Professional Documents
Culture Documents
1 s2.0 S1042443123000690 Main
1 s2.0 S1042443123000690 Main
A R T I C L E I N F O A B S T R A C T
JEL codes: This paper analyzes the performance of peer-to-peer investments, the potential benefits of their
G21 information processing and the investor returns, based on the entire portfolio of the Estonian
G23 platform Bondora. We found that the platform’s scoring model relies on different default prob
G32
abilities across countries and is weak at predicting default within countries. Alternative infor
Keywords: mation could improve the models, but our analysis could not confirm this benefit of the platform.
Alternative finance
The average internal rate of return on closed transactions was –4.17%, and 42% of the loans end
Peer-to-peer lending
with a negative IRR. We concluded that P2P borrowers in the European market are mainly high-
Information asymmetry
Credit rating risk, bank ineligible clients, accepting even loan-sharking level interest rates, which excludes
Scoring model altruistic motives of investors. Even so, investors are not compensated for the credit risk.
1. Introduction
Rapid technological changes in recent decades have contributed to financial disintermediation trends. Alternative financing
providers have emerged or started to increase their market share by promising cheaper services and better financial inclusion through
the exclusion of intermediary institutions (Polasik et al., 2020).
There are still many questions about the relevance, potential benefits and future of direct lending through an online platform. While
the removal of the intermediary layer—and, hence, the regulatory obligations on banks—may be cost-effective and help absorb
underbanked customers, it is not evident whether investors are compensated for the high credit and liquidity risk or whether their
altruistic motives are needed to sustain the business. On the borrower side, it is also unclear whether better access to finance for
subprime customers is in their interest or an access contributes to higher default rates and greater difficulties for the segment
(Gosztonyi and Havran, 2021).
This paper is linked to two strands of research in the literature. The first is the role of alternative information in the risk assessment
of platforms, which has so far been studied mainly in the US market. The second is whether online platforms substitute or complement
banks and serve to improve financial inclusion.
We analysed the loan level data and all cash-flows of Bondora (https://www.bondora.com/en), the sixth largest peer-to-peer
platform in Europe (p2pmarketdata.com). Bondora allows to their clients investing in personal debt starting at €1, and enables bor
rowers to receive funding directly from investors. The investments are denominated in EUR and are available for retail investors
☆
This research was supported by National Research, Development and Innovation Office - NKFIH, K-138826.
* Corresponding author.
E-mail addresses: [email protected] (B. Dömötör), [email protected] (F. Illés), [email protected].
hu (T. Ölvedi).
https://doi.org/10.1016/j.intfin.2023.101801
Received 10 August 2022; Accepted 2 July 2023
Available online 7 July 2023
1042-4431/© 2023 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
mainly from the EU. At the time of our analysis, loans were available to borrowers in 4 countries: Estonia, Finland Slovakia and Spain.
The platform itself provides the online marketplace and some services, such as credit rating of the applicants, without taking any credit
risk.
In this study we address the following research questions:
1. Do P2P lending platforms have an advantage in information processing compared to traditional banks due to the incorporation of
alternative information?
2. What is the performance profile of P2P investments for the lenders?
Our contribution to the literature is twofold. First, while the majority of studies on P2P lending focus on the US market, we analyze
the European market in Bondora’s very detailed database, which includes loan-level portfolio tables and all cash flows since the
platform was launched. Second, in order to assess the performance of the loans, we conduct an ex-post cash-flow analysis, which to our
knowledge has never been done before. By analyzing the cash-flow data of the platform, we can investigate not only the default defined
by the platforms, which ignores all further possible payments, but also the losses realized ex-post.
We found that the platform’s scoring model could not outperform our model, although we used the standard, publicly available
data of the platform. Both our model and the platform’s model rely on differences in default by country of origin, but they are poor at
predicting default within countries.
Although alternative information could improve the predicting power of the models, there is no sign of benefits from using them.
We could not confirm that the platform was able to reduce information asymmetry better than traditional financial intermediaries.
This contradiction between our findings and the results of the literature could be due to the difference between the US and the
continental European markets.
The ex-post analysis of the cash-flows shows that the average internal rate of return (IRR) is negative, and more than 40% of all
transactions end with a negative IRR; thus, there is a net loss of investment. From a regulatory perspective, our results suggest that
platform investors bear an uncompensated, maybe unforeseeable credit risk. The high default rate of non-Estonian borrowers reflects a
decrease in willingness to pay with distance, that can be due to lower crossborder collection efficiency. This inefficiency combined
with loan-sharking level interest rates (77.5% on average in the worst rating segment) may lead to adverse selection of the borrowers.
While harmonisation of regulation and improvement of investor protection is underway (Regulation (EU) 2020/1503 on European
crowdfunding service providers for business is applied from November 2021), further requirements and specifications would be
needed to improve the efficiency of cross-border collection and the transparency of platform investment performance. In addition to
ex-ante data published by the platform, ex-post performance disclosure both in terms of returns and the accuracy of models could
reduce information asymmetry.
The structure of the paper is as follows. The next section presents a summary of the literature, and then in Section 3, we present the
analysis of the portfolio and cash-flow tables of Bondora. Finally, in Section 4, the conclusions are derived.
2. Literature review
The literature on platform lending is relatively new but quite extensive. The emergence of platform lending is frequently explained
by the rapid technological changes of the last decades, which contributed to the disruption of many services in the economy (Goldstein
et al, 2019). However, there is no evidence that the potential benefits of P2P lending platforms disqualify the relevance of banks.
The existence and the rationale of banks are consequences of market imperfections, such as transaction costs, liquidity shocks, and
information asymmetry, which make perfect diversification impossible (Freixas & Rochet, 2008). Although P2P platforms are free of
the considerable fixed costs of branch networks or employers, their cost efficiency is unsupportable if we compare their average cost of
3–4% of the intermediated amount (Morse, 2015) with the long-term stable intermediation cost of 2% that is present for banks
(Philippon, 2016; Bazot, 2018). Platforms primarily only match investors and borrowers; thus, they also do not protect against
liquidity shocks.
