Assessing the Determinants of Corporate Risk-Taking Using Machine Learning Algorithms

Abstract

Given that risk-taking is an essential channel for companies to obtain high returns and realize value enhancement, the goal of this study is to holistically explore the determinants of corporate risk-taking using various machine learning algorithms. Based on the data from Chinese listed companies between 2010 and 2019, we document that the adaptive boosting (AdaBoost) model makes better predictions of corporate risk-taking. We further visualize the importance and influence of the firm basic characteristics, firm performance, and chief executive officer (CEO) characteristics and discover that in the AdaBoost model, the firm basic characteristics, and performance factors, such as the firm’s fixed asset investments, size, and return on equity, are important in predicting corporate risk-taking, while CEO characteristics are less important. Finally, the role of variables in corporate risk-taking varies among large and small enterprises. Overall, our findings deepen the comprehension of what drives corporate risk-taking and provide a potential way for real-world firms seeking to adjust their risk-taking level.

1. Introduction

Corporate risk-taking—a byproduct of a firm’s profit-seeking business operations—is an essential part of investment decisions [1,2]. The literature has provided some insights into the influence factors of corporate risk-taking from the aspects of corporate governance [3,4], ownership structure [5,6], CEO gender [7,8], and CEO overconfidence [9,10]. However, there is still no consensus on the determinants of corporate risk-taking. This study aims to comprehensively explore the key forces behind corporate risk-taking behavior, guiding enterprises to make better decisions on investment projects with high risk and high return.

Higher risk-taking implies that firms tend to invest in projects with higher risk but positive expected net present values [11], which is beneficial to obtaining higher returns and achieving value enhancement [12]. However, high-risk investment projects often come with great uncertainty, especially in the context of the coronavirus disease 2019 (COVID-19) pandemic, which has increased economic instability. Under such complex and changeable circumstances, how firms choose an appropriate risk-taking level based on their business situation is an important question not only for Chinese companies but also for companies in other transitional economies. If enterprises can identify the drivers of risk-taking, it can help them adjust risk-taking, achieving profit acquisition and value creation. Therefore, it is critical to examine the determinants of corporate risk-taking.

Some studies suggest that corporate ownership concentration increases the risk it takes [13]. While Lee et al. [14] found a U-shaped relationship between them. Adams and Funk [15] discovered that there are positive effects of board gender diversity on corporate risk-taking, contrary to the traditional idea that women are more risk-averse [16]. Other studies have examined the influence of firm size [17] and CEO rights [18] on corporate risk-taking. Nevertheless, these studies mostly consider several influencing factors and the conclusions are inconsistent. We cannot determine which factors are the decisive factors of corporate risk-taking based on the findings of previous studies. Moreover, when many factors are considered simultaneously, which factors are most important and whether and how various factors affect corporate risk-taking remain to be further discussed. This paper intends to make up for this research gap.

The goal of this study is to construct a comprehensive corporate risk-taking analytical framework that scrutinizes numerous cross-sectional factors and documents their effect and importance to help firms know the drivers of risk-taking and perform risk management. We mainly discuss the determinants of risk-taking from the perspectives of firm basic characteristics, firm performance, and CEO characteristics. On the one hand, corporate risk-taking requires continuous resource input, which means that it is affected by firm basic characteristics and performance factors, such as firm size, fixed asset investment, income, and return on assets. On the other hand, corporate risk-taking behavior is accompanied by high uncertainty. As the essential decision-maker of a company, the CEO’s personal and cognitive characteristics, will affect corporate risky investment decisions. In addition, we employ a more flexible model to explain their linkages. Machine learning (ML), as a practical new tool for creating theories from data, has captured the interest of academics [19]. ML methods can handle high-dimensional data and do not impose a specified structure on the potential relationships between variables and predicted outcomes, revealing complex patterns between them [20,21]. This flexibility allows us to address complex prediction problems and provides a more accurate analysis of corporate risk-taking.

Based on the above discussion, we construct a sample of Chinese listed companies during the period from 2010 to 2019 and use ML methods to attempt to answer the following research questions (RQs):

RQ1.

Which ML method performs better in predicting corporate risk-taking?

RQ2.

Which factors of firm basic characteristics, performance, and CEO characteristics are more important in predicting corporate risk-taking and their effect?

RQ3.

Do the contribution and effect of influence factors differ by firm type?

The findings and contributions of our study are as follows. First, we contribute to the literature on corporate risk-taking. Previous studies have investigated the effect of certain firm or CEO characteristics on corporate risk-taking separately [22,23], making it not intuitive to judge the decisive factors of corporate risk-taking based on previous results. Our research supplements them by incorporating various firm basic characteristics, firm performance, and CEO characteristics into a unified analytical framework and identifying the determinants of corporate risk-taking. The results indicate that the most important factors in predicting corporate risk-taking are the firm’s fixed asset investments, size, and return on equity based on the AdaBoost model, which performs best in predicting corporate risk-taking. Second, we add to the current literature by expanding on the application of ML algorithms [20,24]. The inconsistent results regarding corporate risk-taking in the existing literature may be due to their failure to accurately identify the driving factors of risk-taking [13,14]. In recent years, ML methods have proven to be exceptionally good at making predictions using complex and large-scale datasets [25], which can help us effectively predict corporate risk-taking and explore the key forces behind it. In this paper, we find that firm basic characteristics and performance factors are more useful in predicting risk-taking, while CEO characteristics play a lesser role. Finally, our study provides suggestions for China and other transitional economies to adjust corporate risk-taking. It is of great significance for promoting the stability and sustainable development of the real economy. We also provide insight into managers and investors into the fact that firm basic characteristics and performance deserve more attention than CEO characteristics in risk management and investment decisions. For example, to take on more risky investments and improve enterprise value, managers should increase fixed asset investments. Additionally, the different roles of risk-taking determinants in small and large enterprises remind managers to formulate accurate management measures according to firm types.

