Big Data in Credit Scoring, Are We Ready?

(1)

Big Data in Credit Scoring, Are We Ready?

Alternative Credit Scores and the Risks That They Potentially Pose

Levan Lobzhanidze 11015551 Master of Law and Finance Supervisor: Mia Junuzović 24th July 2020

(2)

Abstract:

Since the 1930’s credit scoring has gone under multiple major changes. One of them being, in 1989 with the use of algorithms, the Fair and Isaac Corporation created predictive scoring, which has been the root of credit scoring ever since. However, another big change arguably came in the early 2010’s, when companies started utilizing big data in their credit scores. Thus, instead of the companies using the scores based on the five components that were previously commonplace, they went further by using all data that they can manage to amass. This brought huge change to the underwriting process, and made credit scoring easier for both sides – the customer and the company. For customers, the new approach allowed individuals to receive more accurate scores, and even allowed some people that were

previously considered ‘unscorable’ to the loan markets. While for the companies, it decreased default rates, and reduced both costs and the number of losses.

However, given that these companies utilize ‘big-data’, and the fact that legislation was not meant to encompass alternative scoring companies, there are some issues that could potentially materialize. Namely, the initial issues are with regards to transparency and regulatory arbitrage. Meaning, these companies are potentially not transparent enough for lawmakers and customers to understand and seek recourse against. While this further

facilitates regulatory arbitrage, in that, if the companies are too intricate to be understood by outsiders, there might be issues in applying legislation. Furthermore, these two issues might lead to the wider societal issue that is discrimination, in that, they might allow it (through regulatory arbitrage) and mask it (through non-transparency).

And while there currently exists regulation that should control credit scoring

companies (i.e. the FCRA and ECOA), the anatomy of alternative credit scoring companies might allow to side-step legislation without repercussions. Thus, it is important to identify the specific shortcomings of current legislation and propose a solution that will fill the gaps.

The solution that this master thesis proposes is regulatory sandboxes, due to their flexible character and the constant dialogue between the company and the regulator. This approach could potentially decrease the uncertainty on both ends, in turn, contributing to efficient regulation which does not stunt economic growth, and at the same time promotes industry compliance.

(3)

Abbreviations:

• FICO – Fair and Isaac Corporation • ML – Machine Learning

• CEO – Chief Executive Officer • FCRA – Fair Credit Reporting Act • ECOA – Equal Credit Opportunity Act • CRA – Consumer Reporting Agency • FTC – Federal Trade Commission

• OECD – Organisation for Economic Co-operation and Development • FAA – Federal Aviation Administration

• FCA – Financial Conduct Authority

• CFPB – Consumer Financial Protection Bureau • CAS – Compliance Assistance Sandbox

(4)

Introduction

With the onset of digitalization and the internet age, many industries have evolved and are using data in one way or another1. More specifically, some of them use ‘big data’2, and with the help of algorithms3 and Machine Learning4, they utilize the said data in a way that helps ease tasks, or at times even allows them to draw conclusions that were previously impossible5. It has been argued that the use of these methods make companies more exact and efficient6. This brings us to the industry being analyzed in this master thesis – the Credit Scoring industry – which is utilizing data with the same objectives in mind.

Since around the early 2010’s, credit scoring agencies have been undergoing big changes7. Start-ups that do Alternative Credit Scoring are popping up, however, compared to the original companies, the approaches differ quite a bit8. Even though the customers

remained the same – namely, individuals that need credit scores and banks that need to know the creditworthiness of the individual – the way they yield credit scores has become more intricate to say the least9_{. Naturally, the problems that accompany typical data aggregating} companies are present (e.g. Google)10_{, however, what makes these companies stand out is} clout and the influence they can exert on individual lives. This influence is exactly why it is important to examine these companies and determine if they pose threats to individual liberties. And it is wholly important to analyze these companies in common law countries

1_{Michael Belfiore, “How 10 Industries Are Using Big Data to Win Big” (2019)}

<https://www.ibm.com/blogs/watson/2016/07/10-industries-using-big-data-win-big/>; accessed July 19, 2020 2_{“Big data is high-volume, high-velocity and/or high-variety information assets that demand cost-effective,} innovative forms of information processing that enable enhanced insight, decision making, and process automation”.

See: Gartner Inc., “Big Data” (Gartner)

<https://www.gartner.com/en/information-technology/glossary/big-data> accessed July 19, 2020

3_{An algorithm is: “… is a set of precise steps that […], if obeyed exactly and mechanically, will lead to some} desirable outcome. Long division and column addition are examples […]”

See: T.C, “What Are Algorithms?” (2017) The Economist

<https://www.economist.com/the-economist-explains/2017/08/29/what-are-algorithms>; accessed July 19, 2020

4_{“Machine-learning algorithms use statistics to find patterns in massive amounts of data”.}

See: Karen Hao, “What Is Machine Learning?” (2020) MIT Technology Review

<https://www.technologyreview.com/2018/11/17/103781/what-is-machine-learning-we-drew-you-another-flowchart/>; accessed July 19, 2020

5_ibid.

6_{Anna Oleksiuk, “How Alternative Credit Data Can Increase Accuracy in Credit Scoring” (Intellias 22 August,} 2019) <https://www.intellias.com/how-alternative-credit-data-can-increase-accuracy-in-credit-scoring/>; accessed July 21, 2020

7_{id:analytics, ‘Alternative Credit Scores: The Key to Financial Inclusion for Consumers’ (2017) Whitepaper, 4} 8_{Mikella Hurley and Julius Adebayo, ‘Credit scoring in the era of big data’ (2016) 18 Yale JL & Tech,} 146-216; 151

9_{ibid. 151-152}

(6)

where they are prevalent, and more so in the United States, where issues like Redlining have already plagued communities12_.

Thus, the potential issues that can be identified are: regulatory arbitrage, issues with

transparency, and most importantly, discrimination.

Big data has the potential to enhance inequality in manipulable knowledge, which, leaves the public at the mercy of the entities that hold the data13. And in this case, due to the shortcomings of law that allows these companies to abuse loopholes, alternative credit scoring could effectively be shielded from reprisal14. To elaborate, the way that current legislation stands in the United States allows these companies to gather and use as much data as they wish, while not mandating transparency on a level where individuals can seek

recourse to adverse decisions15.

Thus, the starting point of cooperation should be transparency, which, at this point is meagre – simply due to these companies being secretive or hard to understand16. If there is transparency, there is lower chance of regulatory arbitrage, which, based on the way the current legislation is structured, is not the case17. As well as further issues potentially encountered being discrimination, like biased (or incorrectly gathered) data18_{. Thus, due to} the multitude of concerns and the relatively new character of these companies, it is necessary to inspect them diligently and anticipate some of the issues that might materialize. But in order to do that, the potential issues need to be legitimized from legal and ethical standpoints. This brings us to the research question of this master thesis:

What are the implications for Big Data in Alternative Credit Scoring, how prepared is the legal system to dodge or absorb the potential issues, and going forward, are there more efficient alternatives to legislation?

11_{Systematically denying loans to a group of people by using entire neighbourhoods, colour-coded by perceived} risk factor, as a decision-making metric.

