机器学习和非传统数据如何影响信用评分（英文版）.pdf

BIS Working Papers No 834 How do machine learning and non-traditional data affect credit scoring? New evidence from a Chinese fintech firm by Leonardo Gambacorta, Yiping Huang, Han Qiu and Jingyi Wang Monetary and Economic Department December 2019 JEL classification: G17, G18, G23, G32 Keywords: fintech, credit scoring, non-tradtitional information, machine learning, credit riskBIS Working Papers are written by members of the Monetary and Economic Department of the Bank for International Settlements, and from time to time by other economists, and are published by the Bank. The papers are on subjects of topical interest and are technical in character. The views expressed in them are those of their authors and not necessarily the views of the BIS. This publication is available on the BIS website (bis). Bank for International Settlements 2019. All rights reserved. Brief excerpts may be reproduced or translated provided the source is stated. ISSN 1020-0959 (print) ISSN 1682-7678 (online)WP834 How do machine learning and non-traditional data affect credit scoring? 1 How do machine learning and non-traditional data affect credit scoring? New evidence from a Chinese fintech firm Leonardo Gambacorta, Yiping Huang, Han Qiu and Jingyi Wang Abstract This paper compares the predictive power of credit scoring models based on machine learning techniques with that of traditional loss and default models. Using proprietary transaction-level data from a leading fintech company in China for the period between May and September 2017, we test the performance of different models to predict losses and defaults both in normal times and when the economy is subject to a shock. In particular, we analyse the case of an (exogenous) change in regulation policy on shadow banking in China that caused lending to decline and credit conditions to deteriorate. We find that the model based on machine learning and non-traditional data is better able to predict losses and defaults than traditional models in the presence of a negative shock to the aggregate credit supply. One possible reason for this is that machine learning can better mine the non-linear relationship between variables in a period of stress. Finally, the comparative advantage of the model that uses the fintech credit scoring technique based on machine learning and big data tends to decline for borrowers with a longer credit history. JEL classification: G17, G18, G23, G32 Keywords: fintech, credit scoring, non-traditional information, machine learning, credit risk BIS and CEPR. Institute of Digital Finance and National School of Development, Peking University. School of Finance, Central University of Economics and Finance; and Institute of Digital Finance, Peking University. We would like to thank Sebastian Doerr, John V. Duca, Jon Frost, Xiang Li, Julapa Jagtiani and, in particular, one anonymous referee for comments and suggestions. We would also like to thank seminar participants at the University of Basel, the Bank for International Settlements, Bocconi University and the Irving Fisher Committee Central Bank of Malaysia for useful comments. We thank Giulio Cornelli for excellent research assistance. Yiping Huang and Han Qius work is supported by the Chinese National Social Science Foundation (Project 18ZDA091). The views in this paper are those of the authors only and do not necessarily reflect those of the Bank for International Settlements. The authors wish to highlight that the data and analysis reported in this paper may contain errors and are not suited for the purpose of company valuation or to deduce conclusions about the business success and/or commercial strategy of the anonymous Chinese fintech firm. All statements made reflect the private opinions of the authors and do not express any official position of the anonymous fintech firm or its management. The analysis was undertaken in strict compliance with Chinese privacy law. The authors declare that they have no relevant or material financial interests that relate to the research described in this paper. The anonymous fintech firm did not exercise any influence on the content of this paper, but has ensured the confidentiality of the (raw) data.2 WP834 How do machine learning and non-traditional data affect credit scoring? 1. Introduction Financial technology (fintech) is taking on an ever more important role in lending decisions, while lending by fintech companies is gaining a significant share of certain market segments. In the United States, for instance, online lenders now account for about 812% of new mortgage loan originations, with Quicken Loans being recognised as the countrys largest mortgage lender in terms of flow at the end of 2017 (Buchak et al (2017); Fuster et al (2018). China is a country where new fintech credit is relatively well developed, representing around 3% of total outstanding credit to the non-bank sector at the end of 2017 (BIS (2019). New credit scoring models used by fintech lenders differ from traditional models in two key ways. The first is that technology allows financial intermediaries to collect and use a larger quantity of information. Fintech credit platforms may use alternative data sources, including insights gained from social media activity (U.S. Department of the Treasury (2016); Jagtiani and Lemieux (2018a) and users digital footprints (Berg et al (2018). In the case of large technology companies (big techs) with existing platforms, data collection extends to orders, transactions and customer reviews (Frost et al (2019). The second difference is the adoption of machine learning techniques. In contrast to traditional linear models such as the logit model, machine learning can mine the non-linear information from variables. For example, Khandani et al (2010) construct a non-linear, non-parametric forecasting model for consumer credit that is based on machine learning techniques and find that this new model can outperform other models in a range from 6% to 25% of total losses. However, the prediction capability of machine learning models has mainly been demonstrated in applications with a stationary external environment. Their performance also needs to be verified in the case of a structural shock that changes the main relationships between the variables. This paper contributes to the literature by addressing the following four questions: i) Are machine learning-based fintech credit scoring models better able to predict borrowers losses and defaults than traditional empirical models? ii) What is the information content of non-traditional sources such as digital applications on mobile phones and e-commerce platform data? iii) How do the different models perform in the event of an (exogenous) shock? iv) How do the different models perform for customers with a different credit history? The first two questions have also been analysed by other papers, with mixed results. Our contribution is mostly to highlight and explain differences in the results using a more comprehensive set of control variables. The third and fourth questions are completely new and represent the main contribution of the paper. To answer these four questions, we use a unique data set from a leading Chinese fintech company at loan-transaction level for the period between May and September 2017. The fintech firm has requested to remain anonymous but has given us access to a very comprehensive data set. Compared to previous studies, this data set allows us to disentangle the effects of traditional bank-type information (credit cardWP834 How do machine learning and non-traditional data affect credit scoring? 