Updated: Oct 8
How inertia impacts credit modeling methods.
Like general credit modeling techniques, Machine Learning updates credit loss understanding from more data. In supervised learning, the data teaches:
To improve and adapt the ability of the independent variables to predict a dependent variable (feature engineering) and
The appropriate underlying functional relationship best describing the independent and dependent variables. (ML Methods like Architecture and Estimators) (1)
In the following narrative and as a metaphor, I will connect physics to describe how inertia impacts Machine Learning and related modeling techniques.
In the calm environment (matter phase state = solid), best described by normal statistical distributions (kurtosis = 3) and feature independence (low inertia), a particular ML method will be chosen. This will work fine for many years, that is, until it doesn't work. Once the uncertainty of a dynamic environment is added (like when heat is added to melt ice and the matter phase state = liquid), kurtosis and inertia will increase dramatically. This invalidates the current ML method. The problem is, the new phase state will not have enough new data to validate its existence, including to update both the prediction AND the ML method until the damage has been done. (in effect, the ML method itself maintains inertia) This is not a speed of computing problem; it is a phase state transition time to complete problem, aka, The Inertia on Inertia Paradox. This may also create a damaging self-reinforcing cycle, especially if the cause of the uncertainty relates to the modeling methods suffering from inertia. The model can't update, so the issue generating uncertainty continues because the uncertainty signal is not available to the existing ML model. See below for an example.
It takes data to validate a model, it takes time to realize and collect data. The problem is - the current predictor continues to operate until a new predictor is determined. In a fast moving and chaotic environment, the current predictor is very likely wrong and inappropriate.
At the bottom of this article is our Storm Framework. This is helpful for building context to the phase change environment, like when going from calm to stormy, and then back to calm.
An Inertia-On-Inertia Paradox example: Prior to the 2007-2008 financial crisis, there was an inappropriate practice of using past mortgage product loss models (dominated by lower risk fixed rate mortgages) to predict losses on higher risk negative amortization mortgage products. No one believed a mortgage product could have loss rates in excess of credit cards, until they did. The self-reinforcing negative cycle occurred because the true credit loss signal associated with higher risk loans was not available to the models. It was not until the housing market collapsed that the credit risk was revealed.
Perhaps, the message is, credit policy should be stable. It should not be used as a tool to expand homeownership. Credit policy should not contribute to increasing housing demand. Policy makers should focus substantially on housing supply, with credit policies only adapting to available demand. This case is made in our article The affordable housing paradox.
The Storm Framework - a metaphor for uncertainty and risk:
(1) A Survey of Machine Learning in Credit Risk, Joseph Breeden, May, 2020