See: Jordan Pearson ‘AI Could Resurrect a Racist Housing Policy’ (2017) Vice

<https://www.vice.com/en_us/article/4x44dp/ai-could-resurrect-a-racist-housing-policy>; accessed July 20, 2020

12_ibid

13_{Nayef Al-Rodhan. ‘The social contract 2.0: Big data and the need to guarantee privacy and civil liberties.’} (2014), Harvard International Review 16, 3

14_{ibid, 3}

15_{Hurley & Adebayo (n 8) 189} 16_{ibid, 182}

17_{ibid. 195}

(7)

Thus, the structure of this master thesis will be as such, Chapter 1 will contain the

comparison of the old companies to the new, with attention being paid specifically to what

makes these companies alternative. Chapter 2 highlight the current issues being faced, moving into the implications. Chapter 3 will provide an overview of current laws that exist

in the United States to see if the legislation is currently up to the challenge of regulating

these companies. And lastly, the 4th Chapter will contain the possible solutions to fill the gaps in the legislation, whether through more legislation or alternative ways of handling new companies – with the final and proposed suggestion being Regulatory Sandboxes.

(8)

1. How Does Alternative Credit Scoring Function?

1.1 Credit Scoring

Credit scoring is not a new phenomenon, it has been around since the 1930’s starting in the mail order industry and later moved into the personal finance sector19. Essentially, these companies score individuals based on parameters and produce a credit score, which in turn, contributes to the individual’s loan bearing capability20_{. However, lately these}

companies started to utilize predictive credit scoring, based on automated analytical tools21_. The company behind this method is Fair and Isaac Corporation, which in 1989 developed an automated credit-scoring system, commonly known as FICO score22.

1.1.1 Predictive Credit Scoring

The generated score is based on 5 variables: payment history (35%), amounts owed (30%), length of credit history (15%), credit mix (10%) and new credit (10%) is used23. Subsequently, with the help of this scoring model, the company is able to give scores that decrease bias, predict possible defaults, and make it easier for individuals to change mistakes in the system24. Thus, credit scorers employ customer data when creating a predictive credit score, while most banks use these scores upon considering to grant loans25. In effect, this means that most citizens of the United States are subject to predictive scoring in receiving loans26.

With the onset of big data, the new methods of credit scoring have shifted the industry paradigm. Coined as ‘Alternative Credit Scoring’, this approach gathers far more data than the earlier one27. Usually, these companies use Algorithms and Machine Learning (ML) to combine traditional credit data with large amounts of other data to deduce a creditworthiness score28_{. And while these companies have reduced losses with lower income segments more}

19_{Noel Capon, 'Credit Scoring Systems: A Critical Analysis' (1982) 46 Journal of Marketing, 83} 20_{Nick Henry, and John Morris. ‘Scaling Up Affordable Lending: Inclusive Credit Scoring.’ (2018). 3-4} 21_{ibid, 3}

22_{Rob Kaufman, “The History of the FICO® Score”, (2018) myFICO} <https://www.myfico.com/credit-education/blog/history-of-the-fico-score>; accessed July 19, 2020

23_{FICO, 'How Are FICO Scores Calculated? | Myfico | Myfico' (Myfico.com, 2020)}

<https://www.myfico.com/credit-education/whats-in-your-credit-score> accessed 14 July 2020.

See Also: Robinson + Yu, Knowing the Score: New Data, Underwriting, and Marketing in the Consumer Credit

Marketplace 15 (2014), 9 <http://www.robinsonyu.com/pdfs/Knowing_the_Score_Oct_2014_v1_1.pdf> 24_{Cathy O'Neil, ‘Weapons of Math Destruction’ (2016) Penguin Books 117}

25_{Hurley & Adebayo (n 8) 153-155} 26_ibid.

27_{ibid. 148} 28_{ibid. 158}

(9)

than 20%, while simultaneously doubling application approval rates29 – it is still nonetheless interesting to look into what type of data do they actually use and how they use it.

1.2 Alternative Credit Scores 1.2.1 Differences in Data

The phrase that the CEO of Zestfinance – the industry leader in alternative credit scoring30 – postulated is a very good starting point in assessing the differences between the old and the new method. Douglas Merrill said: “all data is credit data”31. Nonetheless, what is “all data”? It seems to be composed of two sets: 1) Baseline Credit Data: usually the data used by the FICO model32. In other words, the borrower’s finances and their bill-paying record33_{. And 2) Fringe Alternative Data (‘Alternative Data’) – which is essentially all other} data34. The latter of which is the main focus of this master thesis, while the company

Zestfinance, due to it being regarded as the industry leader, being the main reference point.

Even though ‘all other data’ sounds straightforward, it is still important to highlight what these companies mean by it. Upon circumspection, the patent of Zestfinance shows 4 sub-types of alternative data gathered:

1) Proprietary data: obtained from data brokers and ranges from something as personal as medical data to something as broad as purchase history35.

2) Public data: data available in the public domain, for example an internet search of the borrower's name with “web crawling and scraping”36.

29_{Tobias Baer, Tony Goland and Robert Schiff, ‘New Credit-Risk Models for the Unbanked’ (2013) McKinsey} & Company <https://www.mckinsey.com/business-functions/risk/our-insights/new-credit-risk-models-for-the-unbanked>; accessed July 19, 2020

30_{Business Wire, “ZestFinance to Provide AI Underwriting for Meridian Link Loans PQ Platform” (2019)} Business Wire <https://www.businesswire.com/news/home/20191017005250/en/ZestFinance-Provide-AI-Underwriting-MeridianLink-LoansPQ-Platform>; accessed July 19, 2020

31_{Quentin Hardy, 'Just The Facts. Yes, All Of Them.', Nytimes.com (2020),}

<https://www.nytimes.com/2012/03/25/business/factuals-gil-elbaz-wants-to-gather-the-data-universe.html> accessed 14 July 2020.

32_{Robinson + Yu (n 23)} 33_{ibid. 3}

34_{ibid. 10-13}

35_{Douglas Merrill, John WL Merrill, Shawn M Budde, Lingyun Gu, and James P. McGuire, J, ZESTFINANCE} Inc, System and method for building and validating a credit scoring function. (2015). U.S. Patent Application 14/276,632. 2-3

See Also: Hurley & Adebayo (n 8) 175

And: Natasha Singer, Mapping, and Sharing, the Consumer Genome, N.Y. TIMES (2012),

<http://www.nytimes.com/2012/06/17/technology/acxiom-the-quiet-giant-of-consumer-database-marketing.html?pagewanted-all&_r=0>

36_{“Web crawlers, or spiders, are programs that automatically browse and download web pages by following} hyperlinks in a methodical and automated manner” and “A web crawler is usually known for collecting web pages, but when a crawler can also perform data extraction during crawling it can be referred to as a web

(10)

3) Social network data: similar to public data, however specifically targeted at social media activity like Twitter posts, Instagram likes or Facebook friends37_.

4) Borrower’s data: data gathered while the potential borrower is applying for the loan, such as social security number, drivers license or even the time the client spent on

the terms and conditions page)38.