The remainder of this paper is as follows. Section 2 reviews the related literature. The data and methodology we used in this study are shown in Section 3. Section 4 shows the results and Section 5 concludes and discusses.

2. Literature Review

Risk-taking, as an important way to improve corporate performance and future development, has triggered a series of studies that explore its influence factors [8,22,26]. In this paper, we mainly review the existing literature from the perspective of the firm and CEO characteristics.

Regarding firm characteristics, several researchers highlight the impact of firm ownership and find that firms with state ownership have the protection of the government and they are more likely to take on risky investment projects [27]. However, Boubakri et al. [5] used the data of 381 newly privatized companies and found that state ownership was negatively correlated with firm risk-taking as government policies that seek to maximize social stability could lead firms to pursue less risky investment projects. Moreover, based on a sample of American financial institutions between 2002 and 2012, Bhagat et al. [17] found that firm size was positively related to risk-taking before and during the crisis period, but negatively after. Faccio et al. [28] found that small and young companies were more willing to take risks. Shahzad et al. [23] examined the different life cycle stages of risk-taking and found that enterprises in the growth and maturity stages had higher risk-taking levels, while those in the entry and recession stages had lower levels. In addition, higher leverage comes with an increased likelihood of financial distress, reducing the risk-taking capacity of firms and managers [29]. Moreover, firm performance also has important influences on risk-taking. Traditional investment theories believe that high profitability generally corresponds to high risks. While some researches reveal a link between firm profitability and risk-taking that is negative [30,31], firms with low profitability may tend to choose more risky investment projects to improve their financial situation. Do et al. [32] argued that companies with poor interim performance took more risks than companies with better interim performance.

More recently, a growing body of work has identified the CEO’s role in corporate risk-taking as the key decision-maker for the company whose personality characteristics and cognitive patterns will shape the firm’s strategic decisions. Some studies look at the influence of CEOs’ age on risk-taking and come to conflicting conclusions. On the one hand, scholars discover that older CEOs have more management experience and accumulated rich social network resources in their careers, which can help CEOs to obtain important resources and market information, enhancing corporate risk-taking [33,34]. Nevertheless, there are also studies showing that individuals’ risk aversion rises with age; compared to young CEOs, senior CEOs seem to be more risk-averse and likely to make less risky investment decisions [35,36]. Regarding CEOs’ gender, scholars generally assume that female CEOs would be more risk-averse than male CEOs [16]. However, it has been shown that women are not always risk-averse [37,38]. For instance, Safiullah et al. [39] discovered that Spanish companies with more female directors had higher risk-taking after the implementation of the “Gender Equality Act”. Ingersoll et al. [22] used a sample of S&P 500 companies from 2009 to 2019 and found that companies with more female executives take more financial risks than those with male executives. In addition, the level of corporate risk-taking is also related to the personality characteristics of CEOs. Narcissistic CEOs with a proclivity for hubris are more willing to take risks [40]. Moreover, CEOs’ power [41], reputation [26], educational experience [7,35], professional experience [42], and social network [43,44] have also been found to affect corporate risk-taking.

3. Research Methodology

3.1. Sample and Data

We gathered a sample of Chinese listed firms from 2010 to 2019 to investigate the impact of firm basic characteristics, firm performance, and CEO characteristics on risk-taking. We obtained all of our data from the China Stock Market and Accounting Research (CSMAR) database. The following observations were removed: (1) Financial companies, as the financial industry does not belong to the real economy and the structure of its financial statements differs from that of non-financial companies; (2) Designated as special treatment (i.e., ST and *ST) firms; (3) Missing data for the main variables. Finally, our sample, which consists of 15,774 firm-year observations for 2329 companies, was finally obtained. To ensure that extreme values do not influence our results, we winsorized all continuous variables at the 1% and 99% levels.

3.2. Variable Definitions

Risk-taking. Following the study of Jiang and Chen [45], we define corporate risk-taking (Risk_taking) as the three-year standard deviation of adjusted return on assets (AdjROA), which is the percentage of earnings before interest and taxes (EBIT) to total assets (Asset) adjusted for the firm’s industry and yearly averages. For example, the rolling standard deviation of AdjROA from t to t + 2 is used to measure the value of Risk-taking at time t. The calculations are as follows:

$${AdjROA}_{i,j,t}=\frac{{EBTI}_{i,j.t}}{{Asset}_{i,j.t}}-\frac{1}{N_{j,t}}\left({\displaystyle \sum }_{n=1}^{N_{j.t}}\frac{{EBTI}_{i,j,t}}{{Asset}_{i,j,t}}\right)$$

(1)

$${Risk\_taking}_{i,t}=\sqrt{\frac{1}{T-1}{\displaystyle \sum }_{t=1}^T{\left({AdjROA}_{i,j,t}-\frac{1}{T}{\displaystyle \sum }_{t=1}^T{AdjROA}_{i,j,t}\right)}^2}$$

(2)

where i, j, and t denote the firm, industry, and year, respectively. N stands for the number of firms in industry j in year t. In Formula (2), T denotes the window period used in this paper, where T = 3, meaning that we use the three-year window and rolling standard deviation of AdjROA to measure Risk_taking.