While P2P lending underperforms, compared with banks, with regard to transaction costs and liquidity insurance, these platforms
may be more advantageous in that they reduce information asymmetry to a greater extent, as they use also soft and sensitive data given
voluntarily and apply big data and artificial intelligence more flexibly than banks (Liu et al. 2020). Another advantage of these
platforms is that they are still free of regulatory restrictions; thus, they can offer high-risk investment possibilities, without risking their
own capital (Davis, 2016).
An often-emphasized argument for fintech companies such as peer-to-peer (P2P) platforms is their flexibility to apply the latest and
most advanced data analysis methods, which ensures them a competitive advantage over traditional financial institutions (Duarte
et al., 2012; Lin et al., 2013; Jagtiani and Lemieux, 2019; Feyen et al., 2021), however, this advantage can disappear, as banks are
incentivized to improve their digital services to compete with the new challengers.
As our study relates to the importance and role of P2P lending, we focus on the strands of research that investigate the substitutive
or complementary nature of this kind of alternative financing and the impact of alternative information in the lending process.
Thakor (2020) summarizes the literature on fintech around four main questions, one of which is the role of marketplace lending in
financial intermediation. He concludes that as banks are unique in providing deposit insurance, P2P lending can complement banks,
2
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
mainly if banks are more capital constrained, serving clientele unable to pose collateral.
The empirical literature on the role of platform lending is quite widespread, investigating the evolution of the market and the
characteristics of P2P loans and comparing them to bank loans. However, only a few theoretical models aiming to situate P2P lending
in financial intermediation are present in the literature. In Merton and Thakor’s (2019) model of financial intermediation, financial
institutions are financed by two types of partners: investors and customers. Investors are willing to take on risk in exchange for an
appropriate risk-adjusted return, while customers demand financial services free of credit risk. The optimal contractual design is
determined by the cost of insulating customers from the credit risk of the intermediary. According to this concept, the distinction
between banks and the market (direct lending) disappears once the above costs, called customer contract fulfillment costs, become
sufficiently large. In the case of financing frictions, direct lending can be an attractive alternative for at least a section of the customers.
Liu et al. (2019) incorporate both social collateral and soft information into their model and show that both can reduce information
asymmetry, making financing available even to small borrowers with limited assets. In contrast to the traditional lending market, low-
risk borrowers can crowd out high-risk borrowers, and P2P platforms complement traditional banks by serving those who are not
targeted for bank lending.
The majority of empirical evidence shows that platform borrowers are mainly underbanked customers with limited access to bank
finance (Das, 2019; Maskara et al., 2021); the unavailability of other financing options encourages borrowers to turn to the platform.
De Roure et al. (2016), examining the German market, also conclude that P2P lending platforms serve an underbanked segment of low-
credit customers, which is out of the scope for conventional banks. This concept of collaboration is supported by Milne and Parboteeah
(2016), who propose that P2P platforms have a complementary function in lending activity. Specifically, they supplement traditional
banks because banks possess a few comparative advantages, which precludes platforms from competing with them. The two financing
forms may cooperate in the future.
In contrast to the above results, Tang (2019) finds, when investigating the unsecured consumer loan market in the US, that a
negative shock in the bank’s credit supply lowers the quality of P2P platforms’ credits. He concludes that the results confirm the role of
platforms as a substitute for bank lending in serving infra-marginal bank borrowers while also complementing bank lending for small
loans. These results also suggest that the credit expansion of P2P lending was based on borrowers who already have access to bank
credit. Cornaggia et al. (2018) also highlight the substitutive role of P2P platforms, stressing that smaller banks suffer losses due to the
decline in loan volumes, while large commercial banks are not affected. Following the restrictions of the COVID-19 pandemic, Najaf
et al. (2022) argue, fintech P2P lending has become the most viable alternative credit option available to borrowers. Moreover, online
services have the potential to augment or replace lending provided by traditional or conventional banking institutions.
From the investor’s perspective, however, platform funding has higher risk and less transparency, and risk management is un
derdeveloped (Milne & Parboteeah, 2016). Although platform investing provides higher interest rates, risk-adjusted interest rates are
comparable (De Roure et al., 2016), and it is questionable why unsecured P2P lending, which is not even covered by deposit insurance,
is beneficial for lenders. The existence of the market under unfavorable lending conditions can be explained by investor preferences, as
in the model of Berentsen and Markheim (2021), where altruistic investors are willing to finance even projects generating negative
expected cash-flow.
Information asymmetry in financial intermediation relates to the problem of the lender having constrained knowledge of the
borrower’s creditworthiness. The consequences of information asymmetry can be considered to be specific forms of transactional costs,
which can be reduced by monitoring and, hence, improving the efficiency of lending. Financial intermediaries create value by
economizing monitoring costs (Diamond, 1984) and by having better access to borrowers’ credit and account history or other public
sources, such as bad debtor registry or legal processes. Banks, using their own capital to finance borrowers, also provide signals about
the quality of the debtor. P2P platforms have little to no access to the previous financial history of the borrowers, and the verification of
this information is also costly and, sometimes, even impossible. As P2P lending platforms do not offer credit from their own sources,
the signaling effect is also less significant than in the case of banks. Additionally, these platforms have the advantage of applying big
data analysis techniques and obtaining “soft information,” which banks are not allowed to gather (Havrylchyk & Verdier, 2018).
Earlier, in the case of crowdfunding campaigns, borrowers and lenders were aware of one another, and the social relationship facil
itated the screening of borrowers. In the present day, this kind of proximity is atypical among borrowers and lenders; however, the
narratives submitted voluntarily by borrowers can contain sensitive information and may have a high impact on investors’ decisions.
The fact that platform lending is the most intense in consumer lending, where informational asymmetry is the highest, also confirms
the importance of the information that banks are prohibited from collecting (Havrylchyk & Verdier, 2018). Information asymmetry
and capital requirements make it unfeasible for banks to finance high-risk customers, even if they are willing to pay higher costs. By
reducing information asymmetry, P2P platforms are seemingly able to reduce credit rationing. Although financial institutions have
also started to use digital techniques to handle low-quality big data, the fact that machine learning tools work as a black box constrains
their applicability for regulatory purposes (van Liebergen, 2017).