Once a sufficient amount of these data types are gathered, a company like Zestfinance has a substantial amount of raw data on its hands, but in order to create a single credit score, the data needs to be structured, classified, and conclusions need to be drawn from it39. Thus, it is exactly at this point where the alternative part of the new credit scoring comes into play.

1.2.2. What Makes It Alternative

An important ingredient in making the credit scores alternative is the use of algorithms and ML. How these two function are as follows:

Initially, the gathered data points (which is called ‘training data’) are translated into a scale, for example, 1 if a person has Facebook, 0 otherwise; or 0 if a person skipped the terms and conditions page, while 5 if they read it diligently40. Meaning, metavariables are created out of simple data points41. After the data-transformation process, algorithms, such as

financial and statistical ones, are used to identify important metavariables for credit scoring42_. These metavariables generate a plurality of independent decision sets describing specific aspects of the borrower, the decision sets are also called ‘input variables’43_{. For example, if a} borrower lives in a specific area but earns more than the relative area, they could get a higher score compared to the neighbours – exemplifying the agility that is employed by this

scraper”. In: Salim Khalil and Mohamed Fakir, 'Rcrawler: An R Package for Parallel Web Crawling and Scraping' (2017) 6 SoftwareX. 2-4;

See also: Merrill, and others (n 35) 3;

India Kerle, Enigma.com (2020) <https://enigma.com/blog/post/what-is-public-data> accessed 14 July 2020. 37_{Merrill, and others (n 35) 4}

See also: Hurley & Adebayo (n 8) 175

38_{Hurley & Adebayo (n 8) 175-176} 39_{ibid 174-176}

40_{Merrill, and others (n 35) 4-5}

See also: Hardy (n 31)

41_{Hurley & Adebayo (n 8) 176} 42_{ibid. 181}

(11)

method44. But algorithms are not enough to tackle this type and amount of data, mainly because with the increase of steps it has to execute, there is also an increase in complexity45_.

This is why at the next and final step, ML, a type of algorithm in itself, is used to finalize a credit score46. The ML process is essentially an optimization routine that seeks to assign suitable weighting to the most important metavariables identified47. In other words, the final model is composed of other models with a goal to finalize a credit score of a customer48. Thus, ML comes in handy because even though it alone cannot analyse unstructured data, its algorithms can process structured data and complete an objective without the need of writing a new program, and if the metavariables are uncommon, it can nevertheless make sense of them and classify them49. To reiterate, after the data is gathered and transformed, algorithms help in structuring and drawing conclusions from the data, while ML helps in producing a final, meaningful, credit score50. This is especially helpful because even though there exist some necessary data types (e.g. social security number), the rest of the data that is gathered is so random that it would be impossible to analyse every applicant with the same algorithm51.

1.3 Benefits of The New Alternative Credit Scoring

Alternative credit scoring has undoubtedly revolutionized the underwriting process52 when it concerns individual loans, and with it has brought numerous benefits53. For the

lenders, it has given them the ability to cluster and group individual segments of a population, has proven to make underwriting more efficient, decreased the default rates, and in general, reached a larger number of individuals54. This can be exemplified by the increase in the

44_{Toon Calders & Indrė Žliobaitė, “Why unbiased computational processes can lead to discriminative decision} procedures”. (2013). In Discrimination and privacy in the information society (43-57). Springer, Berlin, Heidelberg. 46-48

45_{Hurley & Adebayo (n 8) 168-169}

Also: Merrill, and others (n 35) 4-6

46_{ibid. 180} 47_{ibid. 180} 48_{ibid. 181}

49_{Berend Berendsen 'What's The Difference Between AI, ML and Algorithms?', Widget Brain (2020)} <https://widgetbrain.com/difference-between-ai-ml-algorithms/> accessed 14 July 2020.

50_{Hurley & Adebayo (n 8) 181-182} 51_{ibid. 159}

52_{“Underwriting is the process through which an individual or institution takes on financial risk for a fee. The} risk most typically involves loans, insurance, or investments”.

See: Caroline Banton, ‘Underwriting’ (2019) Investopedia

<https://www.investopedia.com/terms/u/underwriting.asp>; accessed July 20, 2020 53_{Henry & Morris (n 20). 1}

54_{Rob Aitken, “‘All Data Is Credit Data’: Constituting the Unbanked” (2017) 21 Competition & Change 274,} 285

(12)

predictive power between 10 to 20%, and in turn leading to better fraud detection . On top of an overall reduction in losses ranging from 6% to 20%56_{. And the ability to score people with} no prior history57_.

The same goes for the borrowers. It has reached a completely new market, called the ‘unscorables’, who are people that do not have a credit history, thus are unable to be scored with the traditional (i.e. FICO) style models58. Furthermore, due to this addition of raw data, the customers are not solely dependent on their borrowing history, thus making the process fairer59. Also, it has allowed for the specialization of loans, meaning that every borrower will not be clustered under the same contract – they will have terms more suitable to their

situation60. And lastly, for the communities in general, these services give greater benefits in a way that they contribute to economic prosperity and greater inclusion in the financial sector61. Thus, not only is this approach more efficient overall, but it also seems to increase social welfare.

Nonetheless, due to the fact that alternative credit scoring meshes two industries that are sensitive to abuse – big data and credit scoring62 – it is necessary to highlight some of the industry specific issues. These include: issues linked to regulatory arbitrage and transparency, as well as potential implications of this method, such as discrimination63. And lastly, due to the fact that these companies are new, it is also necessary to pinpoint the potential failures that legislation might have in addressing them, and try and identify their solutions.

55_{Stephen Jones & Craig Wellman, ‘Artificial intelligence in financial services’, (2019). UK Finance. 22} 56_{Ben Buchanan, ‘Artificial intelligence in finance. The Alan Turing Institute’, (2019). 19}

57_{Sanjoy Malik, 'Council Post: Alternative Data: The Great Equalizer to Lending Inequalities?' (Forbes, 2020)} <https://www.forbes.com/sites/forbestechcouncil/2019/08/14/alternative-data-the-great-equalizer-to-lending-inequalities/> accessed 14 July 2020.

58_{Aitken (n 54) 281}

59_{Henry & Morris (n 20). 11} 60_{ibid. 1-2}

61_{ibid. 22}

62_{Big data has issues regarding data protection. Whereas credit scoring, as the thesis will show, has potential} discrimination issues.

(13)

2. Potential Issues Arising from Alternative Credit Scoring Systems

2.1 Transparency & Regulatory Arbitrage

Whether premeditated or accidental, one of the biggest issues that might arise from alternative credit scoring is regulatory arbitrage, which, coupled with the problems of transparency, might make regulating these companies even harder.

