Machine learning can cut credit risk

It can help predict customers who may not pay up and raise early warning alarms if recoveries may not happen as planned

By Author Ramakrishna Prasad|Raj Swaminathan   |   Published: 15th Feb 2018   12:12 am Updated: 14th Feb 2018   11:50 pm

A Bank is a place that will lend you money if you can prove that you don’t need it – Bob Hope
While this statement is amusing because it is so real, it is neither good banking nor profitable banking.

Ability to pay and willingness to pay are the two sides of the same coin yet poles apart. Will the borrower pay or not pay? Banks, NBFCs and credit card companies have been investing in millions over the years to solve this mysterious question, yet the answer is still eluding. It would be very fair to state that search for the right answer will always be a work in progress. But there is some good news.

Machine learning
Representational Image

Revisit Scoring Models
The first thing that the bank or a credit card company does to measure the creditworthiness of any individual is to use a scoring model. Based on that the risk factor attached to an individual is calculated and the loan/credit card is disbursed. Some customers may turn delinquent over a period and it is only discovered once the customer stops paying.

Let’s assume for a moment that the current models are doing a good job of scoring the customers but then at an operational level, there is a three-way mismatch:
1. Consumer acquisition cost is increasing
2. Delinquency rate is increasing
3. Cost of collections is increasing

While it is hard to accept for a bank or a credit card company, this is the stark reality. When things are not right on multiple dimensions, then it’s time to re-look at the basic decision-making premise — the scoring models. One way is to use Machine Learning models.

Solving the Problem
Today, we have volumes of internal and external data. Machine Learning models can help decipher the relationship between multiple features and help arrive at a much better output. The best part is that these models get fine-tuned with time and with additional inflow of data. We get to understand key customer behavioural trends around frequency and recency of events, which are important for the assessments we need to make. These models can complement the credit risk teams in their decision-making, to start with and over a period develop into primary tools for judgement.

Now that the first hurdle is cleared, it’s time to find a strategy to cross the second hurdle, willingness to pay. There are no easy answers but machine learning can help understand by looking at deeply hidden patterns. There are two ways to solve the problem:

1. Going for breadth – There are fintech companies today that use social media data to understand customers’ profile. This could be a good start but then attempting to gauge someone’s ability to pay and willingness to pay based on the connects, likes and dislikes may not be an entirely robust approach.

2. Going for depth – This is a harder option but can yield better outcomes. Ability to maximise this approach requires in-depth domain expertise and the ability to leverage the power of neural networks. A skill that is not easily available in the market

Going for depth will be a more scalable and sustainable solution than going for breadth. More data of a different nature need not necessarily mean better insights. Plus, it can lead to cost escalations.

Recommendation Engines
Once a consumer joins the eco-system, then the call centre operations take over and engage with the customers. These operations may be internal or may be outsourced. There is ample scope to improve the cost and effectiveness of collections using predictive analytics and improve recoveries. But the path to reap the benefits must be carefully crafted.

In other words, the solution must be customised and built after studying the data. This is where machine learning models and deep learning models like Natural Language Processing (NLP) can be applied again to full effect by designing recommendation engines that can not only predict customers who may not pay up but also raise an early warning alarm if the recoveries are not happening as per plan.

Unlike today, there is no need to deploy people to extract reports. The system will calculate the performance real time. This is where the call centre operations move from being reactive to being proactive.

Know Your Customer Better
For all these years, banks and credit card companies talked of Know Your Customer (KYC). Time has now come to talk of Know Your Customer Better (KYCB).

The ideal scene would be when a customer walks in for a loan or a credit card, there should be an engine that gives the probability of willingness to pay and ability to pay. Based on that, the loan amount, the rate of interest, the product type to be sold are generated by the recommendation engine. So, a go or no-go decision is made, which effectively reduces risk exposure.

Over a fixed tenure, the customer’s interaction or as we say the payment behaviour pattern is observed and a behaviour score is calculated. This score decides whether to sell or not to sell or even what to sell and at what price. This improves the conversions, reduces costs and enhances the profitability of every relationship and, of course, the whole portfolio.

Finally, the collections agencies performance is measured real time to ensure that debt recoveries are moving as per schedule. Since the recommendation engines are real time, there is luxury of time to plan appropriate actions.

Human behaviour constantly changes and there will be Black Swam moments. But if the ability to pay and willingness to pay conundrum can be solved, the damages can be minimised. Banks and credit card companies need to take this path.

(Ramakrishna Prasad is Chief Data Scientist at Indussoft, [email protected], and Raj Swaminathan is CEO, Indussoft, [email protected])