CAUSALITY BETWEEN PEER-TO-PEER LENDING AND BANK LENDING IN CHINA: EVIDENCE FROM A PANEL DATA APPROACH

1. Introduction

Why are the peer-to-peer lending (P2PL) and bank lending (BL) markets interesting? As a new financial intermediary between borrowers and lenders, P2P lending refers to unsecured lending between lenders and borrowers through online platforms without the intermediation of financial institutions. The P2PL system is favored by borrowers and investors due to its declared low lending interest rates and fast liquidity. Over the past years with the spread of Internet technology and the wide use of large data bases, online P2PL as a financial innovation implication has quickly developed worldwide. Online P2P lending has recently emerged as a useful financing alternative where individuals can borrow and lend money directly through an online trading platform without the help of institutional intermediaries such as banks. Many researchers have studied the factors affecting the P2P lending platform. For example, Chen et al. (2013) found that there was significant gender discrimination in the P2P lending market in China, showing that female borrowers were less likely to be funded than male ones, but their default rates were lower. What is more, Lu and Zhang (2018) proposed that platform strength, profitability, risk control, liquidity and transparency could predict the probability of the platform becoming problematic.

P2PL is perceived as increasing overall financial activities, and the increase in these activities is generally considered desirable; namely, the positive impact of P2PL on financial activities is frequently described. Also, P2PL is recognized as having a positive effect on short-term, mid-term or long-term financial and economic growth through different channels. First, P2PL is a new form of loan originating in the credit market. This type of loan market is designed to complement traditional BL to meet the small loan needs of individuals and small and medium enterprises. Second, the P2PL platform can quickly assess and assign risk grades through the use of information technology instead of relying on delegated monitoring with banks as intermediaries. Finally, lenders have the opportunity to handle the financial and personal information provided by the borrowers and directly offer a loan that meets their investment criteria. As a result, lenders can receive the money they need on P2PL for a short period of time which is relative to BL. However, it would be interesting and useful to determine whether the relationship between BL and P2PL platforms is complementary or competitive.

The purpose of this paper is to investigate the causal relationship between P2PL and BL in China’s eight major regions using the regional causality test developed by Kónya (2006) to determine the dynamic and causal relationship between P2PL and BL. This procedure will undoubtedly allow the specific effects of regions to be more readily uncovered. The paper attempts to examine whether there is a causal relationship between growth in P2PL and BL using a bootstrap multivariate panel Granger causality test through a sample of China’s eight major regions over the period from 2014M01 to 2019M12. The rest of the paper is organized as follows. Section 2 briefly reviews the related literature. Section 3 provides an overview of data collection and methodology. Section 4 presents the empirical results. Section 5 shows the robustness check from the study, while the conclusions and research limitation are given in Section 6.

2. Literature Review

The development of P2P lending has attracted the attention of the academic community. It has long been recognized that P2PL has an impact on BL (e.g. Cumming and Johan, 2016; Cumming and Zhang, 2016; Lin et al., 2013). Several researchers (Culkin et al., 2016; Guo et al., 2016; Klievink et al., 2017) have investigated the relationship between P2PL and BL. Based on the past literature, P2PL focuses on the behavior of traders and the risk of trading in this market. As for the risk of trading, the researchers focus on the factors that affect credit risk in the P2PL market and P2PL risk regulations. Lenders can evaluate one-third of the credit risk using both soft and hard information about borrowers (Iyer et al., 2009). In the study of the behavior of traders, the researchers mainly discuss herding behaviors. Because of the risk of information asymmetry, traders exhibit herding behaviors in P2PL (Lee and Lee, 2012). For example, Chaffee and Rapp (2012) described P2PL as risky and that it was therefore necessary to build an evolving regulatory regime for an evolving industry. Herzenstein et al. (2008) found that strategic herding behavior contributed to P2PL bidders individually and collectively. Using a multinomial logit market share model, Lee and Lee (2012) find strong evidence of herding and its diminishing marginal effects as bidding advances. Lin (2009) provides an introduction to P2P lending and presents a preliminary discussion of the credit and risk of online loans. Within the context of the impacts of social networks on behavior, which was a hot topic at that time, Lin et al. (2013) investigated the effects of social relations on P2P lending.

