I’ve been hunting down articles on Covid’s impact on bank analytics, I found the linked piece particularly interesting. (I should say that if OakNorth really can predict SME credit performance during the pandemic then they deserve to achieve a stratospheric valuation.)
Anyway, there’s a discussion in the middle of the piece about forward- versus backward-looking models that piqued my interest. Two quotes, both by OakNorth representatives, stand out:
(referring, presumably, to other peoples’ modelling efforts…)
“…These backward looking models are designed to tell you that you have a problem, they’re not designed to tell you what you do about it”
“Correlations in this cycle are totally different- if you don’t have a forward looking way of looking at risk, you’re not looking at your risk properly”
What exactly constitutes a more “forward-looking” model? I’ve used this term myself (and heard it a million times) but seldom paused to really think it through.
Let’s do that now.
All data, no matter how “Big”, are historical. Every model ever built, be it formal or informal, linear or nonlinear, Bayesian or frequentist, AI or non-AI, is some kind of mathematical transformation of the available historical data.
There’s nothing “forward-looking” about anything so far.
To my mind, the model becomes “forward looking” if it is specified to minimize some kind of ex ante prediction error.
The issue is that even very simple loss forecasting and credit scoring models can meet this criterion. When being built, so long as the modeler’s focus is on performance during an out-of-time holdout sample, they can justifiably claim “forward-looking” status. The models can then be monitored by looking at how well last period’s predictions stood up during the subsequent era.
I suspect that most readers will think of their own models as forward-looking.
If you have two models that both meet this definition, what makes one more forward looking than the other? I’m not sure this even makes sense. The model either meets the definition or it does not. One of the models may be more accurate than the other across a particular forecast horizon but this doesn’t make it more forward-looking.
So, what am I missing here? Is “forward looking” an oft-used but ultimately meaningless descriptor or is my definition off-kilter?