The information P2P platforms collect is mainly based on hard data on the borrower and the credit itself, as well as other local
economic information, such as the location’s criminal statistics or employment rates (Jagtiani & Lemieux, 2018). Borrowers may
provide other data sources, such as an account of the utilities availed, public reports, and alternative lending payment history. The
narratives applicants provide regarding their goals and credit purposes are also a source of soft information; however, their usefulness
is not confirmed (Herzenstein et al, 2008). Emekter et al. (2015) examine the credit risk and performance models of platforms based on
the data of the biggest P2P lending platform Lending Club. They find that besides the platform’s credit grade, a few other variables,
3
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
such as debt-to-income ratio, FICO score (credit score created by Fair Isaac Corporation), and revolving line utilization, also have
significant explanatory power on loan default. The credit grading reflects the riskiness of the loan, but the higher interest rate charged
is not enough to compensate the investors in the worst clientele.
Jagtiani and Lemieux (2019), in their analysis of the loans of Lending Club, find that the correlation between the assigned grade of
the platform and the borrowers’ FICO score declined from 80% to 35% from 2007 to 2014, and the platform’s grades proved to perform
better while predicting loan default. Their results confirm the benefits of alternative data used by fintech lenders. Having examined an
extended sample (2007–2018) of the Lending Club consumer platform, Croux et al. (2020) also confirm the importance of alternative
data on loan default. Das (2019) highlights the importance of alternative data in developing better credit models, which allows lenders
to select creditworthy borrowers from the lower FICO score bucket, who are otherwise excluded from traditional financing. Cumming
et al. (2020) also highlight the role of soft information, proving that the higher the risk the borrower faces (large amount needed or all
or nothing financing form), the higher the length and readability of soft information provided. Hughes et al. (2022) compared the
lending efficiency of LendingClub’s with that of traditional financial institutions and they find that the platfom’s credit evaluation to be
more accurate, which they explain by the use of alternative data and complex modeling capability of fintech credit providers.
Based on the theory and the empirical evidence on P2P platforms presented in the previous sections, we first examined the in
formation processing of the P2P platform Bondora, an Estonian marketplace launched in 2009, to find evidence on the role of
alternative information. Then, to answer our second research question, we analyzed the performance of the platform’s investments.
Bondora provides different datasets, which are updated daily. We used two types of datasets: first, the raw data of the loan book
containing all loans with different applicant-related and other variables (Dataset 1 and 2, detailed description in the next subsection)
and, second, the historical payment table that includes all cash-flow series of each loan (Dataset 3). Table 1 contains the description of
them.
The loan book data were downloaded on October 5, 2020, and on March 15, 2022. The data of the first period were used to build
our benchmark model, while the data of the second period were used for the out-of-sample testing. Historical payments covering the
principal and the interest amount paid in each month, were downloaded on May 24, 2022. These data were used to examine the
performance of platform investments.
First, we analyzed the loan book containing all loans taken between December 10, 2012, and October 5, 2020. The original dataset
covered 151,866 transactions and 112 variables, including different types of information: data regarding the characteristics, financial
background and payment history of the applicant (e.g. the number and the amount of the pervious loans the borrower had before this
application), standard information about the loan request (e.g. interest rate, loan amount, tenor) and technical data (e.g. loan ID, the
form of bidding). The full list of the variables included in the original dataset is presented in the Appendix (Table 11).
During data cleaning framework, inconsistent records were eliminated (when the default date is missing but the loan is more than
90 days past due). It impacted 3,029 rows. Additionally, missing and invalid observations were excluded from the dataset (we
considered a record invalid in case its value was not defined in the Bondore data description). Loans issued within a 12-month period
(from October 6, 2019) were also eliminated, as to our definition of default 12 months after issuance is relevant. After data cleaning,
107,588 observations (loan transactions) remained in the sample.
To support the decision of investors, the platform performs a credit risk assessment for each applicant and assigns a rating based on
its internal evaluation. The methodology is not publicly available; only the final rating is shared with investors. According to the
literature, the rating of the platform is one of the main factors that impact investors’ decisions. Better rating generally results in greater
success in funding (Herzenstein et al. 2008, Emekter et al. 2015, Gavurova et al. 2018). Therefore, it is crucial that the rating
appropriately reflects the risk of the potential borrower. Furthermore, according to the platforms, one of the main advantages of P2P
Table 1
Description of the datasets of bondora used for the analysis.
Dataset 1 Dataset 2 Dataset 3
4
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
lending is their credit risk assessment process, as the platform applies alternative information for their assessment, besides the standard
variables used by banks (among others Popescu, 2016; Croux et al., 2020; Hughes et al. 2022).
According to the platform’s webpage, the current credit risk rating methodology was introduced in December 2014 to improve the
previous rating practice and support risk-based pricing. As stated, the method is in line with the industry’s best practices, commonly
applied by the banking sector. Overall, eight rating bucket grades are determined on a scale from AA (lowest risk) to HR (highest risk).
The rating classification is based on the expected loss (EL) intervals, calculated by the platform with the following formula:
EL(%) = PD*LGD*EAD (1)
where PD is the probability of default, LGD is the loss given default, and EAD is the exposure at default. The data is derived from
three main sources: information provided by the applicant, an external credit bureau database, and behavioral information collected
through the application process on the webpage (Bondora, 2021). It must be noted that this practice differs from banks’ methodology,
where the basis of rating assignment is the probability of default and not the expected loss.
The portfolio table provides information on the performance of the transaction by disclosing the number of days past due (DPD) and
the date of default if the loan is in default. However, these default indicators are not fit our purpose, as industry practice considers a
default to be a delay in payment of more than 90 days in the first 12 months after the loan is issued. Therefore, we created a default flag
accordingly and considered a borrower to be in default if they were more than 90 days in arrears in the 12 months following the
origination of the loan. In the following, we use this default definition for modeling and also when presenting ex-post default fre
quency. Since the portfolio table contains no information on the delayed amount, the materiality threshold is not included in the
default definition. It is important to note that the default defined above does not represent the ultimate loss considered in subsection
3.4, as loans that are 90 days past due in the first year after issuance may recover and meet their obligations later. Similarly, loans that
perform well in the first year may default later, leading to losses.
Table 2 below presents the distribution of the loans examined for the first period (2012–2019) based on the rating category
provided by the platform at origination and the days past due (DPD).
The platform’s credit risk assessment seems to be mostly reasonable. Claims with worse ratings usually showed worse performance.