More recently, Emekter et al. (2015) point out that the credit classification of borrowers plays an important role in loan default and that charging higher interest rates to high-risk borrowers cannot compensate for the risk of loss. Chen et al. (2018) and Xu et al. (2015) expand their horizons to China and explore the impact of borrowers’ social capital on loan outcomes. Using a South Korean platform as an example, Lee and Lee (2012) examine the herding behavior of P2P lenders and use a longer period of study. Zhang and Chen (2017) identify the existence of both rational and irrational herding in the P2P market. Several studies (Clemons et al., 2017; Li et al., 2014; Yum et al., 2012) evaluate the decisions of lenders under different levels of information transparency through big data analysis. Feller et al. (2017) find that there is more complexity and heterogeneity in participant behavior. Guo et al. (2016) design an instance-based credit risk assessment model that can effectively improve investment performance compared to existing methods in P2PL. In terms of the risk of trading, Guo et al. (2016) focus on the factors that affect credit risk in the P2PL market and risk regulation. Lenders can evaluate one-third of the credit risk using both soft and hard information about borrowers (Iyer et al., 2009). A summary of the literature review is presented in Table 1.

3. Data Collection and Methodology

The data for China’s eight major regions were collected from Wang Dai Zhi Jia.¹ Wang Dai Zhi Jia is currently the most authoritative portal site of the P2P lending industry in China.

In this study, the dataset employed included the period from 2014M01 to 2019M12 for China’s eight major regions (i.e. Beijing, Shanghai, Jiangsu, Zhejiang, Shandong, Hubei, Guangdong and Sichuan). The sample period was decided purely by data availability on the measure for the P2PL activities. This paper eliminates the effect of inflation on P2PL and BL, which are transformed into natural logarithms prior to the econometric analysis. The used data included BL to measure the performance of the BL activities and P2PL to measure the performance of the P2PL activities for each region. Besides, this work used control variable for other important variables which, if omitted, would cause the estimated coefficients to be biased. Therefore, this study adopted the P2PL interest rates as control variable. Consequently, this work organizes the panel data for the empirical purpose. Prior to the multivariate panel Granger causality analysis results, the work applied three conventional unit root tests, namely, ADF, PP and KPSS tests. These results are reported in Table 2.

3.1. Testing cross-sectional dependence and slope homogeneity

The first issue in panel causality analysis focuses on accounting for possible cross-sectional dependence across regions. Cross-sectional dependency is the most important issue while dealing with panel Granger causality across China’s eight major regions. Since there are vast regions with various levels of economic growth and an increasing amount of integration within these regions, a shock occurs within one region and influences other regions, such as endowments and geographical differences, which are felt around the regions. Accordingly, it is necessary to conduct a series of cross-sectional dependence tests in which the panel data causality between P2PL and BL in China’s eight major regions is examined in this work. The other issue in panel data analysis is to decide whether the slope coefficients are homogeneous. Causality running from one variable to another due to the imposition of a joint restriction for the whole panel is the strong null hypothesis (Granger, 2003). The panel data analysis concentrates on deciding whether the slope coefficients are homogeneous. Interested readers can refer to the cross-sectional dependence and slope homogeneous tests proposed by Wu and Wu (2018) in detail.

3.2. Multivariate panel Granger causality analysis

The equation system for a multivariate panel Granger causality analysis includes two sets of equations that can be extended in this study. This paper follows Wu and Wu’s (2020) equation systems for multivariate panel Granger causality analysis including two sets of equations that can be written as:

where ${\beta }_1$, ${\delta }_1$, ${\gamma }_1$ and ${\lambda }_1$ are $Nxl{y}_1$, $Nxl{x}_1$ and $Nxl{w}_1$ matrices, and y, x, w and z are $l{y}_{1}x1$, $l{x}_{1}x1$ and $l{w}_{1}x1$ matrices, respectively.

