When the Model Doesn’t Fit

For the more mathematically inclined only, there’s an interesting guest post over at Econbrowser. The post, by Laurent Ferrara (EconomiX-CNRS, University Paris West) and and Valérie Mignon (EconomiX-CNRS, University Paris West and CEPII), considers models of unemployment. Here’s their conclusion:

We clearly see that both ECMs are able to reproduce the main movements in employment during and after the recession. However, there is a persistent gap between the observed employment (green line) and the conditional forecasts from both models (red and blue lines), meaning that the employment is currently well below what it should be according to the models. When comparing both ECM models, taking the non-linear business cycle into account through the non-linear ECM (red line) leads to reduction in the gap. But the contribution of the non-linear cycle to the employment is low (the difference between the blue and the red lines is around 1.2 % in 2010) and tends to diminish (the red line tends to the blue line).

This leads us to conclude that there has been indeed an effect of the Great Recession on the long-term employment. Specifically, on average, since the exit of the recession (2009q2), we get that the employment is 2.7 % below its potential level (according to the non-linear ECM), meaning that, from a structural point of view, around 3 millions of jobs have been lost after the recession. This interpretation is in line with the recent literature on this topic, as pointed out for example by Chen, Kannan, Loungani and Trehan (2011) or Stock and Watson (2012) that put forward various explanations.

As an empiricist I would draw a slightly different conclusion: although their models produce results that are vaguely similar to actual observed behavior they predict neither the degree to which employment has declined nor its lacklustre recovery. Said another way, their model is wrong. Or, at best, inadequate.

When the model doesn’t fit the behavior you’re trying to model, it’s time to change the model.

5 comments… add one
  • I think that is the point of the post, they went from a standard error-correction model (ECM) to a non-linear model to try and get a closer match to reality. In that they succeeded in that the new model does get a better match to reality. However, they failed in that the improvement is not sufficient. What the next step is I don’t know.

  • However, they failed in that the improvement is not sufficient. What the next step is I don’t know.

    Yeah, that’s my point. The approaches that are being used aren’t very close to reality. That suggests a problem with the approaches that are being used. IMO the problem is mostly the assumptions. That’s been my experience with models. You can’t get great results with so-so assumptions.

  • Which is why I still get somewhat annoyed when people say the stimulus worked and point to models that are not unlike the one in question.

    Don’t get me wrong, research like this is good….it is moving things forward, but a level headed analysis would still conclude that macro models are lacking….and these are the same models that many are using to conclude the stimulus worked.

  • When the model doesn’t fit, find another dresser.

  • It’s hip to be square.

Leave a Comment