Machine Learning in Credit Assessment at Capital One

Capital One is banking on machine learning for new product innovations, but the bank should leverage their machine learning resources to develop machine learning for credit risk assessments, using predictive technology to identify good customers using broader sets of data. Through this, machine learning can help expand financial inclusion, to better bank the un- and under-banked.

Since its inception, Capital One has been recognized as a leading-edge financial institution when it comes to its use of technology in banking1. By leveraging big data, Capital One has attracted 45 million customers, making it a top 10 U.S. bank2. But there’s certainly room to grow and new technologies to fuel this growth. One potential innovation for growth is machine learning for customer risk assessment, not only to make better lending decisions but also to expand financial inclusion to un- and under-banked Americans.

Machine learning is on Capital One’s radar. In 2017, the bank hired Nitzan Mekel to build and lead their Machine Learning organization3, which is growing with speed—in the last month alone, nearly 30 “Machine Learning” jobs were posted on the bank’s Glassdoor page4. Beyond being an area of human capital investment, machine learning is receiving the bank’s research dollars, as well3; the research team’s summarized findings are summarized in a dedicated Machine Learning section on the company’s tech blog5 and Domenic Puzio and Jennifer Van from Capital One spoke about “How to Become a Machine Learning Expert in Under an Hour” at the South by Southwest conference6. Beyond research, the bank is incorporating machine learning into product development, citing Eno, one of the industry’s first AI-powered customer service chatbots, and real-time fraud detection as customer-available applications of machine learning at Capital One3. Machine learning seems to be integrating into “almost every facet of [the] business,” says Mekel3.

While exciting progress has been made, there’s more the bank can do to expand machine learning into product development, specifically customer risk assessment. Most banks use few pieces of data, namely credit score, to make lending decisions7, but this may not provide a holistic picture of the customer’s credit-worthiness. By incorporating wider data sets into machine learning models, Capital One could better predict a customer’s fit for financial products. For example, during the Kabbage online application process, customers link their PayPal and Quickbooks accounts through APIs; Kabbage also considers social media data (e.g., Twitter followers, Facebook likes, customer reviews) when making small business loan decisions8.

New and smarter data models would help banks identify potential customers from those automatically rejected because of a low or nonexistent credit score. This is particularly helpful for the 63 million un- or under-banked Americans, who sought financial services products outside of the banking system9. If banks, like Capital One, were to expand the data they use to make lending decision, and make smarter decisions powered by machine learning models, credit-worthy un-banked customers would gain access to capital and banks would gain new customers. It’s a win-win proposition.

Capital One has recognized the ability for machine learning to enhance its credit risk assessments, but development efforts are not ready for product-launch yet. Mekel points to “explainability” as the reason for the delay: “Being in a heavily regulated environment, we want to make sure that we’re not just meeting the regulatory requirements, but that we help set the standard for what fair and ethical machine learning deployment looks like,” Mekel says3. As the models and algorithms become more advanced, it becomes harder to explain to customers and regulators alike, how the underlying models work and with complexity also comes an increased risk of harm from biased, unethical, or unfair outcomes3. Mekel and his team view these as challenges to be addressed before further machine learning applications are deployed at Capital One.

As Capital One more deeply integrates machine learning into consumer product development, the team should consider many things, including data selection and training to eliminate biases; consumer sentiment towards granting broader access to personal data; and partnership with the regulators to make a commercially viable, safe models for lending decisions. To help tackle these challenges, Capital One should consider hiring data ethicists as it seeks to be the standard setter for fair and ethical machine learning. To address the “explainability” issue, Capital One should also seek to proactively educate consumers and regulators alike on how broader data sources are being used to make better business decisions.

There are key questions to be considered as Capital One continues to integrate machine learning into its business DNA. What are potential data sources Capital One should consider when assessing someone’s credit worthiness, and what are the potential sources of biases? What are other ways that Capital One and other financial institutions can use machine learning to close the inclusion gap? I applaud Capital One for its machine learning efforts and its vision to use new innovations to benefit both consumers and the bank; I hope to see machine learning applied to address the issue of financial inclusion in the coming years. (775 words)



1 Capital One Financial Corporation, “Awards and Recognition”,, accessed November 2018.

2 Capital One Financial Corporation, “About Capital One”,, accessed November 2018.

3 Allison Toh, “AI in Your Wallet: Capital One Banks on Machine Learning,” AI Podcast (blog), October 12, 2018,, accessed November 2018.

