國際清算銀行（BIS）：2021年金融科技與金融包容報告（英文版）（19頁）.pdf

《國際清算銀行（BIS）：2021年金融科技與金融包容報告（英文版）（19頁）.pdf》由會員分享，可在線閱讀，更多相關《國際清算銀行（BIS）：2021年金融科技與金融包容報告（英文版）（19頁）.pdf（19頁珍藏版）》請在三個皮匠報告上搜索。

1、BIS Working Papers No 841On fintech and financial inclusion by Thomas Philippon Monetary and Economic Department February 2020 JEL classification: E2, G2, N2. Keywords: fintech, discrimination, robo advising, credit scoring, big data, machine learning BIS Working Papers are written by members of the

2、 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 necessa

3、rily the views of the BIS. This publication is available on the BIS website (www.bis.org). Bank for International Settlements 2020. All rights reserved. Brief excerpts may be reproduced or translated provided the source is stated. ISSN 1020-0959 (print) ISSN 1682-7678 (online) On Fintech and Financi

4、al InclusionThomas PhilipponSeptember 2019AbstractThe cost of financial intermediation has declined in recent years thanks to technology and increased competitionin some parts of the finance industry. I document this fact and I analyze two features of new financial technologiesthat have stirred cont

5、roversy: returns to scale and the use of big data and machine learning. I argue that thenature of fixed versus variable costs in robo-advising is likely to democratize access to financial services. Bigdata is likely to reduce the impact of negative prejudice in the credit market but it could reduce

6、the effectivenessof existing policies aimed at protecting minorities.JEL: E2, G2, N2Stern School of Business, New York University; NBER and CEPR. This paper was prepared for the 2019 BIS Annual ResearchConference. I am grateful to my discussants Manju Puri and David Dorn, to Hyun Shin, Marina Niessn

7、er, and participants at the 2019BIS Annual Research Conference. I thank Marcos Sonnervig for outstanding research assistance.1Fintech covers digital innovations and technology-enabled business model innovations in the financial sector.Such innovations can disrupt existing industry structures and blu

8、r industry boundaries, facilitate strategic disin-termediation, revolutionize how existing firms create and deliver products and services, provide new gateways forentrepreneurship, and democratize access to financial services. On the other hand, they create significant privacy,regulatory and law-enf

9、orcement challenges and they could increase the scope for some forms of discrimination.Examples of innovations that are central to Fintech today include various application of blockchain technologies,new digital advisory and trading systems, artificial intelligence and machine learning, peer-to-peer

10、 lending, equitycrowdfunding and mobile payment systems. In this paper I offer some preliminary evidence and theoretical analysisabout the impact of technological progress in the finance industry.The first question is whether there has been any material change in financial intermediation in recent y

11、ears.To shed some light on this question, I update the work of Philippon (2015) with post-crisis U.S. data. The puzzleemphasized in previous work was that the unit cost of financial intermediation had remained stubbornly close to 200basis points for more than a century, despite advances and large in

12、vestments in computers and communicationtechnologies. The post-crisis data suggests that this puzzle might be diminishing. I find that the unit cost offinancial intermediation has declined over the past 10 years.I then study two issues that are at the heart of the Fintech debate: access to finance a

13、nd discrimination. If weaccept the fact that Fintech brings efficiency gains to financial intermediation, the next question is: how will thesegains be shared? Will Fintech democratize access to financial services or will it increase inequality? I highlighttwo forces that will shape the answer to the

14、se questions. The first force is increasing returns to scale brought bytechnology. I argue that the nature of fixed versus variable costs has changed in a way that is likely to improveaccess to financial services. It may not, however, reduce inequality among all groups.The second force is the use of

15、 big data and machine learning (BDML for short). I argue that this technologyis likely to reduce unwarranted human biases against minorities, but it will probably decrease the effectiveness ofexisting regulations. The tentative conclusion is that Fintech can bring widely-shared welfare benefits but

16、changesin existing policies and regulations are necessary to achieve its full potential.Recent literaturePhilippon (2016) discusses the literature up to 2016 so I will mention here some recent papers.Focusing on residential mortgages, Buchak et al. (2018) study the growth in the market share of shad

17、ow bank andFintech lenders, arguing that it can be explained by differences in regulation and technological advantages. They findthat Fintech lenders serve more creditworthy borrowers (relative to shadow banks) but charge higher interest rates(14-16 basis points), which is consistent with the idea t

