Risky business: why existing models don’t work

“Prediction,” said the Danish physicist, Niels Bohr, “is very difficult, especially if it’s about the future.” Within these pages, we’ve discussed the impossibility of specifically predicting the pandemic, along with the very real possibility of being prepared for the outcomes associated with it. Ultimately, it comes down to the data, and the models they feed. 

For example, if you doggedly stuck to econometric models, you wouldn’t have been able to have a pandemic-ready loss forecast a year before it happened.  Yet if you considered the full range of dynamically-generated scenarios, then you could have. Moreover, you would have also seen that a pandemic-like outcome could have been caused by a combination of other shocks; you would have seen that equity markets could have dropped even more than they did, and not necessarily recover.

Most criticism of models focuses on the econometric approach of “fitting the data” and how you cannot know in advance which periods of data to fit. Clearly, pre-COVID data is not useful during the pandemic, but could 2020 data be useful at a later date? Should COVID, as an event, be included in future models as a dummy regression variable given that an exact repetition will most likely never occur?

In reality, the coronavirus pandemic has revealed these questions to be irrelevant, because an econometric approach is not fit for forecasting risks. Instead of trying to explain and fit all past events into traditional regression models, we should learn from them what the potential impact range could be. 

For example, COVID has shown that the unemployment rate can rise fourfold within a month. So the proper question would be: could this happen in the absence of COVID, perhaps for a very different reason? The answer is: yes, of course it can – just as the dotcom bust and financial crisis taught us credit spreads can jump, equity markets can crash, and liquidity can dry up. 

So instead of forcing all risk models to fit these past crises – or the current one – we need to learn how to calibrate the impact of any and every crisis on risk drivers. Instead of replacing useless models with subjective judgement, risk managers should generate the full range of dynamic scenarios with various combinations of potential shocks. Instead of trying to predict exactly what’s going to happen, financial institutions should consider the real probabilities of potential outcomes that incorporate tail-risks, and adjust their risk appetite accordingly.

Once this becomes the norm, the risk management community wouldn’t have to guess when to switch out of the models that were built in a hurry to fit 2020 data. A transparent, multiple-scenario approach also makes the model governance process faster and easier, without giving up the rigor of model validation.