The tremendous societal benefits of effective credit risk assessment and a fresh perspective on how to improve it

In this article we outline the various benefits of doing effective credit risk assessment, which, although may be well-known, we wish to make an effort to collate them and re-appreciate their grandeur in terms of the impact they can have on the world. Furthermore, we show how increasing the accuracy of the risk assessment has a direct impact on such benefits. This will establish the need for continued attention to this important problem of credit risk assessment, both in industry and academia.

We will also touch upon the aspect of credit scores versus risk assessment, which we believe is a source of major confusion leading to several sub-optimal credit/lending decisions all over the world, both at the retail and institutional levels. We propose a simple yet extremely effective approach towards credit risk decision making via two quantitative signals, one that provides a view of the past and another which provides an estimate of the future. The delinking (of the past and the future) while may not be necessary mathematically for making a good decision, plays an important role in addressing the behavioural loopholes that persist.

We* hope that our proposal helps both the lending institutions and the regulators build and maintain a near-optimal credit ecosystem for a better world.

* I prefer to use “We” instead of “I” when writing articles even when they are single-authored.

1.    Introduction/Background

Credit risk assessment deals with the problem of assessing whether an individual/institution is worthy of being awarded a certain amount of credit. At a business/institutional level, this problem is almost as deep as an investor investing in the institution, albeit with lesser connect to the overall success of the business and to the extent of profit share among other things. Specifically, the problem statement is restricted to the amount of credit, the tenure, the rate of interest, and the schedule of payments. The challenge here is to accurately assess whether the institution can or cannot honour such an agreement, i.e., whether the institution/business is strong enough to meet such a payment schedule. Whereas an investor would evaluate many more dimensions based on the ROI expectations with which they approach such a business and the strategic advantages it would bring to their portfolio. Nevertheless, even a creditor needs to do much analyses in such cases to come up with an effective risk assessment, often case by case. We will not further discuss or speculate about credit risk assessment at a business/institutional level in this article except for our recommendation at the end which is inspired by our experiences in the retail/individual credit risk assessment domain. It is upto the experts to consider it further if it appeals to them.

At an individual/retail level though, things tend to be much more under control/controllable, which closely reflects the state of the society at the point in time where one chooses to analyse or act. The problem statement in this case is: given an individual, assess whether she can be given a certain amount of credit. The problem involves, the purpose for which the credit is sought, the amount of credit, the tenure of the loan, and the payment schedule (often in the form of EMIs). Typical purposes for which a loan is sought by an individual are, agriculture, education, buying a car, motorbike, consumer durable, home, or just personal loans for various other reasons such as medical expenses, holiday trips or marriages etc. One can easily imagine that a better access to credit will have a direct impact on the economic activity in any country/world due to the spur it creates in various activities for which individuals seek credit as mentioned above.

Therefore, the better credit access the people have, the more and faster growth a society can achieve. However, if the creditors/lenders do not perform proper due diligence when awarding such credits, things can quickly degrade making the credit assets – non-performing (NP / NPAs), resulting in losses for the lending institution and at a higher level to the overall economy.

2.    Benefits of effective credit risk assessment

It is useful to understand the basic framework of a lending business to fully appreciate the amount of opportunity and benefits which can be leveraged through proper risk assessment. The mechanism is quite straight-forward, a lending institution obtains license from the regulator (for India – RBI) to start lending, follows guidelines laid down by the regulator when framing their lending policies, and then goes about building a credit book of a certain amount, say, INR 1 billion. It means that the lender has decided on a certain strategy to give out loans worth INR 1 billion, which will also typically be associated with restrictions and requirements such as:

• type of loan, say, personal loans only
• loan amount range of say, INR 50-100K
• book should be built within 3 months
• tenure can range from 12-24 months, and
• a certain interest rate R

The interest rate R is very important and interesting, it is decided based on the estimate of the delinquency/NP rate of that credit book (which also becomes the target delinquency/NP rate for that credit book), which is typically based on past experience/data. (Many other factors also get into play, such as cost of funds, amount spent on collection frameworks, processing fees, etc. which typically are relatively easy to estimate accurately).

Presently, lenders also have books that are continuously built (that don’t have fixed timeframe or total credit limit) whose parameters are continuously updated based on the feedback observed over their active lifetime. Nevertheless, the simple model which we described above helps appreciate the benefits and scope almost equally.