18、hat consumer are willing to pay for better user experience andquick decisions. Fuster et al. (2019) study the differences between Fintech and traditional lenders in the mortgagemarket and find that the former is quicker in processing applications (20% faster), without increasing loan risk.They also

19、provide evidence that Fintech lenders adjust supply more elastically to demand shocks and increase the2propensity to refinance, especially among borrowers that are likely to benefit from it. Their results suggest thatFintech firms have improved the efficiency of financial intermediation in mortgage

20、markets.The advent of Fintech is often seen as a promising avenue for reducing inequality in access to credit. Bartlettet al. (2018) study this issue, analyzing the role of Fintech lenders in alleviating discrimination in mortgage markets.They find that all lenders, including Fintech, charge minorit

21、ies more for purchase and refinance mortgages but thatFintech algorithms discriminate 40% less than face-to-face lenders. Regarding the use of new technologies in creditmarkets, Berg et al. (2019) analyse the information content of the “digital footprint” (an easily accessible informationfor any fir

22、m conducting business in the digital sphere) for predicting consumer default. With data from a Germane-commerce, they find that it equals or exceeds the predictive power of traditional credit bureau scores. Theirresults suggest that new technologies and new data might bring a superior ability for sc

23、reening borrowers.FinTechs are also competing in the market for wealth management. The United States is the leading marketfor robo-advisors. In 2017, it accounted for more than half of all investments in robo-advisors (Abraham et al.,2019). Nevertheless, the amount of assets managed by robo-advisors

24、 is still a small portion of total assets undermanagement, with average client wealth much smaller than the average in the industry (Economist, 2017). Abrahamet al. (2019) argues that because they save on fixed costs (such as salaries of financial advisors or maintenance ofphysical offices), robo-ad

25、visors can reduce minimum investment requirements and charge lower fees.1(In)efficiency of the Existing SystemThe main finding in Philippon (2015) is that the unit cost of financial intermediation in the U.S. has remainedaround 200 basis points for the past 130 years. Improvements in information tec

26、hnologies have not been passedthrough to the end users of financial services. This section offers an update of this work.1.1Financial Expenses and Intermediated AssetsTo organize the discussion I use a simple model economy consisting of households, a non-financial business sector,and a financial int

27、ermediation sector. The details of the model are in the Appendix. The income share of finance,shown in Figure 1, is defined as1yftyt=Value Added of Finance IndustryGDP.1Philippon (2015) discusses various issues of measurement.Conceptually, the best measure is value added, which is the sum ofprofits

28、and wages. Whenever possible, I therefore use the GDP share of the finance industry, i.e., the nominal value added of the financeindustry divided by the nominal GDP of the U.S. economy. One issue, however, is that before 1945 profits are not always properlymeasured and value added is not available.

29、As an alternative measure I then use the labor compensation share of the finance industry,i.e., the compensation of all employees of the finance industry divided by the compensation of all employees in the U.S. economy.Philippon (2015) also explains the robustness of the main findings to large chang

30、es in government spending (because of wars), the riseof services (finance as a share of services displays a similar pattern to the one presented here), globalization (netting out imports andexports of financial services).3The model assumes that financial services are produced under constant returns

31、to scale. The income of the financeindustry yftis then given byyft= c,tbc,t+ m,tmt+ k,tkt,(1)where bc,tis consumer credit, mtare assets providing liquidity services, and ktis the value of intermediated corporateassets. The parameters i,ts are the unit cost of intermediation, pinned down by the inter

32、mediation technology.The model therefore says that the income of the finance industry is proportional to the quantity of intermediatedassets, properly defined. The model predicts no income effect, i.e., no tendency for the finance income share to growwith per-capita GDP. This does not mean that the

33、finance income share should be constant, since the ratio of assetsto GDP can change. But it says that the income share does not grow mechanically with total factor productivity.This is consistent with the historical evidence.2Measuring intermediated assets is complicated because these assets are het

34、erogenous. As far as corporate financeis concerned, the model is fundamentally a user cost model. Improvements in corporate finance (a decrease in k)lower the user cost of capital and increase the capital stock, which, from a theoretical perspective, should includeall intangible investments and shou

35、ld be measured at market value. A significant part of the growth of the financeindustry over the past 30 years is linked to household credit. The model provides a simple way to model householdfinance. The model also incorporates liquidity services provided by specific liabilities (deposits, checking

