Author | Pankaj Kulshreshtha
A couple of decades ago, I was about to complete my PhD thesis in a rather esoteric confluence of AI, social choice theory and behavioural economics. The idea of getting into the PhD program was to get into an academic career that after four years, still felt like a natural choice. Fate if you believe in it, had other plans. A series of events landed me at an analytics center of excellence for a large financial services conglomerate. Those were early days for Analytics (now AI !).The lures of getting some practical experience building predictive models with real data proved hard to resist.
At the analytics CoE, I had the opportunity to work with some of the best minds in analytics for financial services. Dave was one of the senior leaders in the group. He was a distinguished role model for rookies like us. He had the usual PhD type credentials but he was trying to build a lending business with analytics at the center. That was the more inspiring thing about his background. Two decades ago that was as radical as the proverbial geek got. It symbolized geeks itching to get out of the basement and play in the mainstream business.
The crux of Dave's vision was what he called, the perpetual money-making machine (PMMM). This algorithmic machine- PMMM, encapsulated a bunch of propensity (response, activation, right time to call) and predictive (probability of default, spending) models. You prime the PMMM with existing credit card customer data and come up with target lists to cross sell unsecured loans. Then run several evolving experiments that over time refine the PMMM by making the underlying credit underwriting models sharper. All the operations are outsourced and managed by third parties. The more prospects you throw at this machine, the better the machine gets at churning out more profit! Asymptotically (not sure if people still use the word!), we at least have a theoretical construct to run a profitable lending business with just a few geeks at the helm of affairs.
Good bits of this vision were put in practice in a newly minted business unit. Best of our credit modelers and analysts were enrolled to build the heart of PMMM. The results were impressive. A couple of billion dollars were lent in record time as proof of concept. Several folks from Dave’s business went on annual award cruises and champagne flowed in celebratory ceremonies! Soon after the celebrations dimmed, something seemed to have gone wrong. We started noticing delinquencies rising beyond plan levels. And very soon we learnt the one lesson no financial services professional can afford to forget. Lending businesses are fundamentally about collecting. Lending is the easier part. Eventually on this portfolio we had 25% loss to sales. Dave’s business stopped originating and was reduced to a collections shop.
Now let’s fast track to the current euphoria around AI and fintech boom we are witnessing. If you look closely at lending startups, most are built on a few or many of the principles of this PMMM. Compared to two decades ago, lots more data is available and is available in a much more timely manner. That clearly will drive lot better predictions of customer behavior. As the FinTech’s get customers hooked onto their apps, there is novel new data becoming available that can be used to drive customer engagement and experience to a whole new level. The latest machine learning (ML) and artificial intelligence (AI) credit underwriting techniques enable the prediction quality to become lot sharper. I do believe that the timing is right to evolve PMMM paradigm to build new age lending companies. This serves a bigger purpose of increasing credit availability in our societies.
So, what should we do differently this time? How do we avoid the wave of irrational exuberance that can potentially bury all the promise #AI for Credit has? I share two lessons from my experience with Dave’s PMMM.
1. Can’t take out the risk tribal knowledge out of the equation
Basic econometrics teaches us that when you chain multiple models together, the errors spiral out of control. So, as we started getting better at targeting models, adverse selection issues became serious. That resulted in the PD models not predicting well for certain cohorts.
It is important to remember that losses develop over longer cycles and are fundamentally harder to predict because of long cycle macro and micro changes. It is easier to continuously test and refine response models as you see the response within a couple of months. Losses on the other hand mature up to 48 months. So, the methodology of PD model development must have a level of rigor built in. Also, an element of triangulation based on judgement and experience is required to be overlaid as new business plans are designed.
2. Operational issues can throw model predictions off the charts
Reality of the business on the ground changes everyday, while the data scientists build models based on data from a particular cross section of time. It is important to make sure that the operational realities are tied to the data used for credit modeling as much as possible. E.g. if there are inordinate delays in approving the credit applications, it is highly likely that the best of your customers are running away to competition and that over time will give rise to adverse losses.
While I have used insights from my own and a few colleagues’ experiences, please consider the story fictional and illustrative. I also welcome your comments and feedback on LinkedIn or on the Scienaptic blog. I shall attempt to synthesize some of those for collective learning in my subsequent posts. Here is to the hope that lending will find better balance between growth and risk ! Search words:ai credit decisioning ai credit platform ai credit scoring ai credit underwriting ai credit underwriting platform ai decisioning platform ai loan underwriting Ai ml credit platform AI ML credit underwriting ai underwriting platform American Banker's Association artificial intelligence credit artificial intelligence credit platform artificial intelligence decisioning platform artificial intelligence underwriting platform automated underwriting best credit bureau better credit scoring cecl compliance ml ai collection credit management credit bureaus credit decision credit management credit platform credit risk