The benefits of accurate credit risk assessment:

• A more accurate estimate of the delinquency rate of a certain credit policy can result in a much more competitive R which helps the lending institution attract more customers and the customers themselves now availing credit at a cheaper price, which will have ripple effects leading all the way to lower inflation and better growth
• The individuals themselves can have lower payments to make, easing their pain in some cases or encouraging them to take risks more often/dream big in some other cases depending on the purpose of the loan
• At a personal level, many individuals have extremely empathetic and inspiring stories to share where due to better credit risk assessment, large segments of population which were denied credit previously are for the first time getting access to credit, or, getting access at interest rates which are 2 to 3 times lesser than their previous arrangements
• Furthermore, an individual who may be knowingly/unknowingly trying to bite off more than she can chew actually can also get a reality check/feedback from accurate credit risk assessment, where the assessment can act as a financial guide at that moment

The benefits mentioned above are only the beginning of a longer list one can think of. More importantly though, the benefits above cover not just the benefits of having easy or more access to credit but also show the benefits of not having easy or more access to credit when an individual is knowingly/unknowingly getting into the abyss of excess credit without sound support of cashflow and/or other parameters.

The same applies for lending institutions, as they also have the benefits of both reaching out to more population segments through better assessments and also the benefits of protecting themselves by not reaching out to individuals whose future loans are more likely to turn NP than not.

We emphasise the “not” aspect or protection of individuals/institutions (from themselves and one another) as it is often not highlighted enough as a benefit of better credit risk assessment (which is often overlooked as just giving more credit or building a larger credit book).

3.    Proposals for improving credit risk assessment

In this section, we first present a succinct framework for evaluating a credit risk assessment model which makes it easy for decision making as well as for contrasting different models (one may skip this in their first reading and revisit later). Secondly, we highlight the issues raising from confusing credit scores with risk assessment estimates and suggest a way out with new policy guidelines.

3.1. Representing performances of credit risk assessment models

Figure 1 shows a way to represent the performances of risk assessment models. Given a credit book, which would contain users who are all given loans (most likely for similar purposes and similar amounts and tenures) within a bounded period of time, a risk assessment model would try to predict for each user, the probability of the asset given to that user becoming NP (non-performing). This could be approximated to say, the user surpassing non-payment for 90 days past due (dpd) in 12 months on that book (MOB).

Figure 1: % NPAs among Top X % users, a risk assessment model in perspective with an ideal model and a random model

Given the data about users belonging to a credit book of say, Jan 2017, any risk assessment model would output a probability of NP for each user on that book in the range of [0, 1] which are generated solely based on data available about those users as of Jan 2017. To test whether the model is performing well, one can index/rank the user data points (or representative IDs) in the increasing order of these probabilities and analyze the % NPAs in any top x% of users, based on the actual performance of those users on that credit book observed during Jan 2017-Jan 2018.

In Figure 1, note that, there is an overall 4.2% of users whose loans have turned NPAs. Therefore, a random model would have a performance curve which is a horizontal straight line with y-intercept at 4.2%. And, an ideal/optimal model would have a performance curve which follows x-axis until 95.8% (until which point all its top users have zero NPAs) and then starts growing linearly towards y-intercept of 4.2% as each new addition from there on would be an NPA. In reality, any risk assessment model would fare far better than a random model and at the same time not anywhere close to the ideal model (as of today) which one would aspire for. This makes it a very challenging and engaging problem for both industry and academia as the underlying problem is technically difficult, socially beneficial, and has sound business implications.

Figure 2: % NPAs among Top X % users, comparison of two risk assessment models

Figure 2 shows the performances of two risk assessment models (other than the ideal and the random ones) where the Model 2 outperforms Model 1. In fact, the benefits of better credit risk assessment of the new model are huge. For instance, if a lender wants to contain their NPAs at 1.5% levels, then using Model 1 they can only shortlist ~15% of users, whereas using Model 2 they can shortlist ~75% of users which results in 5X more business for the lender. Also, it means ~60% users who were not getting loans previously are now getting the access to credit that they may need in their daily lives, whose qualitative impact on their lives is immeasurable.