36、 accounts,some form of repurchase agreements) issued by financial intermediaries. One can always write the RHS of (1) asc,t?bc,t+m,tc,tmt+k,tc,tkt?. Philippon (2015) finds that the ratiosm,tc,tandk,tc,tare close to one.3As a resultone can define intermediated assets asqt bc,t+ mt+ kt.(2)The principl

37、e is to measure the instruments on the balance sheets of non-financial users, households and non-financial firms. This is the correct way to do the accounting, rather than looking at the balance sheet of financialintermediaries. After aggregating the various types of credit, equity issuances and liq

38、uid assets into one measure, Iobtain the quantity of financial assets intermediated by the financial sector for the non-financial sector, displayedin Figure 1.1.2Unit Cost and Quality AdjustmentsI can then divide the income of the finance industry by the quantity of intermediated assets to obtain a

39、measure ofthe unit costtyftqt.(3)2The fact that the finance share of GDP is the same in 1925 and in 1980 makes is already clear that there is no mechanical relationshipbetween GDP per capita and the finance income share. Similarly, Bickenbach et al. (2009) show that the income share of finance hasre

40、mained remarkably constant in Germany over the past 30 years. More precisely, using KLEMS for Europe (see OMahony and Timmer(2009) one can see that the finance share in Germany was 4.3% in 1980, 4.68% in 1990, 4.19% in 2000, and 4.47% in 2006.3This is true most of the time, but not when quality adju

41、stments are too large.Philippon (2015) provides calibrated qualityadjustments for the U.S. financial system.4Figure 1: Finance Income and Intermediated Assets1234Intermediated Assets/GDP.02.04.06.08Share of GDP18801900192019401960198020002020year.Share of GDPIntermediated Assets/GDPNotes: Both serie

42、s are expressed as a share of GDP. Finance Income is the domestic income of the finance and insurance industries, i.e.,aggregate income minus net exports. Intermediated Assets include debt and equity issued by non financial firms, household debt, and variousassets providing liquidity services. Data

43、range for Intermediated Assets is 1886 - 2012. See Philippon (2015) for historical sources and detailsabout the underlying data.Figure 2 shows that this unit cost is around 2% and relatively stable over time. In other words, I estimate that itcosts two cents per year to create and maintain one dolla

44、r of intermediated financial asset. Equivalently, the annualrate of return of savers is on average 2 percentage points below the funding cost of borrowers. The updated seriesare similar to the ones in the original paper. The unit costs for other countries are estimated by Bazot (2013) whofinds conve

45、rgence to US levels.Figure 2: Unit Cost of Financial Intermediation0.005.01.015.02.025.0318801900192019401960198020002020time2012 DataNew DataUnit CostNotes: The raw measure is the ratio of finance income to intermediated assets, displayed in Figure 1. The 2012 data is from Philippon (2015),while th

46、e new data was accessed in May 2016. Data range is 1886 - 2015.The raw measure of Figure 2, however, does not take into account changes in the characteristics of borrowers.These changes require quality adjustments to the raw measure of intermediated assets. For instance, corporate5finance involves i

47、ssuing commercial paper for blue chip companies as well as raising equity for high-technology start-ups. The monitoring requirements per dollar intermediated are clearly different in these two activities. Similarly,with household finance, it is more expensive to lend to poor households than to wealt

48、hy ones, and relatively poorhouseholds have gained access to credit in recent years.4Measurement problems arise when the mix of high- andlow-quality borrowers changes over time.Following Philippon (2015), I then perform a quality adjustment to the intermediated assets series. Figure 3shows the quali

49、ty adjusted unit cost series. It is lower than the unadjusted series by construction since qualityadjusted assets are (weakly) larger than raw intermediated assets. The gap between the two series grows whenthere is entry of new firms, and/or when there is credit expansion at the extensive margin (i.

