Technology <> Capital | Credit Risk
In my previous note, I had shared that organization structure and talent strategy needs to evolve top-down and left-to-right to truly benefit from technology. From command and control to problem solving teams. (https://www.linkedin.com/posts/activity-6769140107048951808-fGHS).
In today’s post, I discuss the role of technology (including data science) to improve capital allocation and capital formation through better credit risk management. Why credit? Because that perhaps is the single most important asset class in the world (arguably next to real estate). As per some estimates, global debt outstanding could be 100 times India’s GDP in US dollar terms. Imagine what a 1% improvement in debt capital allocation could do (*).
This is a two-part note and I will lay the foundation in this part. Before moving to the technology bit, let us set level on what good credit risk management is about.
In my view, it encompasses three key principles:
1. Ability to unpack and locate exact risk(s) in a transaction
It is not enough to simply talk about ability and willingness to pay and describe it as one amorphous credit score. It is crucial to unpack the score into specific level drivers of ability and willingness of that specific borrower. Further, it is important to locate those specific drivers which are most likely to generate adverse outcomes. In an environment where the dynamics of the underlying population may be changing (e.g. post Covid), these elements assume even a greater importance.
2. Ability to scale risk policy to achieve meaningful scale relative to size of the population/asset
As the portfolio size grows(mathematically), the likelihood of the portfolio demonstrating behavior akin to the underlying population increases. Sub-scale portfolios can result in unreliable outcomes through the cycle, unless they consistently contain only superior risk samples. This is not something that I would bet on — at least not place a very large bet.
3. Ability to ensure granularity of individual bets relative to portfolio size
Imagine you have an underwriting model that has a 90% chance of getting a credit decision right. Now, picture that you only have enough capital to give one loan. It is quite possible that this one loan might belong to the adverse 10% category — hence, your business might shut down even before it begins. If you could give 1000,000 loans, the likelihood that you will make money 90% of the time is higher.
As leaders and investors, spending any amount of resources on technology (and data science) in credit risk may not yield desired results if the investments are not directed towards improving your position on the above stated principles.
Let us now focus on what really happens under-the-hood when technology is used thoughtfully.
Broadly, credit underwriting in India can be plotted between two extremes — fully templated and fully assessed. In fully templated, humans are not really required. Every data is reliable and doesn’t need independent verification. Every rule, nested rules and variations are clearly laid out. No real time judgment is involved. If a loan fits the parameters, then it is a go, else no-go. In case of fully assessed, human judgment is vital. Every loan is custom assessed.
Like technology in operations, India has so far seen the highest technology adoption in score card / model building near the templated end of the risk spectrum. Why? Because it is the easiest to adhere to the three principles I mentioned earlier. How? By deploying technology near this end of the spectrum — which mostly comprises loans to salaried employees with stable jobs or well qualified self-employed professionals with a clear track record.
Underwriting in typical templated segments demonstrate following characteristics:
1. Scorecards can very reliably unpack the overall risk into neat sub-drivers such as repayment track record, credit seeking behavior, debt bearing capacity etc.
2. Scorecard driven policies can scale easily because templated products form a large asset class with a relatively stable underlying population;
3. Templated bets are typically granular given the average ticket size
Therefore, it has made sense for leaders to move away from conventional underwriting models to scorecards near the templated end of the spectrum.
This attempt has been successful in Indian private sector banks and for some large NBFCs which could readily access the customer segment suitable for the three stated rules. Of course, some enlightened public sector banks are also investing here.
When some of the FinTechs attempted this in thin-file, sub-prime or new to credit (NTC) segments, the experiments were not smooth. Not surprising, because, some or all the three conditions are hard to adhere to in the thin-file / NTC segments — even if the products are templated. Especially, if humans are kept out of the loop.
However, in the next wave, it is the assessed end of the credit spectrum where deploying technology can unlock the most value because of two primary reasons:
(a) The ‘human’ factor: Humans are typically not the best at probabilistic decision making. In the assessed end of the segment, this issue is further exacerbated. There is a risk of wrongly interpreting the data and wrongly calculating the probability of default.
(b) Nuances of the ‘assessed’ end of the spectrum: It represents majority of the Indian borrowers (individuals and businesses) who are not perfect digital citizens. Further, they don’t have a massive independently verifiable digital footprint.
This is the next wave of deploying technology to enhance efficiency of capital allocation and capital formation.
How do you do this? More on that in Part-2.
Meanwhile, if i were focused on the templated end of the spectrum either as an employee, CXO or investor — i would ask myself:
How does my underwriting fare against the three principles?
(*) I generally despise this type of math. Take a very large number and multiply with a very small number to still be left with a very large number. Such math can justify almost anything without any real insight. The only situation where I am comfortable with such math is when the numbers have a real meaning and not scenarios / imaginary. In the case of debt, a 1% number in credit cost terms is as realistic as it gets.
PS: All thoughts are personal