As we had no external data on the borrowers, such as the FICO score in the US, we could not compare the platform’s ratings with
banking models, as Jagtiani and Lemieux (2019) do, but our data support their findings.
For a high-level portfolio overview, the loans’ main characteristics were examined. Table 3 provides a summary of the main
statistics by the platform’s internal rating.
The loan amounts are low, with an average of €2,531 and a maximum of €10,632. Loans are concentrated in the lower rating
categories, with more than 80% of the loans falling into the C-rated or even lower one. Risk-based pricing is reflected in the average
interest rate. Even for the AA grade, the interest rate is higher (11.50%) than the average bank interest rate in the Euro area, which is
around 6.9% for the examined period in the retail consumer loan segment (ECB, 2020). Investors’ expected returns are growing
monotonously with the lowering of the rating, except for the HR rating, where the higher risk is not compensated with a higher ex
pected yield. In comparison, the average bank deposit in the Euro area for the same period was around 0.9% for maturity over 2 years
(Euro Area Statistics, 2020). Regarding the default rate (based on the above-described default definition), the portion of defaulted
loans also increases with the worsening of the rating. The average default rate on the whole portfolio is 31%, which is extremely high,
compared to commercial banks’ retail portfolios.
To investigate the performance of the platform’s scoring model, we built a benchmark model using publicly available standard
variables, usually included in the credit risk assessment process of a commercial bank. Then, we compared our results with the
platform’s estimation to find evidence of the role of potential alternative data used by the platform.
For our benchmark model, we selected 12 standard variables related to the financial position of the borrower, their previous loan
history, and a few social features. The variables were selected taking into account the relevance of the data for scoring and their
availability. We performed sanity checks to confirm that the dataset of each variable was complete and valid. The examined variables
Table 2
The Distribution of the Portfolio by Rating and DPD.
DPD AA A B C D E F HR
5
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
Table 3
Descriptive Statistics of Bondora’s Portfolio by Rating as of 2020 October.
Rating Number of Average loan Standard deviation Average Standard deviation Average Number of Average
loans amount (EUR) of loan amount interest of interest rate expected defaulted loans default rate
return
are listed in Table 4. For descriptive statistics, see Table 12 in Appendix. We examined the value set and the distribution of each
variable and performed a few transformations where we found them reasonable.
Even if in a predictive model the variable X has a high explanatory power in the sense that the conditional distributions P(Y|X)
(where Y is the dependent variable) vary with the values of X, it may not be suitable for logit regression, for two main reasons. The
values of categorical variables are usually arbitrarily chosen integers that have no specific meaning, so their weighted sum would be
meaningless in the model. Because the logit function has low values for low PD and high values for high PD, or vice versa, if PD is not a
monotonic function of a variable, it performs poorly in the model. In such cases, some transformation must be applied to the variable to
make it fit the model. The values of the variable are grouped into ’similar’ categories (or split if the distribution is continuous), which
Table 4
Description of the variables of the scoring model.
Variable Description Transformation
6
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
means that the PD is more or less constant within each category, but varies widely between categories. If an intuitive high and low risk
category (as in our case all liabilities) can be identified, a dummy variable is usually created. If there seem to be more than two
homogeneous categories (in our case, the borrower’s country), a WOE transformation is usually applied. The average PD of each
category is calculated and plugged into the inverse logit function WOEi = ln((1 − pi )/pi ), so that a linear relationship between the
values of the variable and the log-odds is established, which is ideal for the logit model. Finally, if a variable has a “wild” distribution, a
smooth transformation (e.g. logarithm) is usually used to handle extreme values and outliers. This is the case for the total income
variable. The transformation used is detailed in Table 4.
As a next step, we estimated the GINI coefficients, which reflect the explanatory power of the variables, results are presented in
Table 5.
By examining the default explanatory power of each single variable country proved to be the strongest, but income and indebtness
characteristics are also impacting the default frequency. Although theoretically possible, it is not realistic to build strong models from
weak variables, so we only included variables with sufficiently high explanatory power. There is no exact rule for when a variable is
strong enough, but GINI above 10% is a good rule of thumb. There are two other variables close to this threshold, total liabilities and
education, which are generally considered very important for retail lendings. We included these in the model and dropped the weakest
three. Based on the correlation of the variables, we eliminated one pair of variables (with lower GINI) whose correlation coefficient
was above 0.5. To build a scoring model, we ran different logistic regressions. We had a group of variables, measuring the indebtedness
of the borrower (AmountOfPreviousLoansBeforeLoan, NoOfPreviousLoansBeforeLoan, PreviousRepaymentsBeforeLoan) and we tried
each of them together and also separated along with other variables in the regression, however the GINI of the model was the same in
the different versions. Therefore we selected the final model based on intutition and economic interpretation of the betas. The final
variables to be used for our scoring model are the country of the borrower; the type of home ownership; total income; the number of
loans taken before the loan; and the total liabilities the borrower has. The results of the regression are presented in Table 6 below.
Based on the regression results, all variables proved to be significant, all of them at the 99% significance level. The impact of income
and total liabilities is in line with our previous expectations, with higher income and lower liabilities reducing PD. Interestingly, if
borrowers had more credit before the current one, their creditworthiness improves. Home ownership has a value of zero if the
borrower owns (wholly, jointly owned, encumbered or unencumbered) the house and a value of one if the house is not owned by the
borrower (e.g. living with parents or renting). Being an owner reduces PD as expected.
Using the beta coefficients of the final model, the score for each observation was estimated. This was transformed into the
probability of default with the following formula:
1
f (y) = y
(3)
1 + e−
where
y = β0 + β1 x1 + β2 x2 + .. + βn xn (4)
The average probability of default for each rating category based on our model (Benchmark PD), the PDs estimated by the platform
at origination (ex-ante), and the real default rates using our default definition are presented in Fig. 1.
Based on our data, the platform underestimates the probability of default for all rating grades. Our model’s in-sample estimation is
a little closer to the observed default rates for the worse rating cathegories than the platform’s results, however, we overestimated the
probability of default for the best rating category.