50、e., new borrowers).Even with the adjusted series, however, we do not see a significant decrease in the unit cost of intermediation overtime.Figure 3: Unit Cost and Quality Adjustment0.005.01.015.02.025.0318801900192019401960198020002020timeRawQuality AdjustedUnit Cost, with Quality AdjustmentNotes:

51、The quality adjusted measure takes into account changes in firms and households characteristics. Data range is 1886 - 2015.As I have argued in the past, the puzzle is why we have not seen substantial productivity gains in financialintermediation. The good news is that, however late, these improvemen

52、ts might be happening now.2A Simple Model of Robo AdvisingI consider a simple model of imperfect competition in asset management services. The model emphasizes the roleof technology, and fixed costs in particular. The key point is that there are two types of fixed costs: fixed costs toset up a busin

53、ess or a system or a platform; and then fixed cost per relationship with each client.4Using the Survey of Consumer Finances, Moore and Palumbo (2010) document that between 1989 and 2007 the fraction of householdswith positive debt balances increases from 72% to 77%. This increase is concentrated at

54、the bottom of the income distribution. Forhouseholds in the 0-40 percentiles of income, the fraction with some debt outstanding goes from 53% to 61% between 1989 and 2007.In the mortgage market, Mayer and Pence (2008) show that subprime originations account for 15% to 20% of all HMDA originationsin

55、2005.6There is a continuum of mass 1 of households whose wealth w is distributed according to the (cumulative)distribution G(w). Households are risk neutral (or, equivalently, returns are risk adjusted) and have access to aninvestment technology with gross return r. The reservation utility of a hous

56、ehold is thus rw. Households also havethe option to hire an asset manager in order to earn higher returns.2.1Traditional Asset Management EquilibriumThere are N asset managers with access to an investment technology with return R r. To be active they needto pay the fixed cost (per active firm). To w

57、ork with a household they need to pay the relationship cost (perclient). The asset management industry is oligopolistic and asset managers charge a fee f (w) to their clients. Irestrict attention to linear fees of the form f (w) = + w, where the intercept covers the fixed cost and where is a markup.

58、 For now the markup is simply a parameter but, as discussed later, it could be a decreasing functionof the number of active intermediaries. The parameters are such that rw, which happens if and only ifw wo=R r .(4)A fraction 1 G(wo) of households hire intermediation services, while the remaining fra

59、ction G(wo) invest bythemselves at the low rate. I consider a symmetric equilibrium where intermediaries have the same number ofclients. The net profit of any intermediary is therefore (N) NZwowdG(w),where wois defined in equation (4). Note that N should be interpreted as the number of asset manager

60、s per capitasince the population is normalized to one. Finally, free entry requires (N) ,with equality if entry is positive. This pins down the number of firms entering the market.Definition 1. Given the cost structure (,), an equilibrium with positive entry solves wofrom equation (4) andN =ZwowdG(w

61、).Welfare is given by the following expression:7W =Zwo0rwdG(w) +Zwo(Rw )dG(w) N=Zwo0rwdG(w) +Zwo(R )w )dG(w)Let us briefly discuss the first best allocation. The planners solution would set N = 1 to save on fixed entry costs,and = 0 to price at marginal cost. This implies wo=Rr. In this equilibrium

62、all asset managers loose money sothe planner would need to impose lump-sum taxes on households to subsidize the financial intermediaries. Assetmanagement clearly improves welfare compared to a situation where all households earn the low rate r, but, asexpected, the decentralized equilibrium does not

63、 achieve the planners outcome because asset managers need tocharge a markup to cover their entry cost, and wo wo.2.2Robo Advisors: A Tale of Two Fixed CostsLet us now introduce a new asset management technology, characterized by stronger returns to scale. More precisely,we assume that robo advisors

64、have access to the investment technology r but a lower cost per client ), but then it can easily manage a large number of clients ( ). In the limit, we can even imagine that= 0. The existing literature has failed to capture the difference between these two types of fixed costs. I willshow that they

65、have different welfare implications.Households now have three options. They can invest by themselves and earn r. They can hire a traditionalmanager.Or they can hire a robo advisor.There are now two cutoffs to calculate w1and w2, one for theparticipation decision (as before) and one for the choice of

66、 the type of asset manager. I assume that both types ofasset managers charge a markup , so robo advisory fees are +w while traditional management fees are +w.Since R R and we know that relatively poor households will choose between autarky and robos. Hencethe first participation cutoffisw1=R r .(5)T

67、he second cutoffis between the robo and traditional advisors: (R )w2 = (R )w2 . Hencew2= R R(6)One can relatively easily accommodate R= R if we introduce horizontal or vertical differentiation between advisorsas in Pagnotta and Philippon (2018), or search costs between households and advisors as in

68、Pedersen (2015) andGarleanu and Pedersen (2018). For simplicity I consider here the case R R that ensures an interior solutionwithout getting lost in the details of strategic interactions between various types of advisors. This could also capture8a convenience yield that investor perceive from inter