We construct a series of variables based on prior studies [5,8,16,39,45], which influence corporate risk-taking from three aspects: firm basic characteristics, firm performance, and CEO characteristics. The detailed definitions of the variables used in this study are shown in Appendix A.

Firm basic characteristics. In this paper, firm basic characteristics refer to the fundamental information variables of a firm, including firm age (Firm_age), initial public offering age (IPO_age), size (Size), ownership (SOE), fixed asset investments (Fixed_assets), sales growth (Sale_growth), debt-to-equity ratio (DER), and leverage (Leverage). Specifically, Firm_age refers to the years between the founding date and the observations [16], while IPO_age refers to the years after a firm’s IPO date [46]. Size is calculated by the natural logarithm of total assets [39]. The value of the dummy variable SOE is 1 if the government is the actual controller of the company, and 0 if otherwise [45]. Fixed_assets is the proportion of fixed assets to total assets [47]. Sale_growth refers to the annual growth of sales from time t − 1 to t [45]. The definitions of DER and Leverage are the percentage of total debt to total equity and total debt to total assets, respectively [30,46].

Firm performance. To measure firm performance, we include firm return on assets (ROA), return on equity (ROE), net profit margin (Net_margin), Tobin’s Q value (Tobinq), market-to-book ratio (MTB), and sales (Sales). Specifically, ROA and ROE are the proportions of net income to total assets and total equity, respectively [16,25], Net_margin is measured as net profit divided by total revenue [25], Tobinq is determined by the ratio of the market value to book value of total assets [32], MTB is a firm’s equity market value to its book value [23], and Sales is the natural logarithm of total sales [46].

CEO characteristics. We consider 15 CEO characteristics that may influence corporate risk-taking, which are CEO gender (CEO_gender), age (CEO_age), duality (CEO_duality), education (CEO_education), tenure (CEO_tenure), board experience (CEO_boardexperience), role as an outside directorate (CEO_outsidedirectorates), production experience (CEO_production), R&D experience (CEO_RD), design experience (CEO_design), human resource experience (CEO_HRM), administration experience (CEO_administration), finance experience (CEO_finance), accounting experience (CEO_accounting), and legal experience (CEO_legal). Specifically, CEO_gender is a dummy variable that has a value of 1 for female CEOs and 0 for male CEOs [16]; CEO_age is measured by the age of the CEO [16]. CEO_duality is assessed by a binary variable that is 1 if the CEO simultaneously works as the chairman and 0 is otherwise [48]. CEO_education is defined by whether the CEO has a postgraduate degree. If the CEO has a postgraduate degree, the value is 1; otherwise, it is 0 [35]. The natural logarithm of the CEO’s tenure is used to define CEO_tenure [48]. A dummy variable named CEO_boardexperience equals 1 if the CEO also serves as a director and 0 if otherwise [35]. The variable CEO_outsidedirectorates is the total number of outside directorates a CEO holds or has held on other firms’ boards [49]. Moreover, we have defined a series of dummy variables based on the CEOs’ professional experience: CEO_production, CEO_RD, CEO_design, CEO_HRM, CEO_administration, CEO_finance, CEO_accounting, and CEO_legal. For example, CEO_finance takes the value of 1 if the CEO has a career experience in finance and 0 if otherwise [35]. Finally, the descriptive statistics of variables used in this study are shown in Appendix B.

3.3. Methodology

We use ML methods to predict corporate risk-taking. Given that there is no agreement regarding the best performing algorithm for predicting risk-taking, 11 ML methods are employed in this study. In terms of linear models, we include linear regression (LinearRegression), ridge regression (Ridge), the least absolute shrinkage and selection operator (Lasso), elastic net (ElasticNet), and Bayesian ridge regression (BayesianRidge). Moreover, we consider ensemble models, all of which are important ML algorithms, including extremely randomized trees (ExtraTree), random forest (RandomForest), extreme gradient boosting (XGBoost), adaptive boosting (AdaBoost), gradient boosted regression trees (GradientBoost), and light gradient boosting machine (LightGBM).

We randomly divide our data into two subsets (i.e., training and test datasets) following an 80%/20% split, which is a commonly used split criterion in previous studies [20]. The training dataset is used to optimize the hyperparameters of each prediction model and identify the optimum combinations of variables to predict risk-taking. The test dataset is used to compare the performance of each prediction to the observed out-of-sample data.

