This piece of mine was published in The Economic Times on 7th November, 2016.
When virtually everyone claims to use “artificial intelligence”, “big data” or “machine learning” to “disrupt”/ “revolutionize” one industry or the other, I wouldn’t blame you for being sceptical about the much-touted AI revolution in Fintech. But, before you throw the baby out with the bathwater, let us delve a little deeper and understand if there is a real problem in lending and if AI can help solve it.
Is there really a problem in the banking system?
Banks and financial institutions, with some notable exceptions, are struggling with bad loans. The average Impaired Asset Ratio – the sum of gross NPAs and restructured advances (a measure of the stress on a lender’s balance sheet) stands at 12% of advances and is slated to rise to 12.5%. Notwithstanding the widespread notion that it is only corporate loans that have turned sour, retail loans, with an IAR of 8.8% and small & micro enterprise loans, with an IAR of 10.5% have contributed significantly to the asset quality woes of the banking system. Clearly, something is not going right, or as the crew of Apollo 13 would have said,” Houston, we have a problem”.
What is causing this problem?
High-profile wilful defaulters, widely blamed by hyperventilating news-anchors for creating this situation, are only a part of the problem. They account for a portion of the delinquencies in the corporate loan books of (mainly PSU) banks. Business cycles that depress commodity prices account for woes in sectors such as steel. Delays in environmental clearances or land acquisition account for some delinquencies in the infrastructure sector. There are several such cyclical and industry-specific factors that explain the sub-optimal asset quality of a portion of the corporate loan books of large banks. However, none of these factors have had a major impact on the asset quality of retail loans or micro-enterprise loans. So, what is stressing these loans portfolios?
More data, more credit but the same old underwriting techniques
Over the years, the amount of data that is available on potential borrowers has risen exponentially. The penetration of credit bureaus in India took off post the 2008 crisis and made available every little detail about a borrower’s credit history ranging from the loans he took to the number of times he missed EMI payments. This vastly improved the ability of lenders to gauge the credit-worthiness of potential borrowers. A quick “bureau-check” for as little as Rs 15 could enable banks to weed out loan applicants with poor credit histories or high levels of indebtedness. This, along with several macroeconomic factors led to a credit boom. The credit outstanding in the banking system grew from Rs. 27,70,012 cr in 2009 to Rs. 64,99,829 cr in 2015 – a whopping 135% increase in just 6 years.
While the amount of data available to lenders increased drastically over time, their underwriting techniques didn’t keep pace with changing times. Lenders traditionally relied on salary slips, bank statements and address verification to gauge the ability of borrowers to repay. The only new variable that they added to the mix was the credit-score provided by credit-bureaus.
What is wrong with traditional underwriting methods?
Traditional loan underwriting methods rely too heavily on credit scores provided by bureaus. While credit score verification is a great first step towards filtering out likely defaulters, credit scores are seldom able to capture complex patterns in loan repayments. For example, one of our pilot exercises with a large Indian lender showed that after adjusting for income and education levels, credit scores could not account for differences in repayment rates in a portfolio of loans made in a large metropolitan city.
Further, credit scores are based solely on credit-history and do not take into account rich data available on new-age platforms, which can potentially give lenders access to data points as varied as online purchases, strength of social connections and travel patterns- all of which, when viewed holistically, allow lenders to get a complete picture of potential borrowers- thus significantly improving their ability to predict loan defaults.
Also, the risk that a given customer presents to a lender can be vastly different for different loan products. For example, a 26-year-old engineer might be able to pay back a two-wheeler loan of Rs 50,000 but be unable to take on the burden of a Rs 1cr home loan. A credit-score cannot differentiate between these two cases.
Lastly, the penetration of credit bureaus in India still has a long way to go, with only about 22% of India’s adults having a credit score, according to The World Bank. Therefore, assessing the risk posed by a potential borrower without a credit score is beyond the scope of what most traditional underwriting models are capable of.
How does Artificial Intelligence help solve these problems?
Artificial Intelligence is uniquely suited to solve many of the problems associated with the traditional loan-underwriting process. Firstly, AI can capture complex patterns in customer data and exploit these patterns to identify likely defaulters. Machine learning – a family of Artificial Intelligence techniques has proven itself in fields ranging from bioinformatics, fraud detection and natural language process to speech recognition. It shows tremendous promise in the domain of finance. For example, our own technology that relies heavily on machine learning, has, during the course of a pilot exercise with a lender, been able to correctly identify up to 61% of the delinquencies that traditional underwriting models missed. More importantly, it was able to do so without significantly affecting loan approval rates. This translated into higher gross returns and improved capital efficiency metrics.
Further, Artificial Intelligence can process large amounts of data that human underwriters would simply not be able to make sense of. Imagine a thousand line items on a person’s bank statements, a hundred items on his credit information report, a couple of thousand data points from his social media footprint and call records running into hundreds. The average lender neither has the ability to access and consolidate this data nor the technology to derive actionable analytics from it. This is precisely the gap that artificial intelligence can fill.
Artificial Intelligence also brings with it, the ability to capture and exploit patterns that are unique to the loan portfolios of different lenders. Different lenders have different client acquisition channels, loan underwriting models and collections processes. These differences result in loan repayment patterns that are idiosyncratic to them – patterns that traditional underwriting models do not account for. For example, we found, during the course of two pilot exercises, that while delinquency rates were positively correlated with age and tenure in the case of one lender, the opposite was true in the case of another lender in the same segment. Our AI automatically generated different models for each of these lenders to account for these differences- something that the traditional underwriting process would have missed. AI can not only capture such patterns but also price loan products based on the risk associated with specific potential borrowers- a concept called Risk Based Pricing.
The AI revolution in lending: Hype or reality?
In the end, it is not the sophistication of technology and the number of buzzwords in a product brochure that make technology successful but how that technology translates into business value for the customer. Therefore, in the true spirit of the adage- “In god we trust, everyone else must bring data”, the relevant questions to be asked are – “By how much did your AI cut delinquency rates?”, “By how much did your AI boost Return on Assets” and more importantly – “How did your AI affect loan approval rates?” The rest, as they say, are mere details.