Essentially, regulatory arbitrage is the strategy of bypassing laws due to regulatory inconsistencies64_{. Meaning, companies taking advantage of loopholes in regulation which} allows them to dodge responsibilities and restrictions, among other things65. And research has shown that when there is a arbitrage opportunity, it is usually undertaken66. Thus, besides identifying the potential social costs of these new companies, this master thesis also seeks to understand whether regulatory arbitrage is possible, and if so, how will it influence

customers. To understand the possibilities of regulatory arbitrage, the potential opportunities need to be analyzed – essentially, examining the nature of these companies and identifying the possible loopholes in legislation. Furthermore, as it will be highlighted in the forthcoming subsection, being that alternative credit scoring is at times a non-transparent activity, it could complicate the situation even further for regulators. And while the purpose of this section is to highlight potential issues, later sections will seek to analyze to what extent and in which cases does legislation allow for these problems to materialize.

2.1.1 Transparency

Currently, one of the most identifiable issues of alternative credit scoring is

transparency. Due to the fact that this technology is cutting edge, the companies try to keep

their data models secret67_{. This is logical, because the moment that their trade secrets become} public data is the moment when these companies cease to be profitable. However, this

secrecy might be accompanied with a detriment, which is at the cost of transparency. Namely, the companies keep the input variables and the weighting of their models as trade secrets, making the whole process ambiguous and unknown to the public68. This sparks a

64_{Magnus Willesson, "What Is and What Is not Regulatory Arbitrage? A Review and Syntheses." (2017)} Financial Markets, SME Financing and Emerging Economies. Palgrave Macmillan, Cham. 71-72 65_{Hurley & Adebayo (n 8) 185-199}

66_{Willesson (n 64) 71-72}

67_{Hurley & Adebayo (n 8) 195-198}

Also: Brenda Reddix-Smalls, Credit Scoring and Trade Secrecy: An Algorithmic Quagmire or How the Lack of

Transparency in Complex Financial Models Scuttled the Finance Market, 12 U.C. DAVIS BUS. L. J. 87 (2011). 94

(14)

multitude of issues. First, if the tools that these companies use to create scores are secret, then there is not much research that can be conducted on their fairness (this applies to customers and regulators alike)69_{. Second, and possibly more alarming, is the fact that the secrecy might} be detrimental for the customers, because they have no means of preparing for adverse decisions and discrimination from the creditors or the scoring companies themselves70. Furthermore, the secretive nature complicates the situation even further on the borrower’s side if they are given adverse decisions, because they cannot rectify what they do not

understand71. Meaning, not only is the individual not able to be prepared in a scenario where they are mis-rated, but they cannot even dispute their score because the data modelling tools are hidden72. This is an issue due to the importance and prevalence of these scores. In the United States, if a person needs liquidity in the form of a loan, they usually need a credit score73; and as the data shows 87% of families aged 36-44 are in debt, making debt an integral instrument in American finance74. Thus, even if these companies follow regulation protocol, the obfuscation and intricacy of their tools might bypass laws in a manner that, from a transparency standpoint, is inconvenient or even disadvantageous for the individual

customer.

2.1.2 Regulatory Arbitrage

And furthermore, regulatory arbitrage might also arise with regards to the terms and definition in the laws75_{. Namely, an alternative credit scoring company could potentially} bypass regulation if they were to anonymize the individuals credit scores76_{. At first glance,} this seems to defeat the purpose of credit scoring; however, the reality is that based on the way the data could be structured, the individual could still be traced back to the anonymized credit score77. Additionally, there could be a potential loophole within the definition of these companies, in that, if they do not supply the credit scores to third party companies, then they do not qualify as credit scoring agencies and might be exempt for regulation altogether78.

69_{Hurley & Adebayo (n 8) 179-180} 70_{ibid, 195-198}

71_{ibid, 198-200} 72_{ibid, 181 & 195}

73_{Hurley & Adebayo (n 8) 148}

74_{Bill Fay, Debt.org, (2020) <https://www.debt.org/faqs/americans-in-debt/demographics/> accessed 14 July} 2020.

75_{Hurley & Adebayo (n 8) 192} 76_{ibid. 185}

77_{ibid, 185-186} 78_{ibid, 186-187}

(15)

Moreover, the way the law is structured currently might make the customers reluctant to pursue a correction to an erroneous judgment on their credit score79_{. Mainly because of the} ambiguity of the data modelling and the burden that customers have to take on to correct an error80. And lastly, there could possibly be limitations to the definition of who the legislation protects with regards to anti-discrimination81. In other words, the definition of what and who counts under discrimination can be limited, in effect, leaving some groups unprotected82. Arguably, the issues concerning discrimination are one of the most pressing when it comes to these companies. This is due to potential discrimination happening in a variety of ways, both knowingly and unknowingly83. Which is further exacerbated by the nature of these

companies, being that they are both new and obscure84.

Thus, this amounts to lack of transparency from the companies, and the current

legislation allowing the companies to engage in regulatory arbitrage. In effect, companies can use loopholes and commit actions (willingly or as a side effect to a goal) that would

otherwise be deemed illegal and/or discriminatory85. This brings us to the following section, in which the possibilities, types, and instances of discrimination will be highlighted in order to identify the existing issues.

2.2 Discrimination: Profiling and Biased Assumptions

A potential issue that could, and has shown to, accompany profiling86 is

discrimination – they are part and parcel87. It is only natural, as classification and division of data point are necessary elements in profiling88. At times, profiling has proven to be efficient, e.g. the aforementioned example about scoring customers with no credit history89. But

nevertheless, profiling has a tendency of ending up being unfair, unethical, and even illegal (in other words discriminatory) – whether done on purpose or by neglect90. It can arise as a 79_{ibid, 190-191} 80_{bid. 198} 81_{ibid, 192} 82_{ibid, 192} 83_{ibid. 193} 84_{ibid. 149} 85_{ibid. 183-201}

86_{“The application of profiles to individuate and represent a subject or to identify a subject as a member of a} group or category”.

See: Bart Schermer, "Risks of profiling and the limits of data protection law." (2013) In Discrimination and

privacy in the information society. Springer, Berlin, Heidelberg. 138 87_{Schermer (n 86) 137}

88_{ibid. 138}

89_{Henry & Morris (n 20) 28} 90_{Schermer (n 86) 138-141}

(16)

problem in instances where people are profiled based on: ethnicity, gender, religion, or sexual preferences91_{. Thus, the goal of this section is to highlight biased (and discriminatory)}

assumptions in data modelling, all the while contrasting it with potential examples of biases in and outside of alternative credit scoring.

Usually, profiling is done through classification of data point into variables, thus, there is a risk that a person is judged based on group characteristics, rather than individual merits92. In other words, group profiles are usually made with statistics, which, could very well hold true for the group itself, but not the individual per se93. An example of this is a neighbourhood clustered as a profile, and that neighbourhood having a 20% higher default rate compared to the average94. This means that an individual’s score could be affected negatively just by residing in a certain neighbourhood95. Furthermore, not only does profiling lead to discriminatory effects, but gathering data, if not done properly, can also lead to biases.

While the main reason why data mining can lead to discrimination is computational models with biased assumptions in the first place96. It is generally assumed that the newly collected data (training data) will follow the same distribution as the reference one (historic data), meaning that the conditions will not change97_{. In other words, computational models} rely on two assumptions, 1) characteristics of the population from which the training sample is collected is the same as that of the one on which the model will be applied, and 2) the training data represents the population well98. If the former is violated, then models may fail to perform accurately99. This can be exemplified by the fact that debt repayment usually follows business cycles in an industry100. Meaning, when there is a boom in the automotive industry, the factory workers are better able to pay off their debts. Whereas during a slump, they have more trouble in doing so. A data model trained in a boom will not yield correct

91_{ibid. 138}

92_{Anton Vedder, "KDD: The challenge to individualism." (1999) Ethics and Information Technology 1, no. 4:} 275-281.

