Credit risk behavioral PD

Home - Article - Credit risk behavioral PD

Hello everybody – we do hope this article finds you well.

Now that some of the loan applications were rejected but others passed the bar successfully (check out our article on Application PD scoring) what does that mean to you?

Two main things:

  • You’ve experienced outflow of cash while disbursing the loan principals and you need to manage the hence emerging risk;
  • You’ve successfully signed a deal with each and every one of your customers meaning mutual trust was achieved. Now you need to manage that trust and even further develop it while, if possible, improving your bottom line.

So we are in the next stage of our customer life cycle (check out our article on Holistic Advanced Analytics within the Finance/ Fintech industry) – portfolio management.

What can we do to manage our customers proactively – meaning while they are still “current” in their deliquency – and why do we need to do that? Here are some real business life ideas:

  • Circumstances change and people financial status might shift negatively in time. Hence, we need to realize proactively customers tendency to start struggling to cover their installments and make amendments – offer restructuring of the loan, for example;
  • Realize up-sale/ cross-sale potential and go for it but only where it fits – maintaining risk level and not ruining customers’ experience;
  • Gauge the risk of your portfolio constantly to observe unwanted shifts. Knowing your current portfolio risk on a monthly bases can actually be a regulatory requirements (Basel, IFRS 9…) and directly affects your provisioning policy and hence profit/ loss balance sheets.

I think we all got the idea how important is to address the measurements of the risk of our portfolio and even more so – address it in as accurate way as possible.

One way to achieve this and based on our experience a great way is by developing an advanced analytics Behavioral Probability of Default model based explicitly on your data -> hence customers -> hence their behavior over your custom products.

Below is a set-up of the analytical exercise where we convert the business case into a Machine Learning (ML) one and the result of the latter is integrated into our CRM strategy:

No alt text provided for this image

As usual, let’s deep dive into each element of the above-describe analytical set-up:

1) Understand (and bring to the surface) your pain-point and hence your business case:

Your portfolio experiences constant need of being gauged in terms of risk exposure and its fluctuations over time. Know that accepting more end customers is great – it increases your market exposure and the potential for profit – but you may also easily turn-out being understaffed. This, in turn, will deteriorate the proper management of all the money already granted. Automation can come to the rescue here.

2) Find the right data for this problem:

Just like with all Data Science exercises – you need data. All the data you used when developing your application PD is valid here as well (you can actually use the Application PD scoring itself as predictor) but you need a couple of things further:

  • Observation point (e.g. monthly month end) delinquency data: days past due (DPD), default flag, etc.;
  • Historical cash-flow transactions data and observation point principal and interest exposure information;
  • Ratios and changes in behavior, e.g.: utilization ratio (of a credit line) and it’s dynamics over time;
  • Anything that proves or is given buy the business representatives as relevant information.

3) Exploratory analysis of your portfolio:

Understand your current portfolio state as well as history – financial and segment-wise. This will give you the opportunity to even further specify the target of your ML model as well as create proper expectations.

Example segment: new vs sequential customers (customer for which this is not the first product with your company).

4) Convert the business case to a Machine Learning one:

In our case – a binary Good/ Bad target needs to be developed. We do advise that the target definition is analyzed and the best is hence selected – 90+ DPD is a great standard approach to go for but your products and customers might prove otherwise.

5) Develop strategy and integrate it into the decision engine:

Building your ML model is not an end in itself (yes, we know we are repeating ourselves but it only stresses how important this is 😊). Make sure you do not miss the bigger picture and see to how your model: integrates in your current strategy, affects your personnel and how they can complement each other, and how it fits into your current/ future IT infrastructure (make sure you do not miss our article on the topic).

Based on our experience with solutions of the kind below we present one way to address the “bigger picture”. The example is for a revolving line of credit product in the retail sector. It aims to pin down customers for which limit increase offer is to be made:

No alt text provided for this image

Only when we combine all-of-the-above – low risk customers with high recent utilization -our limit increase offer starts making sense.

6) Improve your processes, customer experience and bottom line

Now that we are having all in place and fully functional what have we achieved? Well, let’s list some things that we have seen from our practice:

  • Scoring system with high level of interpretability. Business validation of the drivers of customer behavior based on the modelling exercise;
  • Increased conversion rate of the limit increase offer by 30%;
  • Increase in the overall portfolio while maintaining the risk;
  • Automation of the campaign which allows your personnel to spend more time on customer interaction rather than pre-selection for the campaign to be.

Not bad, ah? 😊

With this series our aim is to increase data science coverage and to make data-driven decision integral part of more and more companies around the Globe.


More to read