           Alternatively, if a lender wants to approve/select only 50% from the population under consideration, then Model 2 can help contain the % NP at 1% as compared to 2% of that of Model 1. Furthermore, analysis of the above charts can also be used to design variable interest rates for different segments of population or even personalize interest rates at an individual level, thereby increasing the competition in the industry as well as increasing transparency in passing any potentially lower cost of funds available from central banks to the end customers, as per their individual credit risk assessments.

Over time, many approaches and methods were explored beginning with statistical modeling, statistical learning, and machine learning in today’s time. Even with all these explorations involving increased digitization, new ways of collecting data, representing data and modeling the risk, the task of building credit risk models continues to be challenging and at the same time realistically engaging in terms of surpassing previous benchmark performances, etc. The performance of Model 2 in Figure 2 actually corresponds to one of the state-of-the-art models which is available in India (and the performance of Model 1 corresponds to that of one of the state-of-the-art models from last year), which showcases the huge leaps of advances being made in risk assessment and also the amount of further scope present for improvement. We do not delve into the details of how these approaches work and potential directions for future approaches in this article. However, we do believe that with the current technological progress in terms of data availability and collection, algorithms and infrastructure, there exist very realistic chances of building models which are near-optimal very soon.

3.2. Credit scores vs. Credit risk assessment estimates

We now shed some light on the distinction between credit scores and credit risk assessment estimates, and the need to appreciate the difference and a way forward.

A credit score is traditionally expected to be indicative of whether a person is creditworthy or not, potentially to extend a new credit line in the future. At the same time, it is also expected to be indicative of how well the person has performed on the credit lines extended to her in the past. This presents a contradiction in expectations since a person with good credit history may not necessarily be automatically qualified for a new credit line. Assuming that a person with good credit history will continue to be good with any future credit lines would be a classic example of an error like “hot hand fallacy” introduced by Amos Tversky, Thomas Gilovich, and Robert Vallone (1985) and discussed later by Prof. Richard Thaler, the 2017 Nobel prize winner for Economics (although the concept was originally analyzed w.r.t. random sequences and basketball players). For instance, if someone is paying a number of dues/installments on past loans extended to her with minimal savings left, then it would be risky to extend a new loan to her even though her credit history may be very good. (Indeed, many successful lending institutions do this due diligence beyond the credit scores offered by various bureaus, however, there is no clear regulatory obligation to do the same which we recommend for the future).

In reality, most credit scores which are available today end up becoming a mix of sorts of a person’s past credit behaviour and her future creditworthiness without doing proper justice to either of them. Figure 3 shows the current state of various credit scoring models. For instance, guidelines like, the more the number of credit lines you have the better score you would have, the more the credit line utility one has the better the score will be, have become rules that everyone is getting forced to adjust to, even though such rules do not make any sound logical sense except for the people inside the industry building those models/rules who are stretched with requirements between the two sides of the axis shown in Figure 3. We must move towards a better system where people are not forced to take up more credit cards or adjust their spending patterns just for the sake of getting high credit scores even though they might be paying all their dues on time. And anyone who pays their dues on time on all their credit lines, even if it’s just a single credit line, should get a high CHO score.

Figure 3: Representation of various current credit scoring models on an axis that extends from a CHO model which is purely representing past credit behaviour towards a CRA model which is purely representing future credit risk assessment.

In the following, we will discuss why having two distinct scores one representative of a person’s credit history and the other representative of the person’s future credit risk is beneficial in many ways (than any of the extremes or a single combination as we see today). The reasons go beyond mathematics and touch upon the various use-cases where such scores are used and the human behaviour itself. Also, while the CHO score can be updated regularly and enquired for any number of times without effecting that score itself (which actually happens with many current credit scoring models—the number of enquiries effects the score inversely), the CRA score can just be computed only when a user applies for a particular loan and can include the parameters from her application form and more as of that time.

Here, one may be tempted to suggest- why not completely let go of the credit history observation scores and only have a score corresponding to one’s future creditworthiness. One objection from customers would then be that, if they paid all their dues on time, why are they being represented with a potentially low score. Even though one may try to explain about the way it is measured, i.e., estimating for the future, one may face hard time convincing people. A stronger objection though, is that the credit scores which are representative of the past have their own utility in various places. For instance, when one is moving from one rental home to another, or when someone is looking for new utilities connections, etc., many cases where there is no large credit line being extended, where there is no critical necessity for doing a more complex and expensive credit risk assessment, the score which is representative of the credit history more than serves the purpose.