69、action with a human being instead of a robo. Finally, it isconsistent with the stylized fact that very high net worth households still use mostly human advisors. The conditionfor profitable entry by robots is w1 R + ( )(r + ).We can now characterize the equilibrium when both types of advisors are ac

70、tive.Proposition 1. Under the condition of Lemma 1, the equilibrium with robo advisors is characterized by the cutoffs(w1,w2) in equations (5,6) such that G(w1) poor households save by themselves, G(w2) G(w1) middle-classhouseholds hire robo advisors, and 1 G(w2) rich households hire traditional man

71、agers.The number of roboadvisors isN0=Zw2w1wdG(w),and the number of traditional managers isN1=Zw2wdG(w).Let us now analyse the welfare implication of robo advisors. Welfare is given byW=Zw10rwdG(w) +Zw2w1(Rw )dG(w) N0+Zw2(Rw )dG(w) N1,=Zw00rwdG(w) +Zw2w1(R )w )dG(w) +Zw2(R )w )dG(w).Compared to the

72、planners allocation the participation cutoffw1is still distorted by the markup exactly as in thetraditional equilibrium. The second cutoffw2, however, is the same one that the planner would choose as one canreadily see from equation (6). There is thus no distortion at the robo/traditional advisor fr

73、ontier.The following proposition establishes the key result that if robo entry is profitable , i.e., if we see any roboadvising at all, then there are more households using asset management services in the Fintech equilibrium than inthe traditional equilibrium.Proposition 2. Under the condition of L

74、emma 1, robo advising improves access to asset management services,i.e., w1 w0.Proof. We need to show that w1=Rr w0=Rr. This is true as long as R ( )(r + ) doesnot matter because in equilibrium, the rich subsidize the poor. The rich pay the lions share of fees that serve to9cover the fixed cost of s

75、etting up the robo. Once this cost is paid, poor households benefit from cheaper services.An important lesson here is that the nature of fixed costs matter a great deal for welfare. The welfare properties offixed “coding” costs are fundamentally different from those of fixed costs per client.The fac

76、t that Fintech increases participation does not mean that Fintech reduces inequality among all groups,however. If we think that the cutoffwoused to be above the middle class and that the cutoffw1is now below themiddle class, then inequality between the poor and the middle class increases as the midd

77、le class joins the rich inhaving access to asset management services. Education may also become a stronger determinant of participationthan before if education allows households to gain confidence about the features of the new technology.3Big Data and DiscriminationLet me now discuss the consequence

78、s of the increasing use of “non traditional” data in consumer credit. As in thecase of robo-advising, I describe a simple model and then I ask how Fintech might affect the equilibrium. The basicmodel follows the classic paper of Aigner and Cain (1977). There is a one-dimensional variable q that capt

79、uresthe credit quality of an individual. The distribution of q is normal with mean q and variance 2. I use bars todenote population averages and also, under rational expectations, unconditional beliefs. It is convenient to workwith precisions instead of variances so I define 1 2.All lenders observe

80、a quantitative signal y1about the credit quality of the borrower:y1= q + 1,where 1is normal with mean 0 and precision 1. We can think of y1as a standard credit score. After observingthe signal y1, the conditional distribution of q is normal with meanEq | y1 = q + 1y1 + 1and precision +1. Now imagine

81、 that the population includes two groups, denoted by z: a majority group z = A,and a minority group z = B. If a lender can condition on group membership its conditional expectation becomesEq | y1,z = qz+ 1zy1 + 1z.I assume that the minority group has a weakly lower average credit quality, qB qA. For

82、 simplicity I assume that A= Bis the same for both groups but it is easy to allow for different population variances. A classic argumentin the literature on discrimination is that standard data favor the majority, in the sense that 1B 1A. The lesser10precision of traditional signals could arise from

83、 a higher prevalence of missing data in the minority. For instance,if they are less likely to be employed, minorities have fewer payroll statements. If they are less likely to get a loan,they have shorter histories of repayments.Discrimination can take several forms. Statistical discrimination refer

84、s to the fact that, as long as the signal y1is noisy, the posterior puts a positive weight on the prior. As a consequence, for a given signal y1, a member ofthe majority will receive a better score than a member of the minority.5Under statistical discrimination, however,there is no average bias sinc

85、e the average of a conditional expectation is correct and equal to the true populationmean: EEq | y1,z = qzfor all z. Aigner and Cain (1977) and others have argued that statistical discriminationis unlikely to explain all or even most of the discrimination that we observe empirically. Dobbie et al.

