Using machine learning to improve lending in the emerging markets

Using machine learning to help small businesses and individuals source loans

At the heart of most Western banking systems is an integrated and trusted credit scoring system accessible to all market participants. Because of this system, participants can confidently make lending decisions (both whether to issue a loan and at what price) based on risk models that have been refined over time. In many emerging market nations such credit scoring infrastructure does not exist. As a result, banks in such markets are often unwilling to lend to most consumers and businesses. This trend has led to a funding gap of $2.1-$2.6 trillion dollars for SMEs (Small and Medium-sized Enterprises)[1]. Several firms have stepped up to fill that void with dynamic credit scoring solutions built on machine learning. A notable player is Mines.io.

Why Apply Machine Learning to Credit Scoring?

Machine-learning augmented credit risk modeling allows for a level of nuance not easily attained in traditional models. In traditional credit risk modeling, customers are tagged with easily observed identifiers (new customer, old customer, high earner, etc.) and the credit behavior of these groups is analyzed to discern key trends; with these trends being incorporated into a composite “credit score”.[2] The issue with this approach is that these broader categories don’t necessary offer the level of granularity needed to make the most optimal lending decisions. Consequently, under this approach companies are likely to “leave money on the table” by opting not to lend to entire swathes of the economy or not adequately pricing-in the risk of a specific individual or entity whose activities would distinguish it from its broader peer set. Using machine learning, companies like Mines.io can execute micro-segmentation based on customer behaviors rather that non-behavior identifiers.[3] Additionally, machine learning techniques are also able to “train” models based on additional data sources to improve the predictive power of credit models.[4]

Does the Technology Have any limitations in an Emerging Markets Context?

Regulatory limitations as well as potential data issues limit the ability of lenders to solely rely on machine learning derived-decisions. From a regulatory perspective, lenders are often required to explain very clearly the methodology/rationale for their lending decisions. As such, lenders cannot have a “black box” AI process. Separately, because of potential data quality issues companies must be careful that the source data feeding into their models cannot be easily manipulated. Lastly, in the emerging market context in which companies like Mines.io operate, Companies must be vigilant for backward-looking bias in the data used to generate insights in their models. In regions with rising middle classes many of whom are making money in novel ways (across many small business ventures, etc.), it can be hard to ascertain what aspects of previously successful borrowers led to success and whether those aspects will be salient for new applicants.

What is Mines.io and what is their go-to-market strategy?

Founded in 2014, Mines.io was started by silicon-valley based data scientists keen to build the infrastructure needed to allow financial institutions to more confidently lend to SMEs and individuals, track the credit history of those entities across their entire credit life, and integrate that credit history into an iterative credit algorithm to improve go-forward lending decisions.[5] Having built its system, the key question was how to go to market. From a regulatory perspective, direct lending would likely require banking licenses and compliance/regulatory complexities as well as a more complicated sales process that would take the company farther away from its core data science competencies.

In the short-term, to counter compliance and regulatory complexities associated with directly offering lending products, the Company has instead opted to follow a SaaS model. As part of this strategy, the company has launched a four-fronted product suite selling lending software-as-a-service products to incumbent banks, mobile operators, retailers, and payment processors. To each customer profile, the company offers a white-label service that allows the client to offer lending products to its customers that make lending and rate decisions based on an algorithmic credit engine.[6] To date, the Company has focused on small unsecured consumer loans. In order to monetize, the company charges a specified percentage of interest income from loans originated / assessed by the Mines.io platform.

In the long-term, there are legitimate questions about how much of their underlying capital lenders will be willing to risk on lending decisions made by a 3rd party that does not bare any cost if the loan fails. Second, the Company will also have to answer the question of whether data collected for the purpose of assessing small consumer loan creditworthiness will be applicable for larger loans and other financial products (mortgages, insurance, car loans, etc.) as it looks for additional sources of revenue.

(769 words)

 

[1] Owens, J. and Wilhelm, L. (2018). Alternative Data Transforming SME Finance. [online] Gpfi.org. Available at: https://www.gpfi.org/sites/default/files/documents/GPFI%20Report%20Alternative%20Data%20Transforming%20SME%20Finance.pdf [Accessed 13 Nov. 2018].

[2] Zoldi, S. (2018). How to Build Credit Risk Models Using AI and Machine Learning. [online] FICO. Available at: https://www.fico.com/blogs/analytics-optimization/how-to-build-credit-risk-models-using-ai-and-machine-learning/ [Accessed 13 Nov. 2018].

[3] Zoldi, S. (2018). How to Build Credit Risk Models Using AI and Machine Learning. [online] FICO. Available at: https://www.fico.com/blogs/analytics-optimization/how-to-build-credit-risk-models-using-ai-and-machine-learning/ [Accessed 13 Nov. 2018].

[4] Zoldi, S. (2018). How to Build Credit Risk Models Using AI and Machine Learning. [online] FICO. Available at: https://www.fico.com/blogs/analytics-optimization/how-to-build-credit-risk-models-using-ai-and-machine-learning/ [Accessed 13 Nov. 2018].

[5] MINES – Digital Credit For Emerging Markets. (2018). MINES – Digital Credit For Emerging Markets. [online] Available at: https://www.mines.io/ [Accessed 13 Nov. 2018].

