On the adverse consequences of perceived model complexity

One of the key challenges in building these models is finding the sweet spot between simple models and complex models. In that light, parsimonious model selection methods such as Occam’s razor have become increasingly popular tools for model developers in financial institutions. However, properly applying these principles of parsimony is often made difficult due to the discrepancy between perceived complexity and actual complexity.

In many financial institutions, key decision-makers in upper management (oftentimes without a quantitative background), prefer models that they perceive to be simpler over models which they perceive to be more complex. At first glance this inclination seems to be in line with commonly used parsimonious model selection principles such as Occam’s razor or minimum description length. In many cases, however, the model that is perceived to be more complex in reality contains an amount of parameters that is equal to (or in some cases: lower than) the model that is perceived to be simpler. Next to that, the (supposedly) more complex models are often easier to interpret and implement, capture more sophisticated dynamics and require fewer assumptions than their supposedly ‘simpler’ counterparts. Hence, the discrepancy between perceived and actual complexity often leads to a sub-optimal model choice. To demonstrate this fact, let us first consider a simple example of this phenomenon in the modelling of the loss-given-loss (LGL) and then a more complicated example of this phenomenon in transaction monitoring in the context of detecting financial crime.

Read the entire interview under Download.