$$\begin{array}{l}\left(\begin{array}{c}\hfill x_{1,t}\hfill \\ \hfill ⋮\hfill \\ \hfill x_{N,t}\hfill \end{array}\right)=\left(\begin{array}{c}\hfill {\alpha }_{2,1}\hfill \\ \hfill ⋮\hfill \\ \hfill {\alpha }_{2,N}\hfill \end{array}\right)+\left(\begin{array}{ccc}\hfill {\beta }_{2,1,1}\hfill & \hfill \dots \hfill & \hfill {\beta }_{2,1,l{y}_2}\hfill \\ \hfill ⋮\hfill & \hfill \ddots \hfill & \hfill ⋮\hfill \\ \hfill {\beta }_{2,N,1}\hfill & \hfill \dots \hfill & \hfill {\beta }_{2,N,l{y}_2}\hfill \end{array}\right)\left(\begin{array}{c}\hfill y_{1,t-1}\hfill \\ \hfill ⋮\hfill \\ \hfill y_{N,t-l{y}_2}\hfill \end{array}\right)+\left(\begin{array}{ccc}\hfill {\delta }_{2,1,1}\hfill & \hfill \dots \hfill & \hfill {\delta }_{2,1,l{x}_2}\hfill \\ \hfill ⋮\hfill & \hfill \ddots \hfill & \hfill ⋮\hfill \\ \hfill {\delta }_{2,N,1}\hfill & \hfill \dots \hfill & \hfill {\delta }_{2,N,l{x}_2}\hfill \end{array}\right)\\ \times \left(\begin{array}{c}\hfill x_{1,t-1}\hfill \\ \hfill ⋮\hfill \\ \hfill x_{N,t-l{x}_2}\hfill \end{array}\right)+\left(\begin{array}{ccc}\hfill {\gamma }_{2,1,1}\hfill & \hfill \dots \hfill & \hfill {\gamma }_{2,1,l{w}_1}\hfill \\ \hfill ⋮\hfill & \hfill \ddots \hfill & \hfill ⋮\hfill \\ \hfill {\gamma }_{2,N,1}\hfill & \hfill \dots \hfill & \hfill {\gamma }_{2,N,l{w}_1}\hfill \end{array}\right)\left(\begin{array}{c}\hfill w_{1,t-1}\hfill \\ \hfill ⋮\hfill \\ \hfill w_{N,t-l{w}_1}\hfill \end{array}\right)+\left(\begin{array}{c}\hfill {\epsilon }_{2,1,t}\hfill \\ \hfill ⋮\hfill \\ \hfill {\epsilon }_{2,N,t}\hfill \end{array}\right),\end{array}$$(2)

where ${\beta }_2$, ${\delta }_2$, ${\gamma }_2$ and ${\lambda }_1$ are $Nxl{y}_2$, $Nxl{x}_2$ and $Nxl{w}_1$ matrices, and $y$, $x$ and $w$ are $l{y}_{2}x1$, $l{x}_{2}x1$ and $l{w}_{1}x1$ matrices, respectively.

In the equation systems (1) and (2), $y$ refers to an indicator of P2PL, $x$ is referred to as BL and $w$ refers to the P2PL interest rate as a control variable. $N$ is the number of panel members, $t$ is the time period ($t=1,\dots ,T)$, and $l$ is the lag length. The bootstrap critical values are obtained from 10,000 replications. The selection of the optimal lag structure is determined by minimizing the Schwarz Bayesian Criterion from one to four lags (Kónya, 2006).

4. Empirical Results

The existence of cross-sectional dependence is shown in Table 3. In this study, four different tests (i.e. LM, CD_lm, CD and LM_adj) were conducted. It is clear that the null of no cross-sectional dependency across the regions is strongly rejected for both with and without control variable at conventional significance levels, implying that the seemingly unrelated regression (SUR) method is more appropriate than the region-by-region ordinary least square (OLS) estimation (Zellner, 1962). This finding implies that a shock occurring in one of these regions seems to have been transmitted to the other regions. Table 3 also reports the results from two slope homogeneity tests ($\tilde{\mathrm{\Delta }}$ and ${\tilde{\mathrm{\Delta }}}_{\mathrm{\text{adj}}}$). Both tests reject the null hypothesis of slope homogeneity both with and without control variable, supporting region-specific heterogeneity.

