Some non-obvious learnings from 6 years of building lending products as an finance outsider (tech) #Fintech - a thread 1/n
At its core - lending is game of moral hazard, statistics, information assymetry and adverse select. Of course there are other parts like product, distribution, cost of funds, ops, collections but those are the obvious aspects. Maybe a separate thread. 2/n
Moral hazard - you as a lender have more to loose (cash) in this transaction(loan) than the borrower. Bureau score, reputation, usecase, collections all try to minimise this moral hazard. Remember that In lending your product is MONEY, everybody wants it. 3/n
Statistics - if you *randomly* lend money to 100 people in India then 80 will give you back your money, without any underwriting. But a 20% npa is not an acceptable biz metric (more later) 4/n:
A credit model attempts to find statistical evidence to select a sub group out of this 100 where the probability of default (NPA) is within acceptable limits. The best model selects the largest sub group. 5/n
Credit models are often broad and not fine tuned for a given product and time period. A credit policy attempts to apply a model to a specific product and take a credit decision. 6/n
In other words credit policy attempts to overcome shortcomings (forced and unforced) of the credit model. Various Interactions between the model and policy will probably need a separate thread. 7/n
Credit decisions can be boiled down to two simple aspect - intent and ability. WILL the borrower pay you and CAN the borrower pay you. Not always true but broadly - model covers intent and policy cover ability. (Increasingly untrue nowadays) 8/n
Information assymetry - models and policies need data. If you knew everything about the borrower you will take the perfect credit decision. Getting everything is impossible so we try to make the information gap as narrow as possible (UX permitting). 9/n
Not everyone your model selects will want or take your loan. This response pool (people who actually WANT money) is not randomly distributed. So there is a inherent higher risk in folks who want the money. 10/n
Adverse select - Not everyone who wants money will take YOUR money. Your UX, process, pricing decide who out of the response pool actually takes the loan (through-the-door TTD). Important to know the difference - response risk is not the same as adverse select risk 11/n
Over time the TTD will drift away from the original population over which the model was built. This is measured by the population stability index (PSI). Short term you respond to this by policy changes but over time you will rebuild the model on the TTD once data is avbl. 12/10
Recap - Your model+policy decides which 100 to give a loan too so that not more than 5 will default. Response and adverse will decide how close the TTD is to the original 100. If 20 take your money and that includes the 5 that the model predicts will default then.. 13/n
...you end up with a 25% npa portfolio aka SCREWED. Thanks for reading and I hope this helps n/n.
