, 2022). Thus,

investigating the extent to which treasury bond risk premia can be forecasted by macroeconomic

factors in China contributes to a more comprehensive understanding of bond return predictability.

Our sample period is from March 2006 to December 2022, with two sets of factors. One set

uses yield‐curve data, following Cochrane and Piazzesi (2005), and we incorporate the 1‐year

yield and nine forward rates of bonds with maturities ranging from 2 to 10 years. The other set

exploits information from hundreds of macroeconomic factors, as in Ludvigson and Ng (2009),

and we consider an extensive collection of 102 macroeconomic variables. Methodologically,

building on the literature (Bali et al., 2022; Bianchi et al., 2021; Fan et al., 2022; Huang &

Shi, 2023), we adopt various machine learning methods in the prediction exercises, together

with the OLS regression as a comparison. These include dimension‐reduction techniques such

as principal component regression (PCR) and partial least squares (PLS); penalized linear

regression methods such as the ridge regression (Ridge), least absolute shrinkage and selection

operator (Lasso), and elastic net (ENet); tree‐based models such as the random forest (RF) and

gradient‐boosting regression trees (GBRTs); and the extended neural networks (NNs) used by

Bianchi et al. (2021).

Our empirical analyses yield a number of interesting findings. First, we show that, when we

employ nonlinear machine learning methods, macroeconomic variables substantially improve

the accuracy of bond risk premia forecasts compared with those based only on yields. In

particular, when we incorporate macroeconomic information in the prediction, the out‐of‐

sample Roos

2values of the best‐performing model are 30% higher on average than those of yield‐

based forecasts across maturities. Furthermore, we find that the increased predictive accuracy

by macroeconomic factors translates into greater economic gains for investors. This observation

is different from Zhang et al. (2022), which show that macroeconomic factors have no extra

information for bond risk premia prediction. Our results suggest that unspanned macro-

economic indicators are useful in explaining treasury bond returns in the Chinese market.

2598

|

EUROPEAN

FINANCIAL MANAGEMENT

JIANG ET AL.