The presence of cross-sectional dependency and heterogeneity across these regions provides evidence on the suitability of the multivariate panel Granger causality approach. The results of the multivariate bootstrap panel Granger causality analysis are reported in Tables 4 and 5. Table 5 shows that one-way Granger causality running from BL to P2PL can be found in Shanghai, Jiangsu and Shandong, respectively. The results therefore support evidence for the BL leading hypothesis. To have a clear picture, the different results for P2PL and BL of the sample are displayed in Table 6. A summary of Granger causality between P2PL and BL is made. Also, the results of comparisons for China’s eight major regions (i.e. without control variable) are presented in Table 7.

5. Robustness Check

To reduce the omitted variable bias in the analysis, the P2PL interest rate as control variable is added in this paper. The results from the bootstrap multivariate panel Granger causality analysis are reported in Tables 8 and 9. The results indicate that a feedback exists between P2PL and BL for Shandong; one-way Granger causality runs from P2PL to BL for both Jiangsu and Hubei, and one-way Granger causality runs from BL to P2PL for two regions, namely, Zhejiang and Shanghai. Compared with previous results without control variable, the result shows that the regression coefficients and symbols of the main research variables are relatively stable, and the significance is not changed significantly. This conclusion indicates that the robustness check results in this paper are robust.

There are four points in this paper. First, the empirical results of this study indicate that in the regions such as Jiangsu and Hubei, P2PL stimulates BL, and the BL growth thus depends on P2PL, implying that negative P2PL shocks and P2PL reduction policies may depress BL.

Second, the relationship between BL and P2PL is found in the regions such as Zhejiang and Shanghai, indicating that BL increases the demand for P2PL and then leads to the development of P2PL in these regions. The results provide evidence for the BL leading hypothesis for these regions. Some banks begin to learn from P2PL platforms to develop their business using Internet technologies. However, with the rapid expansion of the scales of P2PL balances, the importance and convenience of P2PL without endorsement or security have been recognized by more and more people, especially for those who value their time. The result is that the potential customers in some banks choose to receive loans from P2PL markets, revealing the competitive relationship between P2PL platforms and banks.

Third, the interaction causal relationship between P2PL and BL is found in the region such as Shandong. The P2PL and BL are endogenous, displaying that these two factors mutually influence each other, and that this reinforcement may have important implications for the conduct of BL or P2PL development policies in the region. The feedback hypothesis argues that the causal relationship between P2PL and BL expansion appears bi-directional, implying that a push in both regions is beneficial. Recognizing the causal relationship between P2PL and BL is of great importance.

Finally, the study findings for the remaining three regions (i.e. Guangdong, Beijing and Sichuan) seem to support the neutrality hypothesis, indicating that neither P2PL nor BL is sensitive to the other. The neutrality between P2PL and BL suggests that P2PL does not exert an adverse impact on BL, and that P2PL is not affected by BL. The neutrality between P2PL and BL is attributed to a relatively small contribution of P2PL to overall output. The empirical result of this study shows a summary of Granger causality and comparisons for China’s eight major regions (see Tables 10 and 11).

6. Conclusions

This paper studied the impact of regional difference in China on the relationship between the P2PL and BL activities. The empirical result shows that the effect of cross-sectional dependence in China’s eight regions on the P2PL and BL activities is significant. The lenders are more likely to lend money to the borrowers who are from the region with higher economic development levels, more traditional financial institutions and higher educational development levels. However, the impact of regional differences on the default rate in the P2P lending is insignificant (Chen et al., 2020a). The empirical results of this study provide evidence for the P2PL leading hypothesis in the regions such as Jiangsu and Hubei, while the BL leading hypothesis relationship supports evidence for the regions like Zhejiang and Shanghai. In addition, the interaction causal relationship between P2PL and BL is found in the region such as Shandong. The empirical findings of this study provide important policy implications for these regions.

Acknowledgment

This study was supported by Department of Education of Guangdong Province (Grant Number 2018WTSCX214).

Notes

¹ Wang Dai Zhi Jia is the first and the largest third-party online lending information platform in China. Here is the website for the data available: http://www.wdzi.com.