3 information) and non-traditional information (obtained from the use of digital applications on mobile phones and e-commerce platforms). Moreover, we can assess the performance of the credit scores calculated by the fintech company using machine learning methods and such large volumes of data. Papers based on data from Renrendai, a Beijing-based company providing P2P financial services (see, for example, Braggion et al, 2019) cannot use credit card transaction information because Renrendais borrowers typically do not have a current account with a bank. Furthermore, unlike other fintech companies, in which borrower information is self-reported by the users themselves (see for example Berg et al (2018), our fintech company is able to read both credit card transactions and digital app information directly from the system (with the users permission). The information is therefore collected more comprehensively to include both credit card information and additional non-traditional information. We analyse personal loans, most of which are repayable in up to one year. We also observe the borrowers repayment record until October 2018 in order to track the status (viable or defaulted) of each loan after origination. This enables us to evaluate the performance of each loan ex post in terms of losses and defaults. In order to answer the third question, we analyse the effects of a largely unexpected regulatory change that occurred in China in the period under review. On 17 November 2017, the Peoples Bank of China (PBoC) the Chinese central bank issued specific draft guidelines to tighten regulations on shadow banking. This regulatory change has led many financial intermediaries to increase their lending requirements, causing credit conditions for borrowers to deteriorate. In particular, the aggregated data indicate a significant increase in the default rate and a drop in lending after the shock. A similar pattern can be observed at our fintech company, which enables us to study how the different models performed during this stress period. The main conclusions of our paper can be summarised as follows: i) The fintechs machine learning-based credit scoring models outperform traditional empirical models (using both traditional and non-traditional information) in predicting borrowers losses and defaults. ii) Non-traditional information improves the predictive power of the model. iii) While the models perform similarly well in normal times, the model based on machine learning is better able to predict losses and defaults following a negative shock to the aggregate credit supply. One possible reason for this is that machine learning can better mine the non-linear relationship between variables in the event of a shock. iv) The predictive power of all the models improves when the length of the relationship between bank and customer increases. However, the comparative advantage of the model that uses the fintech credit scoring technique based on machine learning tends to decline when the length of the relationship increases.4 WP834 How do machine learning and non-traditional data affect credit scoring? 2. Literature review A few studies have started to analyse how credit supplied by fintech firms and their scoring models perform compared with traditional bank lending. Jagtiani and Lemieux (2018a) compare loans made by a large fintech lender and similar loans that were originated through traditional banking channels. Specifically, they use account- level data from LendingClub and the Y-14M data reported by bank holding companies with total assets of $50 billion or more. They find a high correlation between interest rate spreads, LendingClub rating grades and loan performance. Interestingly, the correlations between the rating grades and FICO scores have declined from about 80% (for loans that were originated in 2007) to only about 35% for recent vintages (originated in 20142015), indicating that LendingClub has increasingly used non-traditional alternative data. Using market-wide, loan-level data on US mortgage applications and originations, Fuster et al (2018) show that fintech lenders process mortgage applications about 20% faster than other lenders, even when controlling for detailed loan, borrower and geographic observables. It is interesting to note that faster processing does not come at the cost of higher defaults. Furthermore, fintech lenders adjust their supply more elastically than other lenders in response to exogenous mortgage demand shocks, thereby alleviating the capacity constraints associated with traditional mortgage lending. Buchak et al (2018) compare the pricing of online (fintech) lenders in the US mortgage market with the pricing of banks and (non- fintech) shadow banks; they find that fintech lenders charge a premium of 1416 basis points over bank mortgages. Jagtiani et al (2019) find that fintech lenders in the United States tend to supply more mortgages to consumers with weaker credit scores than do banks; they also have greater market shares in areas with lower credit scores and higher mortgage denial rates. While banks usually incentivise borrowers to pay their loans back by requiring them to pledge tangible assets (eg real estate) as collateral, fintech credit is typically uncollateralised. This makes the use of big data particularly relevant when considering a loan application. Preliminary evidence based on credit data for China suggests that big data can act as a substitute for collateral: the volume of corporate loans supplied by big techs does not correlate with asset prices, whereas bank loans do. The left- hand panel of Figure 1 shows that the elasticity of bank credit to firms with respect to asset prices is close to one for collateralised credit and 0.5 for bank credit to SMEs, whereas credit to small firms it is not statistically different from zero in the case of big tech. Frost et al (2019) use data for Mercado Credito, which provides credit lines to small firms in Argentina on the e-commerce platform Mercado Libre. They find that, when it comes to predicting loss rates, credit scoring techniques based on big data and machine learning have so far outperformed credit bureau ratings. A key question here is whether this outperformance will persist through a full business and financial cycle. Indeed, fintech credit could give rise to new forms of non-prudent risk-taking that needs to be tested in the event of an adverse shock. For example, De Roure et al (2016) find that online lenders in Germany substitute bank loans for high-risk consumer loans. For US consumer credit markets, Tang (2019) finds that online lending substitutes for bank lending by serving marginal borrowers, but complements bank lending in terms of loans. Interestingly, the performance of online lenders seemsWP834 How do machine learning and non-traditional data affect credit scoring? 5 to depend on the quantity and quality of information to which online lenders have access. Some of the literature looks at the informational content of digital soft information and credit performance. Dorfleitner et al (2016) study the relationship between soft factors in P2P loan applications and financing and default outcomes. Using data on the two leading European P2P lending platforms, Smava and Auxmoney, they find that soft factors influence th