The IMF's admission (p41 of this pdf) that "the multipliers used in generating growth forecasts have been systematically too low since the start of the Great Recession" is being interpreted as a victory for Keynesianism, because it means that fiscal austerity is doing more damage than the IMF originally estimated.

However, I suspect it is also a vindication for thinking of economics as a bunch of mechanisms rather than as formal models.

To see what I mean, let's concede that the pre-recession evidence pointed to lowish or even negative (pdf) fiscal mulipliers. Mechanism-based thinking would ask: what mechanisms could generate such a result? Answers would include:

- Fiscal tightening can be offset by lower interest rates.

- Cuts in the government wage bill reduce workers' bargaining power and hence expected profit margins, which can encourage capital spending.

- Lower government spending leads to anticipations of lower taxes, which might encourage increase household and corporate borrowing and spending.

This automatically draws our attention to reasons to doubt that low mulipliers would exist now. Monetary policy is less effective now we are at the zero bound; and an inability or reluctance to borrow limits how far private spending can rise to fill the space left by public spending.

We'd therefore expect to see higher multipliers now, not because Keynesians are always right and Austerians are always wrong, but because the mechanisms generating low multipliers happen not to be powerful here and now.

My approach - actually Andrew Glyn's and Jon Elster's: I steal from the best - contrasts with the idea that economics should be like physics. It is sceptical of whether there are deep parameters that yield eternal true predictions. Granted, there might be a few stylized facts and rules of thumb - my favourite being cubic power laws - but important as these are they are often not sufficient basis for policy or investment action.

If economists can learn anything from physicists, it is not the latter's pursuit of simple models that try to explain everything, but rather the mountain of unglamorous gruntwork they put into gathering data.