3.4. Model Predictive Performance

We assess each model’s predictive performance of corporate risk-taking by its out-of-sample predictive R². The out-of-sample predictive R² is calculated as follows:

$$R_{oos}^2=1-\frac{{{\displaystyle \sum }}_{(i,t)\mathit{\in }{\rm T}}{{(Risk\_taking}_{i,t}-{\widehat{Risk\_taking}}_{i,t})}^2}{{{\displaystyle \sum }}_{(i,t)\mathit{\in }{\rm T}}{Risk\_taking}_{i,t}^2}$$

(3)

where $R_{oos}^2$ is the out-of-sample predictive R², Τ represents a series of predictor variables, ${Risk\_taking}_{i,t}$ is firm i’s real risk-taking in period t, and ${\widehat{Risk\_taking}}_{i,t}$ is firm i’s predicted risk-taking using a specific model on the test datasets.

3.5. SHapley Additive exPlanations (SHAP) Method

To determine how the variables affect corporate risk-taking, whether positively, negatively, or via other more complex correlations. We use the SHAP method, which uses SHAP values to assess the contribution and influence of each variable to corporate risk-taking, thus improving the ML model’s interpretability [20,50]. According to Lundberg and Lee [50], SHAP values are calculated as follows:

$${SHAP value}_i={\displaystyle \sum }_{S\mathit{⊑}N\backslash \left(i\right)}\frac{\left|S\right|\mathit{!}\left(M-\left|S\right|-1\right)\mathit{!}}{M\mathit{!}}[f_x(S\cup \left\{i\right\}{)-f}_x\left(S\right)]$$

(4)

where i is an input feature, N is the set of all input features excluding i, and M is their total number. S is a subset of N, whose number of input features is |S|. $f_x\left(S\right)$ is the predicted value using the set of feature S on data instance x. $f_x(S\cup \left\{i\right\})$ represents the predicted value using the set of feature S plus feature i.

4. Results

4.1. Model Predictive Performance Comparison

We start by examining the prediction performance of ML models in the total sample, then divide the total sample into large and small enterprises subsamples, with the top 30% in the firm size category being considered as large enterprises and all others as small enterprises [24], to further verify the prediction performance.

Table 1. Out-of-sample predictive R² in percentage.

Model	Total Sample	Small Enterprises	Large Enterprises
LinearRegression	6.5826	3.4296	5.0433
Ridge	6.4993	3.7123	5.2760
Lasso	6.5198	3.9324	5.0868
ElasticNet	6.3337	3.7324	5.3546
BayesianRidge	6.3589	3.4596	5.5579
ExtraTree	19.2507	10.0737	17.8735
RandomForest	19.2699	10.3005	17.5704
XGBoost	13.8818	−0.1823	7.1658
AdaBoost	13.7653	17.7049	13.2793
GradientBoost	14.3386	12.6980	11.1379
LightGBM	17.3584	12.5887	18.4043

Notes: this table demonstrates the out-of-sample predictive R² in percentage of machine learning models.

are the results of the model’s predictive performance.

Total sample. As demonstrated in Table 1, in the linear models, we can find that the LinearRegression model, which is the simplest model, has an out-of-sample predictive R² of 6.58%, indicating that even the basic model has some degree of predictive validity. The out-of-sample predictive R² for Ridge, Lasso, ElasticNet, and BayesianRidge do not change substantially when compared to the LinearRegression model. In addition, the prediction performances of the ensemble models are generally better than the linear models. For example, the out-of-sample predictive R² of the AdaBoost model is 13.77%, which outperforms all the linear models we used. We have also discovered that the RandomForest model performs well in predicting risk-taking.

Small and large enterprises. Next, we focus on the model predictive performance for small enterprises (i.e., the bottom 30% of firms by total assets) and large enterprises (i.e., the top 70% of firms by total assets). Table 1 demonstrates that in the small enterprise subsample, the AdaBoost model has the best predictive performance with an out-of-sample predictive R² of 17.70%. While in the large enterprise subsample, LightGBM is the better predictive model and the AdaBoost model also has a good performance.

Although we have used out-of-sample R² to select predictive models of corporate risk-taking, some problems remain as some models have very close out-of-sample R², and the predictive models show inconsistent performance across subsamples. Specifically, in the total sample, the RandomForest model has the best predictive performance. In the small enterprise subsample, the best predictive model is the AdaBoost model, while it is the LightGBM model in the large enterprise subsample. Thus, we cannot identify the optimal predictive model of corporate risk-taking. To address this problem, we use the conditional superiority predictive ability (CSPA) test proposed by Li et al. [51] for alternative model selection.

The CSPA test assumes that the conditional expected loss in the baseline model is not greater than that of the competing model, and thus, compares the predictive performance of the two models. The results are shown in

Table 2. Unconditional superior predicative ability for risk-taking.

Model		One-Versus-One CSPA Tests
		Total	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)	(10)	(11)
(1)	LinearRegression	5 *						R	R		R	R	R
(2)	Ridge	6 *	R					R	R		R	R	R
(3)	Lasso	6 *	R					R	R		R	R	R
(4)	ElasticNet	8 *	R	R	R			R	R		R	R	R
(5)	BayesianRidge	9 *	R	R	R	R		R	R		R	R	R
(6)	ExtraTree	1 *									R
(7)	RandomForest	2 *									R	R
(8)	XGBoost	4 *							R		R	R	R
(9)	AdaBoost	0
(10)	GradientBoost	1 *									R
(11)	LightGBM	1 *									R

Notes: This table shows the results of the CSPA test. R indicates rejection of the CSPA test hypothesis (i.e., the benchmark model outperforms the competing models). For example, in row (2) and column (1) of Table 2, Ridge is the benchmark model and LinearRegression is the competing model for the CSPA test, which results in rejection, thus indicating that the predictive performance of LinearRegression is better than that of Ridge. * indicates significance at the 5% level.