We estimated the ROC curve to check the classification power of the models (presented in Fig. 2). Our benchmark model’s in-
sample performance resulted in a GINI of 44.10%, while the platform’s classification achieved a GINI of 41.08%. The goodness of
these GINI values is hard to judge. In the case of a commercial bank, a retail scoring model is expected to achieve a GINI higher than
80%, but for special, high-risk portfolios, significantly lower GINIs may also be acceptable. On the other hand, Jagtiani and Lemieux’s
(2018) model obtained a GINI of 38% (Area under the ROC curve 69%), even for the best-performing variable set.
The in-sample GINI we could achieve was only slightly higher than the GINI of the platform’s model for the same period. Therefore,
Table 5
The GINI value of the variables used for the scoring model.
Variable GINI
Age 0.035
country 0.387
IncomeTotal 0.153
ExistingLiabilities 0.182
LiabilitiesTotal 0.076
DebtToIncome 0.114
NoOfPreviousLoansBeforeLoan 0.214
AmountOfPreviousLoansBeforeLoan 0.186
PreviousRepaymentsBeforeLoan 0.002
Employmentduration 0.005
Education 0.075
Homeownershiptype 0.149
7
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
Table 6
Results of binary logistic regression.
Variable Estimate Std. Error z value Pr (>|z|)
Fig. 1. Comparison of the ex-ante Model PDs and the Observed Default Rate. Note. Data obtained from Bondora, as of October 2020.
1.0
1.0
0.8
0.8
0.6
0.6
Sensitivity
Sensitivity
0.4
0.4
0.2
0.2
0.0
0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
1 − Specificity 1 − Specificity
Fig. 2. Comparison of the ROC (Receiver operating characterstic) curves in-Sample. Note. The figure on the left shows our model’s curve, and the one on
the right is the curve of the platform’s model. Data obtained from Bondora, as of October 2020.
8
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
we can confirm that the platform’s model performs appropriately. However, the performance of our benchmark model suggests that a
similar result can be achieved based on a classic “banking-like” information dataset. Hence, in our investigation, we could not detect
any sign of the benefits of using alternative data sources or information processing of fintech lenders.
We also performed out-of-sample testing to check the default prediction performance on the independent data horizon. We per
formed the same data cleaning steps for the portfolio table as of March 15, 2022, as mentioned for the benchmark model. As the time
horizon of the benchmark model ends on October 5, 2019 (we filtered out loans issued within 12 months), the validation sample
consists of 50,251 loans starting from October 6, 2019. We used the parameters (WOE binning for the country variable and co
efficients) of the original model and re-estimated the PDs for the new database. The results are presented in Fig. 3, along with the
platform estimation and the observed default rates.
The platform PDs underestimated the real default rates in the best and the worst rating categories, while our model overestimated
them, in a manner similar to the previous period. The GINI of our model is 43.28%, only slightly below the in-sample value (44.10%).
The analysis of the platform’s PD resulted in a GINI of 37.92%, compared to the previous period’s 41.08% The ROC curves are pre
sented in Fig. 4. Interestingly, the drop in GINI was higher for the platform’s estimation, even though their rating is given at the
origination. Thus, this change is not due to the out-of-sample testing, as in the case of our benchmark model, but the changing market
conditions.
As the highest explanatory value—GINI—was classified by the country, it is worth examining the loans on a country level. Fig. 5
presents the country-level default rates of the rating categories, separately for the first and the second periods. Ratings and risk
assessment is adequate for Estonia, but for the other countries the default rate is not a monotonic function of rating. Loans outside
Estonia (foreign origin) are much riskier, reflected in high default rates. There are no loans provided for Slovakian borrowers in the
second period.
Loans out of Estonia not only have lower ratings, but also higher default rates in all rating categories (especially in the first period).
This suggests that the payment discipline of borrowers of foreign origin is much lower, which can be due to the weaker cross border
credit collection.
We tested the performance of our model for each country separately by calibrating a scoring model to country-level data. The
predictive power of the risk models is essentially reduced at the country level; a summary of the GINIs is shown in Table 7. Our country-
level models, which rely on the variables used for the benchmark model above except for the country (total liabilities, the type of home
ownership; total income; the number of loans taken before the loan) performed poorly with GINIs of around 20%. The platform model
performed slightly better in 5 out of 7 cases, but its discriminatory power is also very low.
The decline in the performance of the scoring model at the country level is due to the fact that the rating of borrowers differs
significantly across countries. Consequently, the score of each loan is determined by the country of the borrower and the other var
iables have much less explanatory power. It seems that in the high-risk segment, where platform lending is active, individual defaults
are much less predictable. Since the platform model could not significantly outperform our naive models, and both performed rather
poorly, it seems that additional - alternative - information could help to build a better model. However, based on the data analysed,
there is no evidence that platforms use and exploit alternative information.It is also important to note that the high default frequency
suggests that P2P loans differ significantly from the loans acceptable for traditional financial institutions. Thus, using traditional
techniques for their risk assessment is not appropriate.
To better understand the characteristics of P2P loans and the motives of P2P lending from the lenders’ perspective, in the following
section, we analyze the ex-post performance of platform investments.
Fig. 3. Comparison of the Out-of-Sample PDs and the Real Default Rate. Notes. Benchmark model Data obtained from Bondora, as of March 2022.
9
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
Fig. 4. Comparison of the ROC Curves out-of Sample. Note. The figure on the left shows our model’s curve, and the one on the right is the curve of the
platform’s model. Data obtained from Bondora, as of March 2022.
Fig. 5. Default Rates by Country for the Two Periods. Note. The size of the bubbles reflects the number of loans. Period 1 contains loans issued between
December 2012 and October 2019, Period 2 lasts from October 2019 to March 2021. Data obtained from Bondora, as of October 2020 (period 1) and
March 2022 (period 2).
Table 7
Default Rates and Performance (GINI) of the scoring models by country.
Default rate Number of loans Platform model GINI Benchmark model GINI
Note. Data obtained from Bondora, as of October 2020 and March 2022.
10
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
P2P lending represents high-risk, bank-ineligible loans. Due to the strict regulatory requirements and reputational risks, banks are
unable to provide financing for this segment, but individual or institutional investors may benefit if the risk is compensated by high
interest rates.