93_{Bart Custers, Data Mining with Discrimination Sensitive and Privacy Sensitive Attributes. (2010) In:} Proceedings of ISP 2010, International Conference on Information Security and Privacy, Orlando, Florida, July 12/14. 16-18

94_{Schermer (n 86) 138} 95_{Custers (n 93) 10}

96_{Calders & Žliobaitė (n 44) 46-48} 97_ibid.

98_{ibid. 46}

99_{Mark G. Kelly, David J. Hand, & Niall M. Adams, , The impact of changing populations on classifier} performance. (1999) In Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining. 367-371

100_{15 U.S.C. § 1681m(a) (2012);}

See also: Fair Credit Reporting § 2.3.4.1 (8th ed. 2013), National Consumer Law Center, www.nclc.org/library

[https://perma.cc/8RKG-JHGW]; see also Trans Union Corp. v. Fed. Trade Comm'n, 245 F.3d 809 (D.C. Cir. 2001). § 3.3.6.

(17)

results while analysing a slump101. Whereas for the latter assumption to be satisfied, the training set and the original need to coincide in the variables to a high degree102_{. This means} that variables like: age distribution, male to female ratio and high versus low skilled labour need to have proportional shares103. If the assumption is violated, then the training data does not match the original, and the new model will perform sub-optimally as compared to the original104. Thus, if the two assumptions are satisfied, the data models transferred from the original data set will also transfer the knowledge correctly105. While if the original data is gathered or applied incorrectly, the new models will yield biased results106. There are three scenarios under which the two assumptions can be violated: incorrect labels, sampling bias and incomplete data.

2.2.1 Incorrect Labels

Labels are incorrect when they are inappropriate relative to current times107. One of the main reasons behind this stems from historical discrimination through biased decision-making108. “Sample selection bias exists when, instead of simply missing information on characteristics important to the process under study, the researcher is also systematically missing subjects whose characteristics vary from those of the individuals represented in the data”109. This can happen when certain groups are no longer excluded from certain jobs110, e.g. title of CEO111_{. In a historical context, women have not often occupied that position, and} if being a CEO is a criteria for granting a loan, the algorithm might erroneously conclude that women should not be approved for these loans112_{. The same goes for label changes in time,} meaning that the aforementioned CEO example needs to be accounted for in the data – otherwise, the algorithm can make assumptions today, about conditions that are no longer presently satisfied113.

101_{Calders & Žliobaitė (n 44) 46} 102_{ibid. 46-48}

103_{ibid. 46-48}

104_{Bianca Zadrozny, Learning and Evaluating Classifiers under Sample Selection Bias. (2004) In: Pro- ceedings} of the 21st International Conference on Machine Learning (ICML 2004), 905-909

105_{Calders & Žliobaitė (n 44) 47} 106_ibid 107_{ibid. 48} 108_{ibid. 50} 109_ibid. 110_ibid. 111_ibid. 112_ibid. 113_{ibid. 50-51}

(18)

2.2.2 Sampling Bias

Whereas the sampling can be biased even if the labels are correct – but the groups in question are over or under represented114_{. This can occur when some factors are disregarded} in an analysis115. A real life example that shows the problems with discrimination is the pothole locator app in Boston. The app was meant to locate potholes based on the car speed. However, the app was only available on the Apple app-store, thus, only the neighbourhoods that could afford apple products were the ones to have representative data116. This type of discrimination would go under sampling biases, due to the fact that it omits a certain characteristic/group from the sample distribution117. A parallel can be drawn with the

alternative credit scoring sector and the discrimination that some people might experience do them being a part of a group that is disadvantaged from the start – whether they choose to be a part of that group (not liking Apple) or whether they are a part due to circumstances outside of their control (the person in question bought a cheap phone at the point in time where they could not afford more, although currently – they can).

2.2.3 Incomplete Data

Lastly, the data can be incomplete, which would mean that important characteristics are missing from user profiles118. This can happen due to a myriad of reasons, like privacy reasons or because data is hard to collect119_{. In these circumstances, the classifier would use} the remaining attributes and get the best accuracy out of it – which overestimates the factors that are already present in the dataset120_{. This can be linked back to the final stage of}

alternative credit scoring and how ML estimates the important variables and gives them weights121. Hypothetically, if a company was only able to retrieve the address of the subject which is in a neighbourhood with a high default rate, this would bias the report in a way that might yield a low credit score. While the credit scorer did not manage to obtain the data that shows the individual in question owning multiple houses in better areas. At this point it is

114_ibid. 115_{ibid 51-52}

116_{Kate Crawford, 'Think Again: Big Data', Foreign Policy, (2013)}

<https://foreignpolicy.com/2013/05/10/think-again-big-data/> accessed 14 July 2020. 117_{Calders & Žliobaitė (n 44) 51}

118_{ibid. 52} 119_{ibid. 52-53} 120_ibid.

(19)

important to highlight that alternative credit scoring companies try and mitigate this problem by using ‘all’ data to fill in the blanks.

Based on the notions put forth by alternative credit scoring companies, alternative data collection is supposed to minimize biases and increase default rate accuracy122. However, researchers pointed towards algorithms123 reinstating redlining and similar practices by churning biased or incomplete data to produce seemingly objective results124. Additionally, even FICO has recognized that over reliance on ML can "can actually obscure risks and short change consumers by picking up harmful biases and behaving

counterintuitively"125. This in turn could allow for discrimination. The algorithms will not necessary be the ones discriminating, but if the data is not gathered correctly, if it is not labelled properly and if it does not have enough characteristics to describe a user, it might produce discriminatory results126. Thus, the risks are both present and known; however, whether legislation lives up to the promise of safeguarding individual rights and liberties is to be analyzed in the following section.

122_{Aitken (n 54) 285}

123_{In the form of Artificial Intelligence} 124_{Pearson (n 11)}

125_{Ethan Dornhelm, “Machine Learning's Promises, Pitfalls” (February 13, 2020) American Banker} <https://www.americanbanker.com/opinion/machine-learnings-promises-pitfalls>; accessed July 20, 2020 126_{Calders & Žliobaitė (n 44) 47}

(20)

3. Regulatory Framework of Credit Scoring Agencies

3.1 Current Legislation in the United States

Since there exist federal laws that these new credit scoring companies could (and have been) be clustered under, it becomes necessary to examine them127. Furthermore, due to the scope and goals of this thesis, instead of contrasting the strengths and weaknesses of the legislation, attention will be paid mainly to the shortcomings. Thus, the main federal laws that should cover alternative credit scoring with regards to discrimination are: The Fair Credit Reporting Act (FCRA) and the Equal Credit Opportunity Act (ECOA). This seems to show that there are already backstops put into motion, but reality may paint a different picture – as it will be seen in the forthcoming sections. As a reminder, this master thesis seeks to look into the issues that arise concerning transparency and regulatory arbitrage. And furthermore, if the hypothesized discriminatory behaviour (whether accidental or premeditated) could be

circumvented.