. The CSPA test results show that the AdaBoost model is the best predictive model as it has the lowest rejection frequency when conditioned on industry average risk-taking. The RandomForest and LightGBM models perform well, but when compared to the AdaBoost model, all of them are rejected; thus, we believe that that the AdaBoost model outperforms them in predicting corporate risk-taking.

The AdaBoost algorithm is one of the most popular and extensively researched algorithms developed by Freund and Schapire [52]. It is the first practical boosting algorithm, and its main idea is that if samples are correctly classified, the weight assigned to them in the next round will be lowered. On the contrary, samples that are misclassified will obtain higher weights [21]. In addition, the AdaBoost model does not need to filter features and can avoid overfitting problems. It also has high prediction accuracy and performs optimal learning (theoretically) even in noisy environments [53].

4.2. Which Variables Matter?

After determining that the AdaBoost model has the best prediction performance, we next explore the importance of variables that cover firm basic characteristics, firm performance, and CEO characteristics in the total sample. To gauge the importance of each variable and compare it with other variables, we use the total average reduction in predictive R² when the value of a certain variable is set to zero within every training dataset [24]. Observing

Table 3. Out-of-sample predictive R² in percentage.

Model	Feature	R2
AdaBoost	Firm basic characteristics only	0.4471
	Firm performance only	10.8095
	CEO characteristics only	−273.4586
	Firm basic characteristics plus CEO characteristics	3.5707
	Firm performance plus CEO characteristics	9.5419
	Firm basic characteristics plus firm performance	12.2544
	Firm performance plus firm performance characteristics plus CEO characteristics	13.1284

Notes: This table displays the out-of-sample predictive R² of the AdaBoost model in the total sample by considering different variables.

, we notice that using firm basic characteristics, firm performance, and CEO characteristics at the same time can have a better out-of-sample R². This demonstrates that adding these variables to the corporate risk-taking prediction model helps improve its predictive performance.

Firm basic characteristics. 

Table 4. Relative importance of variables in the AdaBoost model.

Firm Basic Characteristics	R2	CEO Characteristics	R2
Fixed_assets	9.4461	CEO_age	4.3035
Size	8.1247	CEO_tenure	3.0405
Sale_growth	6.1957	CEO_outsidedirectorates	2.6666
Firm_age	5.6597	CEO_administration	0.8726
IPO_age	4.6782	CEO_design	0.7835
Leverage	4.0704	CEO_education	0.7217
DER	4.0615	CEO_HRM	0.6788
SOE	0.5048	CEO_RD	0.6556
Firm Performance	R2	CEO_duality	0.6286
Net_margin	7.6391	CEO_finance	0.5219
ROA	7.1779	CEO_gender	0.4150
ROE	7.0045	CEO_production	0.3391
Sales	6.9890	CEO_accounting	0.2377
MTB	6.5973	CEO_boardexperience	0.0598
Tobinq	5.8831	CEO_legal	0.0429

Notes: This table displays the relative importance of variables based on the out-of-sample predictive R², and the values are expressed in percentages. The total of the feature importance of all predictors, including firm basic characteristics, firm performance, and CEO characteristics, is normalized to 1 in a particular model.

reports the importance of variables. We find that in the AdaBoost model, the most important variable is Fixed_assets. Firm fixed asset investments reflect the abundance of a firm’s internal resources and the potential for development. Additionally, firms with higher levels of fixed assets have more ability to obtain financing to support the firm’s risker activities. Moreover, Size, Sales_growth, and Firm_age are also important in predicting corporate risk-taking, while IPO_age, Leverage, DER, and SOE are less important. The results of the variables’ importance in other ML methods are shown in Appendix C.

In addition, we use box plots for robustness tests. Figure 1 shows the box plots that aggregate the importance of each firm basic characteristic variable across all ML models. Box plots are not affected by outliers and provide an accurate and stable depiction of the discrete distribution of data. We also discover that Fixed_assets is the most influential firm basic characteristic variable for predicting corporate risk-taking.

Firm performance. We consider firm performance variables—ROA, ROE, Net_margin, Tobinq, MTB, and Sales—that may affect corporate risk-taking. Table 4 shows the importance of our firm performance variables in the total sample. In the AdaBoost model, the most important firm performance variable is Net_margin, which reflects a firm’s profitability. We also find that ROA and ROE are also important predictor variables and their importance is less different from that of Net_margin.

Figure 2 demonstrates box plots that aggregate the importance of each firm performance variable across all models. The results show that ROA, ROE, and Net_margin are the top three most influential firm performance variables. Although this result is slightly different from that in the AdaBoost model, it still demonstrates the importance of Net_margin for predicting corporate risk-taking.

CEO characteristics. Table 4 displays the importance of the CEO’s personal characteristics, educational experience, and professional experience variables on corporate risk-taking based on out-of-sample predictive R². In the AdaBoost model, we find that not all CEO characteristics are equally important in predicting corporate risk-taking. The most important variable is CEO_age, followed by CEO_tenure and CEO_outsidedirectorates, while CEO_accounting, CEO_boardexperience, and CEO_legal have less influence.