To understand the effective performance of P2P loans, we examined the dataset of historical payments, with the principal and the
interest amount paid each month by loan ID. The cash-flow table of Bondora downloaded on May 24, 2022, contains 4,343,194 rows,
representing all payments during the lifetime of the platform. First, we ordered the payments according to loan ID. Consequently, we
got the detailed cash-flow of 243,453 loan transactions. Then, we ordered the other details (rating, PD, status, etc.) based on lon ID
from the portfolio table according to each loan.
Bondora assigns a status to each loan: current (transactions in progress), closed or unavailable (NA). For the latter category, other
data is missing and therefore not suitable for analysis. As ex-post analysis is feasible only for closed transactions, so, first, we inves
tigated the loans with a closed status. We calculated the IRR for each loan based on the cash-flow and the historical payment schedule.
62 537 transactions issued between 28 February 2009 and 14 March 2022 were analysed, excluding loans for which no payment was
made other than at the time of disbursement.
Table 8. shows the ex-post performance of closed transactions according to rating categories. The rating was missing for 2,556
transactions. Thus, they are shown separately in the table. Average loan term refers to the difference between the last and the first
payment, Sum CF is calculated by simply summing up all cash-flows.
The worst two categories offer really a high return with high volatility, but surprisingly, the effective loss (negative IRR) in those
segments is not significantly higher than in the much better rating categories. Because of the definition of default, transactions with the
defaulted flag do not necessarily represent an actual loss, and defaulting transactions that are past due for more than 90 days in the first
year after origination may subsequently be recovered and all their obligations settled. Additionally, even non-defaulting transactions
can result in losses if the default occurs outside the first annual period after the contract is signed. The average ex-post default rate
denoted by the platform was 14.11% (see Table 8), much lower than the proportion of loans having non-performance status in both
periods (see Figs. 1 and 3.). Investments with an effective loss—negative IRR—make up even less, specifically, 12.96% of the loans. On
average, investors realizing negative IRR lost 23% of their initial investment. We can conclude that the majority of the defaulted closed
transactions were recovered, ensuring an average IRR of 27.72%.
The picture is not that bright if a closer look is provided at transactions with a “current” status. Although we do not have the original
repayment schedule of the loans, which could be indicative of non-payment and default problems, we also included in the analysis
loans where the original maturity (calculated from the date of issuance and the original loan duration) is exceeded at the time of data
collection or where no payment was made in the last one-year period. We assume here that these transactions can also be considered
closed as there will be no further related payments. This represents 42,598 additional loans that are more problematic, 48% of which
have received a default flag from the platform. In total, including closed status transactions, Bondora’s total closed portfolio consists of
105,135 deals.
The distribution of actual loan term is presented on the histogram in Fig. 6. The P2P loans are mainly short-term transactions, half
of them were paid back below one year, while the average original loan term is around 3.5 years. We also found that the majority of the
debtors (69%) prepaid the claim before the maturity date, as there is no prepayment charge to be paid by the borrower (Bondora,
2021).
The relationship between the annual IRR of the loan and the interest rate priced initially by the platform is shown in Fig. 7. The IRR
distribution is wider for riskier, higher-level interest rates, and the correlation coefficient between interest rate and IRR is low, below
0.09. Although a substantial section of the high-risk borrowers performs well, offering above 100% return for the investors, a maximal,
100% loss can be realized at all interest levels.
The ex-post performance of the extended closed category is shown in Table 9. The average IRR of the portfolio is negative, which
means that investors on average not only do not receive compensation for the risk, but also make a loss on their original investment..
The average IRR is negative even in the best rating categories, with only rating C, HR, and the unrated transactions resulting in a
positive IRR. The IRR dispersion is high, but overall 41.63% of all transactions have a negative IRR and the realised loss is 55% of the
amount invested (for transactions with a negative IRR, the nominal amount of payments received is on average 55% less than the initial
investment). So, despite an average initial expected return of 9.58% − 15.52%, the average realised return is negative in most rating
categories.
The negative IRR of platform loans is particularly striking when considering that the period 2012–2022 was a period of economic
recovery and boom following the 2007–2009 crisis. The non-performing loan ratio ranged between 7.48% and 1.79% (Statista.com),
with a monotonic downward trend after 2015. To assess the performance of P2P lending from an investor perspective, the 10-year
performance of different asset classes is presented in Table 10.
The IRR is not comparable to the ex-post returns of indices or individual assets, but it is clear that all asset classes performed well
over the period and generated a positive risk premium. Thus, the negative performance of platform investments is not due to economic
factors.
Breaking down the results by the country of the borrower, we found that loans for Estonian borrowers had a slightly positive
average IRR of 2.93%, while the average IRR is negative for all other countries. Spain had the lowest IRR with an average of –22.57%,
followed by Slovakia with –15% and Finland with –7.6%%. The distributions are shown in Fig. 8. It seems that the willingness to pay
reduces with physical distance.
The IRR distribution of the analyzed loans is shown in Fig. 9. The green boxes show the distribution of the loans with a non-default
11
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
Table 8
Main characteristics and ex-post performance of closed transactions.
Rating Number of Default Average loan Average loan term (in Average Sum IRR mean IRR St. Dev. P (IRR < 0)
loans rate amount days) CF
Fig. 6. Distribution of Actual Loan Term (in Days). Note. Based on the extended closed dataset of Bondora. Data obtained from Bondora, as of
May 2022.
status, while the orange boxes are the defaulted loans. The standard deviations increase as the rating worsens, but there is no sig
nificant difference in the dispersion of the distributions within a given category. The IRR is positive for about 75% of the non-defaulted
deals and negative for the majority of the loans with a default flag, except for the unrated category, in which even the loans with a
default flag realized a positive IRR in 75% of the cases.
According to the analysis of the realized cash-flows and the IRR, investors do not seem to be compensated for the high risk un
dertaken by P2P loans. The ex-post return is negative on average, confirming the findings of Emekter et. al (2015). The negative IRR
also casts doubts on the economic rationale of platform lending and raises the potential role of other motives such as altruism, as
discussed in Berentsen and Markheim (2021). However, interest rates of up to 260% depending on creditworthiness do not reflect
altruistic motives. It is also worth noting that, in addition to the transactions analysed, there where 4 997 loan IDs that have only one
cash-flow, the initial loan disbursement. We can consider them to be credit fraud, where the borrower had no intention to pay anything
back. Fraud transactions make up more than 2% of all loan IDs.