3.2 Fair Credit Reporting Act (FCRA)

A whitepaper published by a software company Varonis gives an overview of the FCRA128. The reason behind the enactment of FCRA was to regulate Credit Scoring Agencies, their management of consumer credit profiles, their accuracy, accessibility to consumers, restrictions on who can see the data, and provide protections against identify theft129. At the core, the FCRA seeks to protect personally identifiable information130 (PII) that relates to credit information and other personal credit data131.

Initially, it is important to examine who falls under the FCRA. It depends on the type of information involved, the actual or expected use of the information, and whether the information is handled by a consumer reporting agency (CRA)132, in the process of creating a “consumer report”133. The type of information that can be used is broad, however, it will not

128_{Varonis, ‘WHITEPAPER US Data Protection Compliance and Regulations’ VARONIS (2020)} 129_{ibid. 7}

See generally: 15 U.S.C. § 1681(b) (2012).

130_{“…name, address, social security number, or phone number”.}

ibid. 4

131_{15 U.S.C. § 1681a(d)(1) (2012)} 132_{15 U.S.C. § 1681a(f) (2012).}

Used interchangeably with credit scoring agencies and credit reporting agnecies

133_{“Consumer Report”: “any information ... bearing on a consumer's credit worthiness, credit standing, credit} capacity, character, general reputation, personal characteristics, or mode of living”. See: 15 U.S.C. §

1681a(d)(1) (2012).

(21)

be considered to be a consumer report, unless the information concerns an identifiable person134_.

First, a possibility of evading regulation comes from the fact that consumer reports can only be made by CRA’s to third parties135. The definition of CRA’s states: "[a]ny person which, for monetary fees, dues, or on a cooperative nonprofit basis ...[r]egularly engages in whole or in part in the practice of assembling or evaluating consumer credit information or other information on consumers" for the purpose of "furnishing consumer reports to third parties”136. Meaning, if the data is not used by a third party, then the agency gathering the data and giving out the scores can evade regulation. This is further problematized by the fact that some data brokers disclaim responsibility on how the data is used, effectively bypassing regulation by stating that they do not collect data to create consumer reports, thus they should not be regulated as a CRA137.

Furthermore, CRAs must use reasonable procedures to guarantee accuracy of information on the report138. Meaning, besides being true, the information must not be misleading or incomplete139. In the case of an adverse decision on a consumer’s application, the lender has to provide instructions on how to obtain the information in the report140_{. While} the customer should have the right to obtain the report and challenge the accuracy of the info141. This becomes an issue with the new companies, because there is less transparency; not in the sense of the requested report being secretive, but the data points collected being so vast that the burden of accuracy is being shifted on the individual142. Meaning, the companies provide their data models, while the customers are the ones expected to make sense of them. Additionally, FCRA does not limit the type of information used except outdated information like criminal and financial records143_{. Meaning that consumers have few references on what} information is collected on them; which, is made worse by the “all data is credit data”

134_{McCready v. EBay, Inc., 453 F.3d 882 (7th Cir. 2006) (information pertaining "consumer report").} 135_{15 U.S.C. § 1681a(f) (2012).}

136_{15 U.S.C. § 1681a(f) (2012).}

137_{Persis Yu, Jillian McLaughlin, and Marina Levy. (2014) "Big Data: A Big Disappointment for Scoring} Consumer Credit Risk." National Consumer Law Center March.

138_{Hurley & Adebayo (n 8) 188} 139_{15 U.S.C. § 1681e(b) (2012).}

See: See Fair Credit Reporting, (n 100) § 4.2.3.

140_{15 U.S.C. § 1681m(a) (2012);}

see also Fair Credit Reporting, (n 100) § 3.3.6. 141_{15 U.S.C. § 1681g (2012).}

Also: 15 U.S.C. § 1681i(a)(2012).

142_{Hurley & Adebayo (n 8) 189} 143_{15U.S.C.§ 1681c}

(22)

approach from Zestfinance . In other words, a customer that is not educated on tackling a data modelling errors – which is unlikely because there are no limits on the data that can be used – might be rendered helpless in the face of a denied loan. Additionally, not only is it hard to study these data models, it might even be impossible because there is no way of preparing for these errors beforehand. This is due to the data models being trade secrets, and there being practically zero notice given as to what type and how the data will be utilized145. Thirdly, courts have pointed out that reports concerning property, and not the owner themselves, can fall outside of FCRA146. Meaning, by clustering the owner under the property, the reports can sidestep regulation. Furthermore, these reports can also bypass regulation by deleting the personal information (e.g. names, addresses) of customers and anonymizing the individual ID’s147. This poses as a problem because it is still possible to identify the individual, even after these actions148. And although the Federal Trade

Commission (FTC)149 has noted that even if the information is not tied to the customer, it can still qualify as a consumer report, the question still remains on if there is a vivid line dividing identifiable and non-identifiable information150. Based on this loophole, a company can make and sell a customer report, while dodging regulation because the reports are not directly liked to individuals.

Thus, based on the intricate data modelling and the ambiguous character of these companies, there seem to be a multitude of problems unravelling. First, the issue with the definition of CRAs; the alternative credit scoring companies can evade regulation by either separating themselves from the responsibility of how the data is used, or by not acting as third parties and making the loans themselves151_{. Second, even if it is established that a} company is a CRA and the subsequent CRA made an error in a consumer report, once big data steps into the picture, it becomes harder for a lay person to do justice to their own report,

144_{Quentin Hardy, 'Just The Facts. Yes, All Of Them.', Nytimes.com (2020),}

<https://www.nytimes.com/2012/03/25/business/factuals-gil-elbaz-wants-to-gather-the-data-universe.html> accessed 14 July 2020.

145_{Yu, McLaughlin and Levy (n 137) 20}

146_{Fuges v. Southwest Fin. Serv., Ltd., 707 F.3d 241, 253 (3d Cir. 2012)} 147_{Hurley & Adebayo (n 8) 185}

148_{Julia Angwin & Mike Tigas, Zombie Cookie: The Tracking Cookie That You Can't Kill, PROPUBLICA} (2015), http://www.propublica. org/article/zombie-cookie-the-tracking-cookie-that-you- cant-kill

[https://perma.cc/7H5W-JKY4].

149_{The regulator overlooking credit scoring agencies.}

150_{FEDERAL TRADE COMMISSION, ‘40 Years of Experience with the Fair Credit Reporting Act: An FTC} Staff Report with Summary of Interpretations, reprinted in National Consumer Law Center, Fair Credit Reporting, at Appx. D, Part V’ (8th ed. 2013), www.nclc.org/library [https://perma.cc/8RKG- JHGW]. 151_{Hurley & Adebayo (n 8) 183-199}

(23)

due to the vastness and intricacy of the reports152. Furthermore, it seems to be hard, if not impossible, to prepare oneself and become knowledgeable of these reports, due to the very same problems, coupled with the fact that the data gathered can be dynamic and

ever-changing153. And lastly, there is another way to sidestep regulation, which is by anonymizing personal data. And the issue here stems from the fact that researchers have pointed towards a possible abuse of loopholes, due to the data still being traceable to the individual154. Thus, whether it is regulatory arbitrage or the burden that the individual bares, it seems that the FCRA does not present itself ready to fully govern these new credit scoring companies.