[6] MINES – Digital Credit For Emerging Markets. (2018). MINES – Digital Credit For Emerging Markets. [online] Available at: https://www.mines.io/ [Accessed 13 Nov. 2018].

Previous:

CircleUp – Identifying the next “Big Thing” using Machine Learning

Next:

CLOUDFLARE: USING MACHINE LEARNING TO SECURE AGAINST CYBER ATTACKS

Student comments on Using machine learning to improve lending in the emerging markets

  1. This seems like a great product especially because its solving a salient need for SME’s in emerging markets. I’d be curious to see if this same type of predictive modeling could be leveraged by investors who want to assess the “riskiness” of a new venture. Although I’m sure this is apart of their secret sauce, I’m curious to know what information is fed into the model that is starkly different from the information used by creditors. As a consumer, I’d be careful to only offer access to things that tell the best credit story about me so I wonder how they’re planning to solve for that in their algorithm, assuming they can’t access information without someone’s permission.

  2. A neat application of machine learning technology to solve a major pain point in emerging markets. I agree with your analysis that the lack of granularity emerging market banks employ in developing individual credit scores leaves a major portion of the population unbanked or under-banked and also limits the banks’ growth potential, and that machine learning can help address this problem. In response to your question around whether banks will accept the risk of a third party making lending decisions, I do not think this will be an issue in the long-term. It is at the discretion of the bank to make a lending decision based on a variety of factors, and the third party credit score is only one of them. Furthermore, there are many companies attempting to tackle the same problem, such as Lenddo and ZestFinance, and these companies will inevitably be purchased by large banks who understand and can integrate the technology into their lending operations.

    In terms of limitations in the emerging markets context, one of the main hurdles is lack of access to customer behaviour data. Most emerging markets have huge informal economies, such that information about a customer’s business doesn’t ever enter the digital realm, where companies like Mines.io obtain their information. In addition, micro-segmentation presents the risk of discrimination in lending decisions, and it’s possible that major swaths of borrowers will continue to be left out if the algorithms do not adjust for biases that are present in the training data.

    [1] Tech Emergence Website, https://www.techemergence.com/artificial-intelligence-applications-lending-loan-management/, accessed November 2018.

  3. I enjoyed reading the piece, definitely a very interesting angle.

    I agree with the limitation the author has raised, especially around regulatory limitation as well as data access. Regulatory limitations and data issues will definitely limit the ability of lenders to solely rely on machine learning derived-decisions. I also wonder how much manipulation borrowers in emerging markets can make to secure lending if it is solely vetted by an algorithm.

    In certain markets, I do agree ML can help with those lending decisions. I believe Ant Financials (Alibaba subsidiary) applied ML and built a platform as well as their own credit rating at least on the consumer front, they then have the service of “Huabei” to provide financial solutions on consumer purchases. The scores “Zhima Xinyong” are now widely accepted at various other services since the majority of the people in China has it and it provides some benchmark. So it is another interesting example to check out if you are interested.

    Good article overall and I enjoyed reading your perspective!

  4. Interesting article. As the author of the article mentions, using data improves the quality of the credit assessments by being more granular, though there are a lot of challenges related to regulation. On top of that, selecting which data to be used as a proxy for credit scores can be another challenge. I have worked closely with a startup that used data mining to improve the quality of credit scores in emerging markets in Asia. They used the data from uber and grab drivers as a way of improving their credit scores. So if a driver worked several times a week for several hours, it meant he was hardworking and therefore more capable of paying back debts. However, there are several limitations with this approach (e.g. if a driver has another job) that need to be accounted for in order for this to be a reasonable proxy.

  5. I enjoyed reading your article. Having worked in a number of frontier markets, I have witness how limitations in credit scoring infrastructure has starved SMEs from much needed capital. I agree that machine learning can be leveraged to compute credit scores that can form the basis for lending to this segment. Among many reasons, the barriers to loan products for SMEs stems from how expensive it is to serve this market. Given the limited amount of consolidated data on companies in this segment, data gathering and analyses is very time consuming and expensive. Also, there is no consensus on the basis upon which credit score should be calculated and school of thoughts on credit scores are very dynamic.
    Specific to Mines.io, I agree that regulations that require a banking license can make this product expensive to execute. However, I am weary about the product diffusion in the market as my predication is that adaption would be low. A key concern for banks, its potential customer, is the credibility of this model. Machine learning becomes intelligent overtime and one of the ways it learns is from how past credit scores have been useful in predicting credible individuals to given loans. Who bears this learning cost and who provides the data for the machine to learn. I think this two points would pay a key role in the decision process of banks. Expecting a bank to adopt an unproven model is a herculean task and expecting a bank to pay for a value that they are instrumental in creating may be an uphill battle.

    Thank you for a great read!

  6. Great read! This is one of my favorite applications of ML, especially because of the impact it can have on the developing world. I’m very curious to know what exactly are the “behaviors” that they use to assess credit quality in a non-traditional way, and how they go about measuring them. Depending on how easy it is to track these behaviors, a way to prove the effectiveness of the product to prospective customers would be to back-test the AI-driven lending decisions against a financial lender’s existing portfolio and see if it can outperform it.

    If you want to explore another company, check out Tala (https://tala.co/). They are a startup based in LA and have issued around $500mm in loans to customers in underserved countries by using alternative data signals.

Leave a comment