Stochastic projections are notoriously difficult to fully interpret because of the sheer number of variables, not to mention the fact that the phenomena we’re trying to model are tremendously complex. Yet this level of analysis can translate into very real capital benefits. For this reason, modelers are seeking ways to make the structure of models simpler, and allowing complex relationships to emerge rather than modeling them directly.
Often building a model of an observed behavior directly creates a programming challenge. For the modeler, it’s almost impossible to determine the dependency structure at the outset, and if inaccurate assumptions are made about causal relationships the model will fall flat. The recent economic crisis highlighted the point that factors do not connect the same way in all conditions.
Risk is inherently an emergent property. Yet the techniques currently used are fundamentally at odds with how behaviors are produced in reality, making the models extremely hard to build, hard to understand, and potentially misleading in their outputs. However, an alternative technique, known as “agent-based models,” is producing results by letting the interactions between factors take place naturally within the model. Therefore the final aggregate outcome is generated by the interaction of risk drivers, not characteristics correlated as risks.