|
|
|
| Research on an Early Warning Model of Corporate Bond Default and its Economic Mechanism Based on Machine Learning |
| JIANG Fuwei, LIN Yihao, MA Tian
|
School of Finance, Central University of Finance and Economics;
School of Economics, Minzu University of China |
|
|
|
Abstract In recent years, the bond market has played an important role in serving the real economy, optimizing resource allocation, and supporting macroeconomic policy regulation. However, since China terminated rigid payments in 2014, bond defaults have occurred frequently. In this context, the identification of bond default risk has become a new and key issue for the capital market and economic development. At the same time, financial technology (fintech) is becoming an important method to enhance the prevention and control of financial risks. In this context, this paper proposes the use of fintech, such as big data and machine learning, to develop an early warning model for bond default risk that fits the current context. This paper systematically explores the performance of big data and machine learning models in predicting bond default risk.
In terms of data, in this paper, we examine the general enterprise bonds, enterprise bonds, medium-term notes, and commercial papers issued by China's A-share listed companies in the interbank and exchange markets. We construct a macro and micro mixed big dataset with a total of 1,245 variables, including 15 macroeconomic indicators, 70 enterprise characteristic variables, and 12 bond characteristic variables. Specifically, this paper adds macroeconomic indicators that reflect the willingness and ability of local governments to rescue enterprises subject to bond defaults. In addition, we construct enterprise characteristic variables based on six categories of indicators, including valuation and growth, investment, profit, inertia, transaction friction, and intangible assets. Furthermore, we cross-multiply macro and micro indicators to construct interactive indicators. In terms of model construction, we select 10 machine learning models, including PCA, PLS, Ridge, LASSO, ENet, SVR, RF, GBDT, XGBoost, and AdaBoost. Based on the above models, we examine an early warning model of bond default risk based on big data and machine learning, and explore the economic mechanism behind machine learning.
The empirical results show that a machine learning model can predict China's bond default risk better than the classical Altman, Merton, and credit rating models. Moreover, nonlinear machine learning models perform better. The above conclusions remain valid under the modified Diebold-Mariano statistic test. We find that machine learning can distinguish differences in bond risk more effectively than the portfolio analysis method of empirical asset pricing. In addition, we determine that the advantages of machine learning models over benchmark models increase over time, and that bond default risk prediction gradually requires an increasingly complex modeling process. Finally, this paper establishes an “improved credit rating” based on machine learning spread predictions and finds that increasing rating discrimination can improve the effectiveness of ratings.
This paper further explores the economic mechanism behind the model. First, the heterogeneity analysis finds that the machine learning model has stronger predictive ability for bonds with low ratings, long issuance maturities, and high coupon rates, issuances by non-state-owned enterprises, and bonds issued in the interbank market. Moreover, its predictive ability is stronger during periods of higher economic policy uncertainty than during other periods. Second, the variable importance analysis reveals that indicators related to valuation and growth, investment, profit, intangible assets, and bond characteristics provide good warning signals of default risk in the context of machine learning, but the role of the inertia and transaction friction indicators (other than stock liquidity) is relatively insignificant. Third, machine learning models can achieve accurate predictions through default bond identification, short-term signal identification (bond trading volume), and long-term feature identification (financing constraints, internal control), and their sensitivity to “negative information” is better than that of classical models.
The contributions of this paper are as follows. First, this paper promotes the application of big data and machine learning in finance. From a theoretical perspective, we point out that classical models lack dynamic time-varying parameters and variable diversity, and therefore proceed to empirically test the effectiveness of using machine learning models. Second, this paper expands the research perspective on bond default risk. We use the continuous variable “credit spread” to quantify bond risk, and enrich bond default risk identification from a high-dimensional perspective. Third, this paper deepens the discussion on credit rating in China's bond market and finds that improving rating discrimination can improve rating effectiveness. Based on our research, future studies could further expand the prediction indicators for bond default risk from additional perspectives, such as text analysis, and use more advanced machine learning and deep learning models to improve the accuracy of predictions of bond default risk.
|
|
Received: 03 November 2022
Published: 16 July 2024
|
|
|
|
|
|
|
|