3.3 Equal Credit Opportunity Act (ECOA)

ECOA was enacted in 1974 to prohibit discrimination from creditors against the applicants on the basis of sensitive characteristic like religion, race, national origin, sex, or marital status155. The core purpose of this law is to act as a means for individuals and classes of consumer to challenge lending decisions and policies that either overtly discriminate, or lead to discrimination156. In order to file a claim, the challenger can either allege “disparate treatment” by highlighting that based on the aforementioned sensitive characteristics, they were singled out and subsequently treated unfairly157. Or, they can allege “disparate impact”, meaning, a policy that seems neutral actually had less favourable terms for members of a protected class in contrasted to other borrowers158. On paper, this act seems to be able to challenge discrimination with regards to CRAs, but again, these companies have gone through a big change with the help of big-data. In reality, customers may find it harder to challenge biases based on disparate treatment or disparate impact if the lender justifies their actions based sophisticated credit scoring algorithms and a multitude of data points159.

The first potential shortcoming brought forth is the limited scope of protection. Namely, ECOA only prohibits discrimination on the above mentioned grounds, while, for example, sexual orientation is not mentioned160_{. Some courts have expanded the prohibition} 152_{ibid, 188-189} 153_{ibid, 183-199} 154_{ibid, 186} 155_{15 U.S.C. § 1691(a)(1) (2012);} See also, 12 C.F.R. § 1002.1 (2016);

Treadway v. Gateway Chevrolet Oldsmobile Inc., 362 F.3d 971, 975 (7th Cir. 2004). 156_{Hurley & Adebayo (n 8) 189-190}

157_{Ricci v. DeStefano, 557 U.S. 557, 557 (2009) (construing the disparate treatment test in the context of an} employment discrimination suit under Title VII of the Civil Rights Act of 1964).

158_ibid.

159_{Solon Barocas & Andrew D. Selbst, Big Data's Disparate Impact, 104 CALIF. L. REV. 22} 160_{Rosa v. Park W. Bank & Trust Co., 214 F.3d 213, 215 (1d Cir. 2000)}

(24)

on sex discrimination as encompassing gender norms, but nevertheless, consumers

themselves may run into difficulties trying to challenge this type of discrimination161_{. The} issue of limited scope is only exacerbated by the fact that proving discrimination under ECOA is burdensome, because the plaintiff has to prove either that the lender made their decision on “a discriminatory intent or motive”162, or by showing that the decision had a “disproportionately adverse effect on minorities”163. All of this, coupled with the thousands of data points and intricate date modelling of the alternative credit scoring companies, might actually mask overt discrimination164. Or, alternatively, as provided in previous sections, it could perpetuate the already existing biases in the data.

The second potential shortcoming comes from the fact that even if a plaintiff can successfully challenge a decision based on disparate impact, the lenders can avoid liability if they prove “business necessity” by stating and explaining a “valid interest served”165. The lender must prove that the policy was “related” to its objectives or business goals166. Thus, with a loose reading of “related” in addition to the fact that courts have disagreed on what constitutes as a business justification167, the lenders could prove business necessity and shift the burden back to the plaintiff to provide a policy or practice that would be as effective in meeting the business goals and not produce a disparate impact168.

Thus, based on these two scenarios, lenders can implement policies that could single out specific groups by using neutral proxies on purpose or accidentally, and keep the biased data as is. Under both of these instances the ECOA is helpless. Furthermore, if a lender seeks to challenge a decision, they might not even be able to because of a similar problem as in the

161_ibid.

162_{Ricci v. DeStefano, 557 U.S. 557, 557 (2009)} Disparate treatment.

163_{Ricci v. DeStefano, 557 U.S. 557, 557 (2009)} Disparate impact.

164_{Barocas & Selbst (n 159) 22}

165_{Texas Dep't of Housing & Cmty. Affairs v. Inclusive Communities Project, Inc., 135 S.Ct. 2521-2525} (2015).

Also: Ricci v. DeStefano, 557 U.S. 557, 587 (2009).

166_{Ricci, 557 U.S. at 578 (“If an employment practice which operates to exclude minorities cannot be shown to} be related to job performance, the practice is prohibited.")

167_{Ricci v. DeStefano, 557 U.S. at 578 (2009) ("the 'touchstone' for disparate-impact liability is the lack of} 'business necessity': If an employment practice which operates to exclude minorities cannot be shown to be related to job performance, the practice is prohibited."

Also: Timothy C. Lambert, ‘Fair Marketing: Challenging Pre-Application Lending Practices.’ (1998) Geo. LJ

(25)

FCRA, namely, the models have become so intricate and vast that they might mask the wrongs of the decisions. And lastly, if a case were to arise that a customer successfully challenges a discriminatory action, the lenders can still dodge responsibility by claiming business necessity.

3.4 Overall Regulatory Climate and Potential Solutions

Overall, there seem to be difficulties that regulators might face when applying federal laws such as the FCRA or ECOA to alternative scoring. This is especially concerning

because there are multiple issues at hand, namely: 1) insufficient transparency, 2) the potential for regulatory arbitrage, and 3) discrimination in the form of biased data. It is noteworthy to highlight that these challenges depend on each other and potentially have an interconnected relationship. For example, insufficient transparency is an issue that could lead to regulatory arbitrage. This is so, because obfuscation hinders meaningful research by the consumers or the regulators, which leaves the door open for activities that may be value maximizing, but socially harmful169. The hidden nature of the activities might contribute to arbitrage, through which, the companies could discriminate against individuals170. Even if the law allows it (through a loophole or otherwise), this does not mean that unfavourable

treatment of individuals should be the accepted norm. This goes back to the first point, in that, it might be hard to identify gaps if there is little transparency. Furthermore, as provided above, unfavourable treatment has a multiple different ways of materializing. Thus, due to this interconnectedness, a legal solution needs to encompass all of the challenges that alternative credit scoring poses, rather than addressing only some.

The following subsections will elaborate on four possible solutions to address the aforementioned problems. First, more legislation, possibly aiming at alternative credit scoring companies. Second, Self-regulation, which is essentially an approach between a regulator and the company, under which the company is left to comply or report on some laws on their own terms171_{. Third, banning these type of companies/activities altogether. And fourth, regulatory} sandboxes, which are essentially testing grounds for innovative products where companies and regulators stand as counterparties172. Even though all of these approaches could possibly

169_{Hurley & Adebayo (n 8) 195-196} 170_{ibid, 196}

171_{Neil Gunningham & Joseph Rees, (1997). Industry Self-Regulation: An Institutional Perspective. LAW &} POLICY, 19(4), 366, 370

172_{Wolf-Georg Ringe & Christopher Ruof, ‘A Regulatory Sandbox for Robo Advice’, (2018) ILE Working} Paper Series, No. 14, University of Hamburg, Institute of Law and Economics (ILE), Hamburg. 4

(26)

do justice in addressing the issues of alternative credit scoring, it is necessary to highlight why this master thesis opted to pursue the last one.