Figure 3 demonstrates the aggregate importance of each CEO characteristics variable in all models and displays that CEO_age is also an important variable in predicting corporate risk-taking.

4.3. The Variable Effects on Corporate Risk-taking

We use the SHAP approach to visualize the overall contribution and effects (positive or negative) of the firm basic characteristics, performance, and CEO characteristics on corporate risk-taking using the AdaBoost model in the total sample and subsamples (i.e., small and large enterprises).

Table 5. Top ten important variables of corporate risk-taking.

Total Sample	Category	Effect
Fixed_assets	Firm basic characteristics	Positive
Size	Firm basic characteristics	Negative
ROE	Firm performance	Negative
ROA	Firm performance	Negative
Firm_age	Firm basic characteristics	Negative
Sales	Firm performance	Negative
MTB	Firm performance	Negative
Sale_growth	Firm basic characteristics	Positive
Net_margin	Firm performance	Negative
DER	Firm basic characteristics	Positive
Small enterprises	Category	Effect
Fixed_assets	Firm basic characteristics	Positive
ROA	Firm performance	Negative
IPO_age	Firm basic characteristics	Positive
Size	Firm basic characteristics	Negative
MTB	Firm performance	Negative
ROE	Firm performance	Negative
Sales	Firm performance	Negative
Leverage	Firm basic characteristics	Positive
Net_margin	Firm performance	Positive
Sale_growth	Firm basic characteristics	Positive
Large enterprises	Category	Effect
ROE	Firm performance	Negative
Firm_age	Firm basic characteristics	Negative
Fixed_assets	Firm basic characteristics	Negative
Sales	Firm performance	Negative
Net_margin	Firm performance	Negative
Size	Firm basic characteristics	Negative
ROA	Firm performance	Negative
IPO_age	Firm basic characteristics	Negative
Leverage	Firm basic characteristics	Positive
Sale_growth	Firm basic characteristics	Positive

Notes: This table shows the importance and influence of the top ten variables on the prediction of corporate risk-taking based on SHAP value in the total sample and subsamples (small and large enterprises).

summarizes the top ten variables, which enables us to pinpoint the motivating factors behind corporate risk-taking in the total sample and subsamples (small and large enterprises) by computing the variable’s SHAP value. The detailed overall importance and effect of all variables on risk-raking are shown in Figure 4. The horizontal axis represents the SHAP value of variables. The larger the SHAP value, the more important this variable is for predicting corporate risk-taking. The vertical axis represents each variable, and they are ordered from highest to lowest based on their importance.

4.3.1. Total Sample

In Table 5, we find that in the total sample, Fixed_assets, Size, ROE, ROA, and Firm_age are the top five most important variables in predicting corporate risk-taking, while CEO_education, CEO_RD, CEO_accounting, CEO_boardexperience, and CEO_legal have less importance. By investigating the categories these variables belong to, we can find that although CEOs are the key decision-makers of firms whose characteristics, risk preferences, and cognitive patterns can influence a firm’s investment decisions and behavior of taking risks [8,39], the importance of CEO characteristics in predicting corporate risk-taking is less than the firm basic characteristics and performance, meaning that the firm characteristics are the main drivers behind corporate risk-taking behavior. Next, we specifically analyze the effect of variables in each category.

Figure 4 also reflects how each variable affects corporate risk-taking. Variables in red (blue) indicate that they have a positive (negative) effect on corporate risk-taking. In the firm basic characteristics, Fixed_assets is the most important variable; its red color means that Fixed_assets is positive with corporate risk-taking. Firms need fixed asset investments to support their performance and investment decision, which has an important positive impact on corporate risk-taking. The color of the “Size” bar is blue, indicating that Size has a negative effect on corporate risk-taking; these results coincide with earlier studies that demonstrate that large enterprises are less risk-taking, while smaller enterprises are more aggressive and tend to make riskier investments [30,47]. Firm_age is negatively associated with corporate risk-taking; that is, younger firms are more risk-seeking, which is similar to the findings of Faccio et al. [28]. Contrarily, IPO_age is positive; the reason may be that IPOs can provide companies with funds to undertake new project [54]. Other firm basic characteristics variables used in this paper all have positive effects on corporate risk-taking. In particular, firms with state ownership tend to take risks more, which is consistent with the results of the studies of Dong et al. [27] and Zhu and Yang [6], providing an explanation for the conflicting findings in the literature. However, it should be noted that SOE is not important in predicting corporate risk-taking.

In the firm performance category, the color of all the bars is blue, including Net_margin, ROA, ROE, Sales, MTB, and Tobinq, which means that these variables all have negative effects on corporate risk-taking. These results imply that the better the firm performance, the lower the risk-taking, agreeing with the results of Khaw et al. [30] and Do et al. [32].