86、(2018)argue that the misalignment of incentives inside the firm can also lead to biases. They find that the short-termincentives given to loan officers create a long-run bias against immigrants. I therefore consider the stronger issue ofprejudice in lending decisions.3.1Traditional LendersTraditiona

87、l lenders meet face-to-face with borrowers. This has two consequences: they observe directly the type z,and they generate another signal u = q+uabout credit quality with precision u. With two signals the conditionaldistribution of q is normal with meanEq | y1,z = qz+ uu + 1zy1 + u+ 1z,and precision

88、+ 1+ u. I assume that loan officers have a biased belief about the minority. They perceive theaverage quality to be qB= qB . We can think of as arising from prejudice or negative stereotyping. Theconditional expectation of the lender is thenE(y1,u,B) = ( qB ) + uu + 1By1 + u+ 1B.Note that the averag

89、e statistical bias is simply EhE(y1,u,B) | Bi qB= +u+1B. The harm potentiallyimposed on the minority is stronger if the prejudice is large and if the signal is noisy (1Bis small).6RegulationsRegulators often impose constraints on the use of group membership in lending decisions.Forinstance it is ill

90、egal in the U.S. to make lending decisions based on race or gender. We can capture this idea byassuming that regulations prevent lenders from conditioning explicitly on z. The score of the traditional lender then5Suppose for simplicity that 1B= 1A. Then Eq | y1,A Eq | y1,B =q( qA qB)q+1.6I consider

91、for simplicity the case of risk neutral lenders but it is worth pointing out that the results only get stronger if lenders are riskaverse. Risk averse lenders dislike lending to minorities because the residual conditional variance is higher. If we assume for example thatlenders have mean-variance pr

92、eferences and hence care about the score per unit of uncertainty Eq | y1,z/Vq | y1,z = qz+uu+1zy1,then we see that a low value of 1zdirectly hurts the minority.11becomesT (y1,u;B) = ( q ) + uu + 1y1 + u+ 1,which does not depend explicitly on z = B. The subjective bias remains and on average we haveE

93、T (y1,u;B) | B = qB ? d? + u+ 1,whered q qBreflects the regulatory constraint that prevents statistical discrimination based on group status.Consistent with the empirical literature I assume that the negative bias has not been fully eliminated by regulations,i.e., 0 d u, in which case the term is ne

94、gative. Because Fintechlending is more precise it puts less weight on priors and it reduces the regulatory subsidy that the minority enjoys.Notice however that the sum of the two effects, d2u +1+2, is always positive. This leads us to the followingPropositionProposition 3. Unbiased univariate Fintec

95、h lending reduces biases arising from prejudice against the minority.Even though it can undermine the effectiveness of existing policies the net effect is alway positive for the minority.The proposition highlights the fundamental benefits of new lending technologies which come from two sources:no pr

96、ejudice and more precise signals. Note that the higher precision of the signal would be helpful in any case,even if it was used by traditional lenders.Remark 1. Even in the absence of Fintech lenders, giving traditional lenders access to the signal y2decreases biasesagainst the minority.If tradition

97、al lenders also observe y2they will form the posterior ( q)+uu+1y1+2y2 +u+1+2and the bias will bereduced to (d) +u+1+2. If 2is high the bias becomes small. We turn next to the case where big data containsmultiple signals.3.3Multivariate Unbiased Fintech LendingThe model above does not capture one pe

98、rvasive worry that policy makers have about big data, namely that itcan inadvertently discover proxies for group membership. As Barocas and Selbst (2016) write “Approached withoutcare, data mining can reproduce existing patterns of discrimination, inherit the prejudice of prior decision makers,or si

99、mply reflect the widespread biases that persist in society. It can even have the perverse result of exacerbatingexisting inequalities by suggesting that historically disadvantaged groups actually deserve less favorable treatment.”Let us then assume that, in addition to y2as a proxy for q, big data a

100、nalysis also generates a proxy for z.The important point here is that this second signal is a byproduct of the machine learning approach. The BDMLsystem is looking for information, and to the extent that z is informative and that proxies for z can be constructedin the data, then the system will find

101、 it and use it. I assume that the information takes the form of a signal s aboutz. It is easy to model an imperfect signal but the intuition is the same if the signal is perfect and the derivationis simpler. I therefore assume that the ML system constructs a perfect, albeit indirect proxy for z. The

