Variables displaying a wrong way sign contribution to a model are too often dismissed without further consideration
If sign remains wrong after all due diligence it should be welcomed as an opportunity to deepen our understanding of the model
Long run multiplier effects
Takes the example of the unemployment rate as unique factor to model delinquency
The initial impact of rising unemployment rate is larger than the long term impact of unemployment on default probability
Should thus expect the lagged unemployment terms to take negative signs compared to the instantaneous unemployment variable which takes a larger positive coefficient
by including the lagged effects we can boost the scale of the short term effect giving higher stressed projections
In many situations it is enough to demonstrate that the total effect of a shock “goes the right way” and that the model fits the observed data well
Omitted variables
Management actions related variables are correlated with economic outcomes
Omitting those variables will bias the effect of the economic variables on performance
Moody’s view is that they should be included for stress testing and model should take into account of past management shifts
Ceteris paribus conditions
Multicollinearity is a fact of life
The FED’s scenario reflect the fact that a rapidly rising unemployment rate will be accompanied by commensurate shifts in all other aspects of the economy.
The direct result of rising cost of money should be to increase losses
Intuitively we should see positive signs on interest rate variables in as standard PD regression model
Higher interest rates are symptoms of a booming economy
Strong correlations are generally far more effective in reducing prediction errors than weak causations
Interest rates in credit loss models have the advantage of being reasonably easy to forecast
Model with negative interest rate coefficients should not overly trouble stress-test model validators
Conclusion
Banks generally do not move in lock step with the economic cycle
Models that are not permitted to capture these timing effects and casual relationships are next to useless for assessing bank capital adequacy
In a situation where model validators and Fed regulators are insisting on tiny models that assume coincidence with the economic cycle is an elaborate form of economic window dressing
We need to move to a situation where stress test models at least attempt to shine light on sometimes complex underlying processes even if the ban manager
s are sometimes left scratching their heads about the data.