Why we need good models

We live in an ever-changing world frequently rocked by unexpected events. When this happens, models can fail and forecasts go awry, as they did during the 2008 financial crash.

Once again, the accuracy of models is being tested as the world battles with the Covid-19 pandemic. This new crisis reiterates the need for robust models that can account for sudden changes and reveal unexpected features.

Two new papers “Robust Discovery of Regression Models” and “The Value of Robust Statistical Forecasts in the Covid-19 Pandemic” by Jennifer Castle, Jurgen Doornik, and David Hendry focus on what makes a good model, why forecasting goes wrong, and the importance of robust forecasting in a pandemic.

In the first paper, the authors focus on what they term “model discovery” where a general-to-specific approach is used to ensure everything that matters is captured. It means large numbers of variables are included, but those deemed insignificant can be discarded and the model made more specific. 

If the model detects a shift – something within the data which is unanticipated – it will account for the change and update its course. This, along with the detection of outliers and other variables will permit the modeller to allow an explanation to be based off the data instead of their own, possibly biased, assumptions.

In their second paper, statistically based models are compared to epidemiological models to understand how accurate both approaches are at forecasting the impact of the pandemic.

The authors conclude both types of models have a role to play.

While the epidemiological models are able to assess policy implications or analyse different scenarios in order to inform policy, they can be slow to adapt to sudden shifts like lockdowns. If not accounted for, these shifts will disrupt the model estimates, and the resulting forecasts will be wrong. Robust forecasts adapt quickly; however, they fall short on creating future policy scenarios.

On the other hand, as confirmed by their first paper, statistically based econometric models can account for past shifts and are therefore able to address the economic implications of the pandemic, such as unemployment levels during the furlough scheme.

While both papers focus on different methods of econometric modelling to produce useful forecasts, their ultimate goal is the same. What makes a model robust and how it can be used to better understand the world around us is arguably now more important than ever.

  Angela Wenham, Office and Communications Manager, Climate Econometrics

  David Hendry, Co-Director, Climate Econometrics 

  Jurgen Doornik, Researcher, Climate Econometrics 

  Jennifer Castle, Associate, Climate Econometrics