4. Conclusion
This paper investigated the possible explanations for the rapid growth in P2P platforms’ market share in the credit market. From a
theoretical perspective, the strongest argument for the relevance of P2P lending is the reduction in information asymmetry through the
use of alternative data and P2P platforms’ information processing. We analyzed the loan-level data of an Estonian platform, Bondora,
to find evidence on the benefit of using alternative information based on the performance of the credit risk model of the platform. We
found that the grades assigned by the platform are in line with the default risk of the borrower, but the assigned default probabilities
underestimated the real default rate in all segments. Our benchmark scoring model performed slightly better than the platform’s model
both on the in-sample and out-of-sample data; however, we used traditional explanatory variables: age, debt-to-income, home
ownership, employment duration, country of origin, and existing liabilities. When looking at loans by country of origin, the perfor
mance of the models decreases significantly, indicating that both our model and the platform’s model primarily capture different
12
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
Fig. 7. Realized IRR as a Function of the Interest Rate. Note. Based on the extended closed dataset of Bondora. Data obtained from Bondora, as of
May 2022.
Table 9
Main characteristics and ex-post performance of the extended closed dataset.
Rating Number of Default Average loan Average loan term (in Average Sum IRR mean IRR St. Dev. P (IRR < 0)
loans rate amount days) CF
Table 10
10-years Performance of different investments.
Investment type 10 yearAnnual return Volatility
Note. MSCI data as of 31th March 2023, Bitcoin statistics based on daily values available at investing.com.
country-level creditworthiness. Therefore, alternative data may be needed to improve the models, but our results do not confirm that
the platform can incorporate them.
By analyzing the ex-post performance of “closed” transactions, we found a significantly high average IRR of 27.72%. However, if
we extend the analysis to transactions with a “current” status that are not expected to generate further cash flows, we find that the IRR
is negative on average, with 41.63% of all transactions ending in a net loss (negative IRR).
There are huge differences in credit performance across countries. Foreign (non-Estonian) borrowers are not just lower rated, but
their credit performance is significantly worse even in the same rating category, indicating the inefficiency of crossborder collection
and higher information asymmetry. We conclude that the high credit risk reflected in the extremely high default rates is associated
with net ex-post loss of the investment. Although the level of the interest rates excludes the altruistic motives of the investors, they are
13
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
Fig. 8. IRR Distribution According to Country. Note. Based on the extended closed dataset of Bondora. Data obtained from Bondora, as of May 2022.
Fig. 9. Distribution of IRR by Rating Category and Default Status. Note. Based on the extended closed dataset of Bondora. Data obtained from Bondora,
as of May 2022.
insufficient to compensate investors for the credit risk. The loan-sharking level interest rates on the other hand may lead to adverse
selection on the borrower side.
Platform loans represent high-risk transactions that may not be acceptable for a traditional financial institution; thus, P2P lending
complements traditional financial intermediation. However, the negative ex-post performance casts doubt on the rationale of P2P
investments, even for market participants free of capital burden and reputation risk. Regulation is now better focused on the segment,
but more transparency, disclosure of ex-post performance of loans and models and improvement of the crossborder collection would be
needed to reduce information asymmetry and adverse selection.
14
B. Dömötör et al. Journal of International Financial Markets, Institutions & Money 86 (2023) 101801
Barbara Dömötör: Conceptualization, Methodology, Validation, Formal analysis, Supervision. Ferenc Illés: Methodology, Soft
ware, Validation, Formal analysis, Visualization. Tímea Ölvedi: Conceptualization, Methodology, Software, Validation, Formal
analysis, Investigation, Visualization.
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to
influence the work reported in this paper.
Data availability
Acknowledgement
This research was supported by National Research, Development and Innovation Office - NKFIH, K-138826
Appendix
Table 11
List of all variables in the raw portfolio table.
15
B. Dömötör et al.
Table 12
Descriptive statistics for the variables in the scoring model.
Age Income Debt to Country Employmentduration Education Existing Liabilities No of previous Amount of Previous Home
total income liabilities total loans before previous loans repayments ownership
loan before loan before loan type
Minimum 18 0 0 – – – 0 0 0 0 0 –
Maximum 75 1 012 019 4 607.82 40 145 042 25 44 417 34 077
References
Bazot, G., 2018. Financial consumption and the cost of finance: Measuring financial efficiency in Europe (1950–2007). J. Eur. Econ. Assoc. 16 (1), 123–160. https://
doi.org/10.1093/jeea/jvx008.
Berentsen, A., Markheim, M., 2021. Peer-to-peer lending and financial inclusion with altruistic investors. Int. Rev. Financ. 21 (4), 1407–1418. https://doi.org/
10.1111/irfi.12333.
Bondora (2021). What are the prepayment risks? Available: https://support.bondora.com/en/what-are-the-prepayment-risks.
Cornaggia, J., Wolfe, B., & Yoo, W. (2018). Crowding out banks: Credit substitution by peer-to-peer lending. Available at SSRN 3000593. 10.2139/ssrn.3000593.
Croux, C., Jagtiani, J., Korivi, T., Vulanovic, M., 2020. Important factors determining Fintech loan default: Evidence from a lendingclub consumer platform. J. Econ.
Behav. Organ. 173, 270–296. https://doi.org/10.1016/j.jebo.2020.03.016.
Cumming, D.J., Leboeuf, G., Schwienbacher, A., 2020. Crowdfunding models: Keep-it-all vs. all-or-nothing. Financ. Manag. 49 (2), 331–360. https://doi.org/
10.2139/ssrn.2447567.
Das, S.R., 2019. The future of fintech. Financ. Manag. 48 (4), 981–1007. https://doi.org/10.1111/fima.12297.
Davis, K., 2016. Peer-to-peer lending: structures, risks and regulation. JASSA 3, 37–44. https://search.informit.org/doi/10.3316/informit.419817919644101.
De Roure, C., Pelizzon, L., & Tasca, P. (2016). How does P2P lending fit into the consumer credit market? Bundesbank Discussion Paper No. 30/2016. 10.2139/
ssrn.2756191.
Diamond, D.W., 1984. Financial intermediation and delegated monitoring. Rev. Econ. Stud. 51 (3), 393–414. https://doi.org/10.2307/2297430.