However, as discussed before, using a score which is representative of the past for extending new credit lines into the future would not be sound. As many successful lending institutions already follow, it is advisable to have a separate risk score which estimates one’s ability to repay future loans extended to her. Even with the institutions which have such a mechanism in place, the motivation to develop such scores didn’t necessarily stem from the above observations, rather it stems from the observation that the scores available from various bureaus were not performing well (or doing very badly in some cases) in estimating NPAs upfront. In such cases, we hope that the above discussion helps understand why such scores were not performing well and cannot be expected to perform well for estimating future NPAs unless the above proposed distinction is implemented.

Let us take a more detailed look at each of the four cases that arise with having two scores:

1.    High CHO score and high CRA score. A person has a good credit history and also has many other parameters in right place to support a good future credit extension as well.

2.    High CHO score and low CRA score. A person has a good credit history, however, the support for a new credit line extension is extremely weak. This case would generate a lot of confusion with current credit scoring models forcing them to give a score somewhere in between depending on which part of the axis they fall into as in Figure 3. Such a confusion neither serves use-cases like getting new rental homes or utility connections well (which should ideally be straightforward for such a person) nor serve a purpose like getting a new loan (which should not be extended ideally, whereas some current models might suggest otherwise).

3.    Low CHO score and high CRA score. A person who has missed on her payments in the past (maybe a while ago) might end up with a low CHO score (depending on the model), while her current parameters may reflect a high CRA score. We believe though that this possibility is less likely since good CHO models would already try to incorporate the intensity of credit defaults and the timing of them, as would good CRA models which usually have correlations with such parameters from the past. (This type of cases if observed in large numbers may mean either the CHO model or the CRA model being used (or both) is faulty/needs to be improved.)

4.    Low CHO score and Low CRA score. An individual who has credit defaults in the past and no indicators showcasing support otherwise for the future either.

We hope that regulators note this distinction and consider enforcing lenders to measure and note down quantitatively, two distinct metrics about each applicant/application, one that is representative of the past credit behaviour and one that estimates the future creditworthiness. A low future score should definitely act as a warning signal for any new credit line extension.

Before we conclude, we also wish to recommend similar adoption of two distinct scores even for lending at an MSME/big business/institutional level, and the enforcement of the same by the regulatory authorities. This again can help avoid over optimism leading to higher NPAs, such as the ones described by Prof. Raghuram Rajan, Ex- RBI Governor, in his note to the parliamentary committee of India on NPAs -- “One promoter told me about how he was pursued then by banks waving checkbooks, asking him to name the amount he wanted.” Such instances if not well-deserved, will automatically face the reality check when the lender tries to put the second quantitative estimate about the future potential of the candidate/application, especially knowing that the same would have to be explained in the future in case the asset turns NP. Just by enforcing such a metric forces people to think realistically about the future creditworthiness and the associated modeling instead of just the past (while also relieving the pressure of assigning a high score for the future just because someone might have a good credit history).

4.    Conclusion

We have presented various societal and economic benefits of good credit risk assessment, and the need for continued attention in industry as well as academia to better perform the same. With newer technologies enabling increased digitization, gathering of large amounts of data and new types of learning algorithms and infrastructure fuelled by the AI revolution coming in, this important problem can be solved in ways that were not previously envisaged and presents an opportunity worthy of our attention. We further present an effective way of representing the performances of various credit risk assessment models which helps in decision making in various use-cases as well as in contrasting multiple models with one another.

Lastly, we argued that there is a need for two distinct scores one that is representative of a person’s past credit behaviour and another which estimates the person’s future creditworthiness both of which solve different sets of purposes and using a single combined score would fall sub-optimal in serving either of those sets of purposes. We hope that our suggestions appeal to the lending community, the regulatory authorities, as well as to the people in better understanding and harnessing the power of credit for a better world.

Note: At this point we have not completed a thorough literature survey to contrast with other publications on the subject, which we plan to introduce in a later version of this draft.

Disclaimer: The views expressed in this article are my personal views and may not represent the views of my employer.