With regard to the CEO characteristics, almost all variables have positive effects on corporate risk-taking except CEO_administration, CEO_education, CEO_RD, and CEO_production, which have negative effects. For example, the impact of CEO_age on corporate risk-taking is positive; that is, older CEOs tend to have more experience and capabilities, which can help them to obtain resources and market information, enhancing corporate risk-taking [33], while younger CEOs may prefer conservative investment strategies to avoid losses caused by decision-making mistakes [31]. In terms of CEO_outsidedirectorates, if a CEO also serves as the outside directorate in other firms, the increased CEO power and social networks encourage an optimistic risk attitude in them and greater discretion on corporate investment decisions, improving the firm’s risk behavior [41,43]. For CEO_tenure, Wang and Poutziouris [55] explored the link between the tenure of CEOs and corporate risk-taking. They believe that with the increase in tenure, managers become more familiar with the situation of the industry and increase their risk behavior. CEO_gender, a CEO characteristics variable that has been paid more attention to in previous studies, has a positive impact on corporate risk-taking, thus supporting that female CEOs are not always risk-averse [22,39].

4.3.2. Large and Small Enterprises

In the large enterprise subsample, ROE, Firm_age, Fixed_assets, Sales, and Net_margin are important predictive variables, while Fixed_assets, ROA, IPO_age, Size, and MTB are important predictive variables in the small enterprise subsample. Moreover, Table 5 demonstrates that Fixed_assets is an important variable for both small and large enterprises. Firm_age is a critical variable in large enterprises, while less important in small enterprises. Additionally, MTB is vital in predicting risk-taking for small enterprises but less so for large enterprises. In addition, the predictive performance of CEO characteristics on risk-taking is relatively weak among large and small enterprises weak compared to other variables.

Then, we compare the differences in the variable’s impact on risk-taking for large and small enterprises. Fixed_assets and IPO_age have positive effects on small enterprises, while they have negative effects on large enterprises. In other words, increasing the firm’s fixed asset investments and promoting listing firms as soon as possible are helpful in enhancing the level of risk-taking, but these mainly affect small enterprises. If the same measures are taken for large enterprises, their risk-taking level will be reduced. Furthermore, Net_margin has a negative effect on risk-taking for large enterprises and a positive influence on small enterprises. One possible reason for this result is that smaller and lower net profit firms are more likely to enhance firm performance through high risk-taking.

4.3.3. Robustness Tests

To guarantee the correctness of our findings, we used the bootstrap method, which randomly selects several samples from the total dataset and forms a new training dataset, to further test the overall importance and effect of variables on corporate risk-taking. The results are shown in

Table 6. The SHAP values of the top ten important variables of corporate risk-taking.

Total Sample	Category	SHAP Value	Effects	Sign
Fixed_assets	Firm basic characteristics	1.0553	0.3130	Positive
Size	Firm basic characteristics	1.0076	−0.7559	Negative
ROE	Firm performance	0.9919	−0.4278	Negative
ROA	Firm performance	0.8996	−0.4496	Negative
Firm_age	Firm basic characteristics	0.8583	−0.6920	Negative
Sales	Firm performance	0.6254	−0.5165	Negative
Sale_growth	Firm basic characteristics	0.5989	0.4305	Positive
Net_margin	Firm performance	0.5852	−0.2070	Negative
MTB	Firm performance	0.5302	−0.3582	Negative
Leverage	Firm basic characteristics	0.4827	0.7540	Positive
Small enterprises	Category	SHAP value	Effects	Sign
Fixed_assets	Firm basic characteristics	1.9806	0.3480	Positive
ROA	Firm performance	1.3139	−0.2492	Negative
ROE	Firm performance	0.9385	−0.7008	Negative
Size	Firm basic characteristics	0.9341	−0.7123	Negative
IPO_age	Firm basic characteristics	0.9220	0.8805	Positive
MTB	Firm performance	0.8456	−0.7019	Negative
Sales	Firm performance	0.8190	−0.3223	Negative
Net_margin	Firm performance	0.6956	0.2678	Positive
Sale_growth	Firm basic characteristics	0.6665	0.1323	Positive
Leverage	Firm basic characteristics	0.6576	0.8503	Positive
Large enterprises	Category	SHAP value	Effects	Sign
ROE	Firm performance	0.9759	−0.2915	Negative
Firm_age	Firm basic characteristics	0.8756	−0.5913	Negative
Net_margin	Firm performance	0.7770	−0.4371	Negative
Fixed_assets	Firm basic characteristics	0.7712	−0.0643	Negative
Sales	Firm performance	0.7642	−0.4005	Negative
DER	Firm basic characteristics	0.6779	0.8604	Positive
Size	Firm basic characteristics	0.6475	−0.5721	Negative
ROA	Firm performance	0.5957	−0.5168	Negative
IPO_age	Firm basic characteristics	0.5781	−0.5108	Negative
Sale_growth	Firm performance	0.5474	0.4613	Positive

Notes: this table shows the importance and influence of the top ten variables on the prediction of corporate risk-taking based on SHAP values in the total sample and subsamples (small and large enterprises).

. In the total sample, we can find that Fixed_assets, Size, and ROE are also important variables in predicting corporate risk-taking. Additionally, the importance of CEO characteristics in predicting corporate risk-taking is less than the firm basic characteristics and performance. From the effects of variables, in the total sample, we can also discover that Fixed_assets have a positive effect and Size and ROE have a negative effect on corporate risk-taking. In small and large enterprises, Fixed_assets is also a vital predictive variable, and has positive effects on the risk-taking of small enterprises, while it has a negative effect on large enterprises. Overall, our main results are robust.