Duarte, J., Siegel, S., Young, L., 2012. Trust and credit: The role of appearance in peer-to-peer lending. Rev. Financ. Stud. 25 (8), 2455–2484. https://doi.org/
10.1093/rfs/hhs071.
Emekter, R., Tu, Y., Jirasakuldech, B., Lu, M., 2015. Evaluating credit risk and loan performance in online Peer-to-Peer (P2P) lending. Appl. Econ. 47 (1), 54–70.
https://doi.org/10.1080/00036846.2014.962222.
Euro Area Statistics (2020). Bank interest rates - Deposit. Downloaded: 2020.12.02. Available: https://www.euro-area-statistics.org/bank-interest-rates-deposits?
cr=eur&lg=en&page=0&charts=M..B.L22.L.R.A.2250.EUR.N+M..B.L22.F.R.A.2250.EUR.N+M..B.L22.K.R.A.2250.EUR.N&template=1.
European Central Bank (2020). Statistics. Downloaded: 2020.12.02. Available: https://www.ecb.europa.eu/stats/html/index.en.html.
Feyen, E., Frost, J., Gambacorta, L., Natarajan, H., Saal, M., 2021. Fintech and the digital transformation of financial services: implications for market structure and
public policy. BIS Papers.
Freixas, X., Rochet, J.C., 2008. Microeconomics of banking. MIT press.
Gavurova, B., Dujcak, M., Kovac, V., Kotásková, A., 2018. Determinants of successful loan application at peer-to-peer lending market. Economics & Sociology 11 (1),
85–99. https://doi.org/10.14254/2071-789X.2018/11-1/6.
Goldstein, I., Jiang, W., Karolyi, G.A., 2019. To FinTech and beyond. Rev. Financ. Stud. 32 (5), 1647–1661. https://doi.org/10.1093/rfs/hhz025.
Gosztonyi, M., Havran, D., 2021. Highways to Hell? Paths Towards the Formal Financial Exclusion: Empirical Lessons of the Households from Northern Hungary. Eur.
J. Dev. Res. 1–34 https://doi.org/10.1057/s41287-021-00434-9.
Havrylchyk, O., Verdier, M., 2018. The financial intermediation role of the P2P lending platforms. Comp. Econ. Stud. 60 (1), 115–130. https://doi.org/10.1057/
s41294-017-0045-1.
Herzenstein, M., Andrews, R.L., Dholakia, U.M., Lyandres, E., 2008. The democratization of personal consumer loans? Determinants of success in online peer-to-peer
lending communities. Boston University School of Management Research Paper 14 (6), 1–36.
Hughes, J.P., Jagtiani, J., Moon, C.G., 2022. Consumer lending efficiency: commercial banks versus a fintech lender. Financial Innovation 8 (1), 1–39. https://doi.org/
10.1186/s40854-021-00326-1.
Jagtiani, J., Lemieux, C., 2018. Do fintech lenders penetrate areas that are underserved by traditional banks? J. Econ. Bus. 100, 43–54. https://doi.org/10.21799/
frbp.wp.2018.13.
Jagtiani, J., Lemieux, C., 2019. The roles of alternative data and machine learning in fintech lending: Evidence from the LendingClub consumer platform. Financ.
Manag. 48 (4), 1009–1029. https://doi.org/10.21799/frbp.wp.2018.15.
Lin, M., Prabhala, N.R., Viswanathan, S., 2013. Judging borrowers by the company they keep: Friendship networks and information asymmetry in online peer-to-peer
lending. Manag. Sci. 59 (1), 17–35. https://doi.org/10.1287/mnsc.1120.1560.
Liu, Z., Shang, J., Wu, S.Y., Chen, P.Y., 2020. Social collateral, soft information and online peer-to-peer lending: A theoretical model. Eur. J. Oper. Res. 281 (2),
428–438. https://doi.org/10.1016/j.ejor.2019.08.038.
Maskara, P.K., Kuvvet, E., Chen, G., 2021. The role of P2P platforms in enhancing financial inclusion in the United States: An analysis of peer-to-peer lending across
the rural–urban divide. Financ. Manag. 50 (3), 747–774. https://doi.org/10.1111/fima.12341.
Merton, R.C., Thakor, R.T., 2019. Customers and investors: a framework for understanding the evolution of financial institutions. J. Financ. Intermed. 39, 4–18.
https://doi.org/10.1016/j.jfi.2018.06.002.
Milne, A., Parboteeah, P., 2016. The business models and economics of peer-to-peer lending. European Credit Research Institute. https://doi.org/10.2139/
ssrn.2763682.
Morse, A., 2015. Peer-to-peer crowdfunding: Information and the potential for disruption in consumer lending. Annu. Rev. Financ. Econ. 7, 463–482. https://doi.org/
10.3386/w20899.
Najaf, K., Subramaniam, R.K., Atayah, O.F., 2022. Understanding the implications of FinTech Peer-to-Peer (P2P) lending during the COVID-19 pandemic. Journal of
Sustainable Finance & Investment 12 (1), 87–102. https://doi.org/10.1080/20430795.2021.1917225.
Philippon, T. (2016). The fintech opportunity (No. w22476). National Bureau of Economic Research. 10.3386/w22476.
Polasik, M., Huterska, A., Iftikhar, R., Mikula, Š., 2020. The impact of Payment Services Directive 2 on the PayTech sector development in Europe. J. Econ. Behav.
Organ. 178, 385–401. https://doi.org/10.1016/j.jebo.2020.07.010.
Popescu, G. (2016). Alternative data usage by FinTech lenders. Available: https://cointimes.info/2016/06/20/over-12000-variables-of-alternative-data-for-36-
billion-in-alternative-loans/.
Tang, H., 2019. Peer-to-peer lenders versus banks: substitutes or complements? Rev. Financ. Stud. 32 (5), 1900–1938. https://doi.org/10.1093/rfs/hhy137.
Thakor, A.V., 2020. Fintech and banking: what do we know? J. Financ. Intermed. 41, 100833 https://doi.org/10.1016/j.jfi.2019.100833.
Van Liebergen, B., 2017. Machine learning: a revolution in risk management and compliance? J. Finan. Transf. 45, 60–67.
Cited websites
17