102、 Fintechlenders conditional expectation is thenEq | y1,y2;B = qB+ 1y1+ 2y2 + 1+ 2which is now an unbiased estimator of qB. Therefore Eq | y1,y2,B = qB ET (y1,u;B) | B.13Proposition 4. Even if Big Data and Machine Learning lead to indirect proxies for group membership, Fintechlending still reduces bi

103、ases against minorities.Even in the most extreme case, unbiased Fintech lending produces pure statistical discrimination. To the extentthat minorities suffered from prejudice in the pre-fintech era, their welfare should increase when Fintech lendersenter a market. This is consistent with the evidenc

104、e in Bartlett et al. (2018) who find that Fintech algorithmsdiscriminate 40% less than face-to-face lenders.These results assume that 0 d , i.e., that policy did not over-shoot its targets and some negative biasremained before the introduction of the new lending technology. While this seems to be em

105、pirically plausible onaverage, it does not rule out the possibility that in some location j the minority is doing so poorly thatdj qj qBjis larger than , even though the opposite is true on average across all locations. In that particular location creditmarket regulations lead to positive redistribu

106、tion. Since the use of big data makes the credit market more neutral,it reduces redistribution and it could potentially hurt the minorities in some specific locations.Broadly speaking, however, the analysis suggests that Fintech lending is likely to reduce discrimination as long asthe algorithms the

107、mselves do not suffer from prejudice. Let us therefore consider next the case of biased algorithms.3.4Biased Multivariate Fintech lendingSuppose that Fintech engineers suffer from the same prejudice as loan officers and export their bias into theiralgorithms. Once the BDML system constructs a proxy

108、for z it applies a penalty to the true quality of minorityborrowers just as face-to-face officers did. This is arguably an extreme and unrealistic assumption because it ismore difficult to induce biases in an algorithm than during face-to-face meetings, and the evidence in Bartlett et al.(2018) sugg

109、ests that algorithms discriminate less. But this extreme assumption is helpful to make the point andhighlight the key issue. The conditional and now biased expectation becomesEq | y1,y2,B = ( qB ) + 1y1+ 2y2 + 1+ 2,and on average we haveEhEq | y1,y2,B | Bi= qB + 1+ 2.In this example Fintech lenders

110、also suffer from prejudice. If we compare with traditional lenders we getEhEq | y1,y2,B | Bi ET (y1,u;B) | B = ? d? + u+ 1 + 1+ 2,= + u+ 1?2 u + 1+ 2d?.Even thoughd the term 2u +1+2d can be negative is 2is small: Fintech lending can now potentially harmthe minority. On the other hand, if 2is signifi

111、cantly larger than uFintech lending helps the minority despite its14prejudice. Why is that? As before, Fintech lending increases precision. The impact of the prejudice depends onthe precision of the credit scoring signal y2. Even if there is prejudice, its impact is small if 2is large because theBay

112、esian part of the mechanism carries more weight.Proposition 5. Biased multivariate Fintech lending can decrease the welfare of the minority, but this becomesincreasingly less likely as Fintech algorithms become more precise.There are two keys lessons here. The first lesson is that for Fintech to hur

113、t the minority two issues must interact:(i) BDML needs to build a direct proxy for group membership; and (ii) the algorithm itself must contain prejudice.If only one issue is present, Fintech improves welfare for the minority. It is only in the biased multivariate case thatFintech can be detrimental

114、. The second lesson is that even in the pessimistic case, when Fintech lending becomesprecise enough it always improves welfare for the minority. Even when engineers suffer from prejudice and somehowembed this prejudice into their algorithms, the prejudice decreases with the precision of the credit

115、scoring model.4ConclusionFintech is likely to decrease the costs of financial intermediation, but also to create new regulatory issues. Inthis short note I have highlighted two important forces. In the case of robo-advisors, I have argued that the newpattern of fixed costs is likely to improve parti

116、cipation by relatively less wealthy household. This may not lowerinequality across all groups, however. In the credit market, I have argued that alternate data sources are likely toreduce non-statistical discrimination. To the extent that minorities were hurt by prejudice or negative stereotyping,th

117、ese minorities should gain from the use of alternate data sources. On the other hand new data can reduce theeffectiveness of existing regulations.15ReferencesAbraham, F., S. L. Schmukler, and J. Tessada (2019). Robo-advisors: Investing through machines. World BankPolicy Research Working Paper (13488

