**2. Literature Review**

In the 1950s and 1960s, (Arrow 1964; Debreu 1959) were the first to explore optimal contracts under uncertainty, and laid the foundation for contract theory. In the late 1960s and 1970s, Gorge Akerlof, Josef Stiglitz, and Michal Spence formed the incentive theory as a branch of contract theory, and introduced the concepts of "hidden information" and "hidden actions". The asymmetric information problem under the incentive theory has been prolongedly discussed in modern contract economies. Credit rationing (Stiglitz and Weiss 1981) and information signaling (Spence 1973) were the two major branches of the discussion.

One major class of the contracting problem lies in hidden information, which is also regarded as adverse selection. It describes a situation in which one party to the contract has private information that the other does not. When the contract is made by the party that lacks private information, the uninformed party needs to screen the information possessed by the informed party. This is the so-called screening problem. If the contract is offered by the informed party, it is a signaling problem, since the informed party can signal the information they have through the type of contract offered. (Akerlof 1970) used the automobile market as an example to explain the situation in which one party has private information, and regards the second-hand automobile market as a market for "lemons", since the seller has private information about the condition of the car, and thus they have the incentive to sell cars of below-average-quality. Therefore, the entire market quality has been dragged down, but due to the asymmetric information, the buyer can only bargain according to the average price, so would only like to buy lower-quality cars, which results in above-average-quality cars exiting the market. This situation, when low-quality

products replace high-quality products, resulting in the entire market quality declining, is so-called adverse selection. In the loan market, this refers to a situation in which high-risk borrowers are usually those who are most eagerly looking for money, and most likely to obtain the loan. Thus, how to mitigate adverse selection and how to efficiently use signals to screen the borrower becomes a crucial and heated discussion topic. Credit appraisal is the application of screening in the financial market; the borrower has private information about the quality of the project and the incentives of paying back. Our research investigated the efficiency of the screening mechanism in online lending and a possible approach for improvement.

Empirical research concerning credit analysis in peer-to-peer lending can be divided into two groups. One is targeted at analyzing the trust of the lenders. This research area studies how lenders screen borrowers, or what the determinants are for the success of loan funding. The other trend investigates the borrower's repayment behavior, which indicates their creditworthiness; in other words, the potential factors that may signal the possibility of default.

From the perspective of lenders, according to (Debreu 1959), "The role of soft information in trust building: Evidence from online social lending" is representative of the literature analyzing lenders' trust. Data was used from Germany's largest P2P platform, Smava, to analyze trust-building between borrowers and lenders. The interest rate was used as a proxy for trust level. They introduced the concept of soft information as the personal information the borrower was willing to disclose. The results showed that communicating personal information increased lenders' trust, but the impact was small and limited to educational and professional information. In addition, if the borrower used statements aimed at arousing pity, they were given a higher interest rate, indicating a loss of trust. (Herzenstein et al. 2008), on the other hand, more comprehensively summarized the determinants of success in P2P lending into several groups: demographic characteristics, including gender, race, and marital status; financial strength, including credit ratings from credit bureaus, debt ratio, and house ownership; effort indicators, i.e., the effort to increase reputation, mainly through group activity and loan description; and loan decision variables, i.e., the loan features, such as amount, interest rate, and duration. Their results showed that all variables representing financial strength had a significant influence on funding success except house ownership, which was insignificant. Credit ratings from A to E were all positively related to success, except high-risk grading, but debt-to-income ratio was negatively related to success. Results for demographic characteristics showed that women were more likely to receive funding, which was opposite to expectations; marital status was not significant in the decision to grant a loan. African Americans; racial identity had a negative effect on loan funding success. The effort to include a picture had no significant influence on success, but the effort to join in group activity and give a loan description had a positive effect.

Besides these two representative works which summarized the determinants of success in funding applications, a large group of researchers examined the impact of a specific screening variable on the success of the loan application. (Barasinska and Schäfer 2014) analyzed the impact of gender on the possibility of successful funding on German P2P platform Smava; (Gonzalez and Loureiro 2014) and (Pope and Sydnor 2011) analyzed whether a profile picture would influence funding success; similarly, (Duarte et al. 2012) analyzed appearance and funding success; (Greiner and Wang 2009), (Herrero-Lopez 2009), and (Lin et al. 2013) focused on the impact of social capital on loan success; (Wang et al. 2019) led the analysis of the impact of video information on loan success. Researches in this field provided evidence of the screening determinants from the lender's perspective, but lack the comparison with the borrower's repayment behavior. This may be due to data limitation, but without this comparison we cannot diagnose the efficiency of these determinants. Looking from the lender's perspective can only provide information about the lenders' preference but cannot show whether these preferences correctly recognized the borrower's creditworthiness. Our research is based on the determinants previous studies

provided, but in addition we compared the results with the borrowers' repayment behavior to explore the real efficiency of the screening mechanism of the lenders.