5. Discussion and Conclusions

This study uses ML methods to explore the determinants of corporate risk-taking from firm basic characteristics, firm performance, and CEO characteristics using Chinese listed companies during the period from 2010 to 2019. Our results show that, first, the AdaBoost model performs better at predicting corporate risk-taking than the other ML algorithms, and considering firm basic characteristics, firm performance, and CEO characteristics at the same time is helpful in improve ML models’ predictive performance. Second, using the SHAP method, we have found that in the total sample, the most important predictors for risk-taking are the firm’s fixed asset investments, size, and ROE, implying these variables have better predictive power for corporate risk-taking. From the perspective of the impact of these variables on corporate risk-taking, we have discovered that the firm’s fixed asset investments positively affect corporate risk-taking, while the firm’s size and ROE have a negative effect. Moreover, CEO characteristics, such as CEO accounting experience, board experience, and legal experience, play a lesser role in predicting firm risk-taking. Finally, the variables’ contribution and effect on corporate risk-taking are heterogeneous in large and small firms based on their SHAP values. The firm’s fixed asset investments are important in predicting risk-taking for both small and large enterprises. However, the firm’s age is only important for large enterprises’ risk-taking, while the firm’s market-to-book ratio is only vital for small enterprises’ risk-taking. For the variable’s impact, our results demonstrate that the firm’s fixed asset investments, net profit, and IPO age have positive effects on risk-taking for small enterprises, but negative on large enterprises. This study provides significant theoretical and practical implications for understanding corporate risk-taking.

5.1. Theoretical Implications

Our research contributes to the growing literature on corporate risk-taking [8,22,23,29]. The conclusions of the literature are inconsistent. Specifically, corporate risk-taking is affected by many factors, but it remains unknown how important each factor is, and whether and how various factors affect corporate risk-taking when many factors are considered simultaneously. This study contributes to filling this gap by exploring the drivers behind corporate risk-taking from the perspectives of firm basic characteristics, performance, and CEO characteristics, and analyzing their complex relationship with corporate risk-taking. Specifically, we find that CEO characteristics are relatively less important for predicting corporate risk-taking when compared to firm basic characteristics and performance, which provide novel insights into the determinants of corporate risk-taking.

Second, this paper complements the literature on applying ML methods to risk prediction [24,25]. ML algorithms, as a new technique, perform well in classification and prediction, especially with big data. Using ML methods to predict corporate risk-taking and the SHAP approach to interpret the ML model, our research analyzes the contribution and impact of various predictors on corporate risk-taking, providing essential and novel insights into the drivers of corporate risk-taking. Overall, the findings show that the firm’s fixed asset investments, size, and ROE have better predictive power for corporate risk-taking, whereas CEO accounting experience, board experience, and legal experience play lesser roles.

Finally, this paper further explores the differences in the importance and influence of various factors on corporate risk-taking among large and small enterprises, improving the analysis of risk-taking from the perspective of firm size. We find that firm size negatively affects risk-taking, supporting the view that large enterprises are less risk-taking, while smaller enterprises are more aggressive and tend to make riskier investments [30,47]. Moreover, the determinants of corporate risk-taking vary by firm size. Firm fixed asset investments are important in predicting risk-taking for both small and large enterprises. However, firm age is only critical for the risk-taking of large enterprises, whereas firm market-to-book ratio is only critical for small enterprises.

5.2. Practical Implications

This research has significant practical implications in helping China and other transitional economies increase their corporate risk-taking and promoting stability and the sustainable development of the real economy. As scholars have emphasized, taking risks can promote a firm’s capital accumulation and improve its market competitiveness [11,12]. However, high-risk investment projects come with a great level of uncertainty; the failure of investment could cause heavy losses for firms or even bankruptcy [56]. Therefore, it is necessary to explore the driving factors of corporate risk-taking and help companies, especially those in transitional economies, choose a reasonable risk-taking level in complex and changeable environments. This study emphasizes that firm basic characteristics and profitability are critical in predicting risk-taking, while CEO characteristics have less of an impact. These findings can help managers realize that they should pay more attention to the effect of firm characteristics, such as the firm’s fixed asset investment, size, and ROE on risk-taking, rather than the effects of CEO characteristics.

Moreover, our findings can provide valuable information for risk management in different companies. We find that there are differences in the contributions and effects of various factors on corporate risk-taking for large and small enterprises. Managers should choose the appropriate level of risk-taking according to their scale. For example, we emphasize that increasing firm fixed asset investments, net profit, and promoting firms to go public can help improve the risk-taking ability of small enterprises, but negatively impact large enterprises.

5.3. Limitations and Recommendations

Although this study reveals important findings, it has several limitations. The sample used in this paper is China’s listed firms; the results are applicable in that region and may not be generalized. Future research can consider including diverse data from other economies to solve this problem. Furthermore, corporate risk-taking may be industry-specific. We have analyzed the overall risk-taking of Chinese listed firms and do not distinguish the heterogeneity of risk-taking in different industries. Future research can consider using ML methods to predict corporate risk-taking in specific industries. Finally, there may be additional omitted variables that are associated with both the predictors and outcome variables. Future research can consider more factors in the framework of corporate risk-taking based on this paper.