4 Glassdoor, “Machine Learning Engineer Jobs”,,11_KO12,37.htm, accessed November 2018.

5 Capital One Tech, “Machine Learning”,, accessed November 2018.

6 Capital One Tech, “Become a Machine Learning Expert in Under an Hour”, Capital One Tech (blog), February 28, 2018,, accessed November 2018.

7 Experian, “How lenders make—and monitor—credit decisions”,, accessed November 2018.

8 Darren Dahl, “The Six-Minute Loan: How Kabbage is Upending Small Business Lending and Building a Very Big Business”, Forbes, May 6, 2015,, accessed November 2018.

9 FDIC National Survey of Unbanked and Underbanked Households,


PayPal – Bringing Down the Bad Guys, One Transaction at a Time


3D-Printing: A Nuclear weapon for startups willing to disrupt the vending machine business.

Student comments on Machine Learning in Credit Assessment at Capital One

  1. I thought this was a fascinating post on the use of machine learning in banking. I was reminded of our FRC case on Handelsbanken, which makes lending decisions based off a holistic picture of the borrower and differs from the U.S. market which is largely reliant on credit scores. If U.S. regulations allow for the inclusion of ML in lending decisions, Capital One seems well positioned to take advantage of this. I thought your question about bias was particularly relevant given that using past lending decisions as training data might be risky due to biases within the data both in terms of who applies for and who ultimately receives a loan. Beyond hiring data ethicists, what else do you think Capital One could do to work on bias elimination?

  2. This is definitely an interesting case where machine learning might have both financial utility for CapitalOne in addition to potentially some social impact by reaching an underserved population. I share your concern that just as Amazon tried to use machine learning to combat gender bias in hiring (and their machine kept recommending candidates who matched the profile of past hires), CapitalOne may face similar challenges with biases implicit in the data they use as inputs. I would be curious to see how data ethicists would work. Great article!

  3. KW, I would love to know what the initial estimates are for additional revenue and cost-savings for Capital One by implementing machine learning to predict new customers. I would imagine it’s current B2C business and customer acquisition costs are quite expensive. Just as an example, a friend of mine from a competitor credit card company who is the Director of B2C quoted me $1 per mail to new customers. If machine learning is able to identify the right factors that predict a good client, the business opportunity is massive. To your point around data ethicists and management of biases, use of the data in this matter could be deemed exploitative and potentially discriminatory — a significant risk that the firm cannot ignore. In your research, did you find any indication of how Capital One plans to manage those risks?

    Good job!

  4. I agree that the financial services industry is ripe for innovation driven by machine learning. I agree that the most obvious solution in personal banking is in analyzing customer creditworthiness, but I think there are other applications as well. For instance, machine learning can be used for things such as monitoring for and identifying fraudulent activity by being able to analyze immense amounts of data and uncovering patterns and trends that humans would not be able to. From a customer perspective, machine learning using natural language processing can be used to identify intent in some data sources such as emails. For example, if a customer emails their bank, a machine learning algorithm could derive the intent of the email and then direct it to the most appropriate person or department within the organization to handle it. These are just a couple of other exciting examples where machine learning can be applied to consumer banking.

  5. Great article! I love the idea of exploring machine learning to supporting inclusion initiatives. I wonder what type of activities and further benefits Capital One would be able to tap into by exploring these markets further. It seems as though machine learning would be an excellent avenue to accelerate financial inclusion by developing a better understanding of trends.

  6. I really enjoyed this article! In general, I think it makes a ton of sense for Capital One and other banking institutions to leverage machine learning in their credit assessments. While I agree that they should consider adding data sets beyond the typical credit score into the algorithm, I have concerns about the use of social media data such as Twitter followers and Facebook likes. My main concern stems from the fact that we already spend way too much time on social media, and I’m afraid that linking social media data to major financial decisions (like whether you can get a mortgage to buy a house) will only amplify that. It actually reminds me of a Black Mirror episode called “Nosedive” where social media ratings impact your socioeconomic status. I think the key is that we should strive for unbiased, objective data in these financial decisions, but social media data feels more like a popularity contest which is very subjective.

  7. Nice work on this article. Your comment about using ML to open up credit/banking access for the un/der banked got me thinking a bit. In particular, that segment is probably the segment of the population with the fewest datapoints to draw on to make a credit decision (unlikely to own property, hard to verify rent, credit scores not reliable, etc.).

    While that is a barrier to Capital One using ML effectively today due to the narrow dataset on each individual, we are likely to see a future where more aspects of our financial lives continue to move online (e.g. rent payment, car payment) and into one place (with broader adoption of services like Intuit’s Mint). If Capital One continues to push the envelope, they have an opportunity to be the first mover in the unbanked segment in the future.

Leave a comment