118、1).Aigner, D. J. and G. G. Cain (1977). Statistical theories of discrimination in labor markets. ILR Review 30(2),175187.Barocas, S. and A. Selbst (2016). Big datas disparate impact. California Law Review 104, 671732.Bartlett, R., A. Morse, R. Stanton, and N. Wallace (2018). Consumer-lending discrim

119、ination in the era of fintech.Working paper.Bazot, G. (2013). Financial consumption and the cost of finance: Measuring financial efficiency in europe (1950-2007). Working Paper Paris School of Economics.Berg, T., V. Burg, A. Gombovi, and M. Puri (2019). On the rise of fintechs credit scoring using d

120、igital footprints.Working paper.Bickenbach, F., E. Bode, D. Dohse, A. Hanley, and R. Schweickert (2009, October). Adjustment after the crisis:Will the financial sector shrink? Kiel Policy Brief.Buchak, G., G. Matvos, T. Piskorski, and A. Seru (2018). Fintech, regulatory arbitrage, and the rise of sh

121、adowbanks. Journal of Financial Economics 130(3), 453 483.Dobbie, W., A. Liberman, D. Paravisini, and V. Pathania (2018). Measuring bias in consumer lending.Economist, T. (2017). Silicon speculators.Fuster, A., M. Plosser, P. Schnabl, and J. Vickery (2019). The role of technology in mortgage lending

122、. The Reviewof Financial Studies 32(5), 18541899.Garleanu, N. and L. H. Pedersen (2018). Efficiently inefficient markets for assets and asset management. TheJournal of Finance 73(4), 16631712.Mayer, C. and K. Pence (2008). Subprime mortgages: What, where, and to whom? StaffPaper Federal ReserveBoard

123、.Moore, K. B. and M. G. Palumbo (2010, June). The finances of american households in the past three recessions:Evidence from the survey of consumer finances. StaffPaper Federal Reserve Board.OMahony, M. and M. P. Timmer (2009). Output, input and productivity measures at the industry level: The eukle

124、ms database. The Economic Journal 119(538), F374F403.Pagnotta, E. and T. Philippon (2018, May). Competing on speed. Econometrica 86.Pedersen, L. H. (2015). Efficiently Inefficient: How Smart Money Invests and Market Prices Are Determined.Princeton University Press.Philippon, T. (2015). Has the us fi

125、nance industry become less efficient? on the theory and measurement of financialintermediation. The American Economic Review 105(4), 140838.Philippon, T. (2016). The fintech opportunity. NBER Working Paper.16Previous volumes in this series 840 February 2020 Operational and cyber risks in the financi

126、al sector Iaki Aldasoro, Leonardo Gambacorta, Paolo Giudici and Thomas Leach 839 February 2020 Corporate investment and the exchange rate: The financial channel Ryan Banerjee, Boris Hofmann and Aaron Mehrotra 838 January 2020 The economic forces driving fintech adoption across countries Jon Frost 83

127、7 January 2020 Bad bank resolutions and bank lending Michael Brei, Leonardo Gambacorta, Marcella Lucchetta and Bruno Maria Parigi 836 January 2020 FX spot and swap market liquidity spillovers Ingomar Krohn and Vladyslav Sushko 835 December 2019 The Cost of Steering in Financial Markets: Evidence fro

128、m the Mortgage Market Leonardo Gambacorta, Luigi Guiso, Paolo Emilio Mistrulli, Andrea Pozzi and Anton Tsoy 834 December 2019 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 83

129、3 December 2019 Central Counterparty Exposure in Stressed Markets Wenqian Huang, Albert J. Menkveld and Shihao Yu 832 December 2019 Hedger of Last Resort: Evidence from Brazilian FX Interventions, Local Credit and Global Financial Cycles Rodrigo Barbone Gonzalez, Dmitry Khametshin, Jos-Luis Peydr an

130、d Andrea Polo 831 December 2019 Believing in bail-in? Market discipline and the pricing of bail-in bonds Ulf Lewrick, Jos Maria Serena and Grant Turner 830 December 2019 De jure benchmark bonds Eli Remolona and James Yetman 829 December 2019 Central banking in challenging times Claudio Borio 828 December 2019 The currency composition of foreign exchange reserves Hiro Ito, Robert N McCauley 827 December 2019 Bank Loan Supply during Crises: The Importance of Geographic Diversification Sebastian Doerr, Philipp Schaz All volumes are available on our website www.bis.org.