3.4.1 More Legislation

With more legislation, there might pop up the following issues: first, the nature of these companies makes them difficult to legislate against173. This is because these companies do not have restrictions on the data they use174. Meaning, if there were to be restrictions on some type of data today, tomorrow the companies might find different data, which, could lead to the same discriminatory results. In other words, the dynamic and open ended nature of these companies makes it hard to legislate against them. Second, arriving at the point of understanding of how these companies function might prove difficult, mainly because there is the issue of secretive data modelling175. In other words, because the activities of the

companies are cutting edge and competitive, they are incentivized to keep their activities to themselves and not divulge the info176. This lack of transparency, in turn, makes it harder for customers to prove unfair treatment and for regulators to identify the gaps and apply the laws177.

3.4.2 Self-Regulation

A possible second solution considered by this master thesis is based on industry self-regulation. In one of their reports, the Organisation for Economic Co-operation and

Development (OECD) states that industry self-regulation “[…] concerns groups of firms in a particular industry or entire industry sectors that agree to act in prescribed ways, according to a set of rules or principles. Participation by firms in the groups is often voluntary, but could also be legally required. The groups can be wholly responsible for developing the self-regulatory instruments, monitoring compliance and ensuring enforcement, or they can work with government entities and other stakeholders in these areas, in a co-regulatory

capacity”178. In this case, even if the most tightly regulated version of this approach is

173_{Hurley & Adebayo (n 8) 189} 174_{15U.S.C.§ 1681c (2012).}

Also: Hurley & Adebayo (n 8) 189

176_{Matthew A. Bruckner, Preventing Predation & Encouraging Innovation in Fintech Lending. (2019).} Available at SSRN 3406045. 6

178_{OECD, ‘Role and Use in Supporting Consumer Interests’ (2015) Self-Regulation, Industry.} DSTI/CP(2014)4/FINAL 11

(27)

undertaken, meaning self-regulation being legally required and monitored by government entities, it might still pose some issues with regards to alternative credit scoring companies. Namely, even though there is an argument to be made about the companies themselves being better able to understand their own products, there nevertheless have been issues that

materialized179. An example was the 737 Max scandal, in which the Federal Aviation Administration (FAA) delegated Boeing to partially oversee the certification of its 737 Max plane software, which, at this point is believed to be the key reason on what led to the plane-crashes180. Some government officials have said that the problems were caused due to the FAA doing “safety on the cheap” with allowing self-regulation, while others pointed towards the close relationship with regulators and the industry (possibly hinting at regulatory

capture181), and calling for the relationship to be questioned and changes made182. And

furthermore, critics in academia cite the same reasons for distrust of self-regulation, in that, it fails to fulfil its theoretical promises and serves the industry rather than public interest183. As John Braithwaite has put it, “Self- regulation is frequently an attempt to deceive the public into believing in the responsibility of a[n] irresponsible industry. Sometimes it is a strategy to give the government an excuse for not doing its job''184_{.Thus, it seems like the transparency} problems might still be prevalent even after this approach. And with alternative credit scoring being this fresh and with so much influence over individual’s lives, it becomes necessary to control it at its infancy, before it becomes large enough to have leverage over the regulators and wide consequences regarding individual liberties.

3.4.3 Shutting Down the Industry

The third possible solution could be shutting the industry down, but as mentioned previously in section 1.3 “Benefits of the new method”, the stats paint a picture that argues for the benefits that alternative credit scoring bring185. Thus, shutting down the industry will

179_{Gunningham & Joseph (n 171) 366, 370.}

180_{Sinead Baker, ‘FAA boss says it let Boeing partly self-regulate the software thought to be behind both fatal} 737 Max crashes’. (2019)

181_{Regulatory capture is the process through which regulated entities end up manipulating the state agencies} that are supposed to control them.

See: Ernesto Dal Bó, "Regulatory capture: A review." (2006) Oxford Review of Economic Policy 22, no. 2:

203-225. 203 182_{Baker (172)}

183_{Gunningham & Joseph (n 171) 370.}

184_{Braithwaite, John. "Responsive regulation in Australia." (1993), Business regulation and Australia’s future,} 91

(28)

not only discourage innovation in this sphere, but will also arguably decrease social welfare186_.

This brings us to the focus of the master thesis – regulatory sandboxes. And as the text will show later on, this approach has already been considered by the United States’ government and has already been implemented187. Thus, the following subsection will give an overview, highlight how it could regulate alternative credit scoring companies efficiently, and finally, provide the potential shortcomings of this approach.

186_{Henry & Morris (n 20) 1}

187_{CFPB, “Granted Applications” (Consumer Financial Protection Bureau)}

<https://www.consumerfinance.gov/policy-compliance/innovation/granted-applications/>; accessed July 22, 2020

(29)

4. Regulatory Sandboxes

4.1 Why a Sandbox?

Emergence of disruptive technologies oftentimes put established regulatory

paradigms and business practices under a question mark188. Due to this, regulators might face issues when applying laws that were essentially meant for a different industry at a different time (i.e. the differences between original and alternative credit scoring companies)189_{. Thus,} regulation should pay attention to risks that new phenomena pose, because it is the

materialization of these risks that trigger intervention190. However, good regulation does not only focus on addressing risks, but also identifies and promotes desirable market

developments191. While too much regulation potentially stifles innovation, competition and consequently economic development192.

The risks in question, are connected to the vague and non-transparent nature of these companies; the regulatory arbitrage opportunities that they might exercise; and the influence that they exert on individual lives, potentially leading to discrimination. But nevertheless, these firms have also shown to be useful and regulating them with only the issues in mind might hinder their growth, in turn hurting social welfare193. Thus, the companies that understand this are placed in a “grey zone”194 that makes the reactions of the regulation authority unpredictable195. Effectively, placing the companies under regulatory uncertainty196.

Regulatory uncertainty has proven to show problems such as: barriers for market entrants – due to the fear of sanctions upon entry; barriers for investors – who factor in ‘potential risks’ of the companies in the valuation, even if these risks might never materialize and the investors themselves are not well placed to assess them197; and finally, another form of regulatory arbitrage in a sense of “forum shopping”198 – which, in our case, if the

regulatory solution is not federal, the companies could establish themselves in states that

188_{Ringe & Ruof (n 172) 33} 189_{ibid. 4} 190_{ibid. 13} 191_{ibid. 7} 192_{ibid. 21} 193_{ibid. 24} 194_ibid. 195_ibid. 196_ibid.

197_{FCA, ‘Regulatory Sandbox’ (2015) Financial Conduct Authority. 5}

198_{Forum shopping is a practice adopted by litigants to get their cases heard in a particular court that is likely to} provide a favorable judgment.

See: US Legal Inc., “Forum Shopping Law and Legal Definition”

Big Data in Credit Scoring, Are We Ready?