From the perspective of borrowers, (Santoso et al. 2020) used data from three Indonesian P2P platforms to analyze the determinants of loan interest rates and default status. As an inconsistency in the existing literature, they also observed that factors such as age and gender have different results on three different platforms. The paper investigated the relationship of the chosen determinants with default probability and the loan interest rate. However, they did not link these two results together and further investigate the phenomenon behind and the origin of the problem. Our paper's target is to fill this gap and analyze whether borrower signals are correctly diagnosed by lenders. (Dorfleitner et al. 2016) studied the effect of soft factors derived from descriptive text on the probability of successful funding and probability of default on two European P2P lending platforms. Their results showed that typos, text length, and keywords evoking positive emotions are significantly related to funding success but have no impact on default probability. Their research provided the first evidence of linguistic factors in credit analysis; however, they only focused on linguistic factors and did not further investigate the misdiagnosis of other soft factors when comparing lenders' judgment and borrowers' real behavior.

The first paper that touched on the efficiency of the lenders' diagnosis is that of (Iyer et al. 2016) in the 2016 paper, "Screening peers softly: Inferring the quality of small borrowers", they used the advantage that they had acquired the true credit scores of the borrowers from the credit bureau while the lenders on the prosper platform only had information about the credit grading. As a predictor, they used the final interest rate collected by the borrower to assess whether the lenders on the platform would use the details available to assess the borrower's true credibility. The results showed that, within one credit category, the lenders were able to infer one-third of the variation in creditworthiness that was captured by credit scores. Their results also suggested that, on top of the traditional financial factors, non-standard "softer" information was also used in analyzing the borrower's credit risk, especially for lower credit rating borrowers. Although the paper diagnosed the fact that lenders on the platform had one-third of the ability to infer the real creditworthiness of the borrower, it also indicated that misspecification existed since only one-third had been captured which implied that two-thirds hadn't. This paper opened the first debate on whether the usage of soft information would compensate for the traditional credit analysis model and add more choice for credit model development after the 2008 financial crisis. However, this paper did not delve into the specific determinants which resulted in the misspecification. Our paper is an extension of that of (Iyer et al. 2016), whereby we provide empirical evidence for the misspecification of the lenders' screening mechanism in P2P lending.

We further compared the literature on these two trends, and found inconsistent results for the same variable in different models; for example, gender was insignificantly correlated with success in (Pötzsch and Böhme 2010) but significantly correlated with success in (Zhang et al. 2017), (Herzenstein et al. 2008) and (Pope and Sydnor 2011). At the same time, female gender was shown to be positively related to default in (Santoso et al. 2020) but negatively related to default in (Ge et al. 2017) and insignificantly related in (Pope and Sydnor 2011). Moreover, the results of (Dorfleitner et al. 2016) showed that typos, text length, and keywords evoking positive emotions were significantly related to funding success but had no impact on default probability. People who mentioned education in their loan descriptions were more likely to obtain loans (results significant), but mentioning education was shown to be insignificant in predicting default. However, in (Liao et al. 2015), people with higher degrees of education had a lower probability of default (significant) but were not more likely to get funding (insignificant). In (Freedman and Jin 2008), mentioning education in loan descriptions had an insignificant influence on funding success but people who did so were significantly less likely to default. Mentioning car ownership was not significantly related to success but was significantly and positively related to default. In addition, mentioning family was significantly and positively related to

success but also significantly and positively related to default. Due to these inconsistencies, we doubt whether investors can truly diagnose the credit signals given by borrowers. If there are misdiagnoses, which factors resulted in these mismatches?

Thus, we come up with our hypothesis:

**Hypothesis 1:** *Investors on the P2P platform can correctly diagnose the credit signals the borrower provide and efficiently screen out low credit borrowers;*

**Hypothesis 2:** *Investors can more efficiently diagnose hard financially related signals than soft socially related signals.*
