Friday, January 23, 2009

Swift sudden collapse

The financial crisis (for instance in Iceland) has sometimes been described as "swift" or "sudden". This is also the case with the U.S. While the collapse itself can be seen as swift or sudden does it also mean as some commentators have indicated - that it was "unexpected"? Swiftness or suddeness does not mean unexpected. A series of missteps or wrong turns can accumulate and the economy can collapse suddenly.

The story in Fortune on Iceland indicates that this may well be the story. As far back as 2006:
In April 2006 the rating agency Fitch abruptly downgraded its outlook for Iceland, citing concerns about the banks. Investors panicked, and the currency and the stock market both plunged 25% in a matter of days.

Compounded by what the article characterizes as "personality issues":
... says a senior official at the ECB in Frankfurt, "if you want to put in place a swap agreement with us, you call [ECB president Jean-Claude] Trichet and make an appointment to come to see him. Then you fly to Frankfurt to discuss your request and explain how it is supposed to work. You need to win him over. Oddsson didn't even try."

Those who say that the subprime crisis could not be predicted or was unforseeable are incorrect since various indicators have been showing that the economy was under some stress for some time. (Roubini has warned of the impending crisis, for instance.) Of course, the timing of the collapse if that is what we're after is unpredictable.

In many economic models such as early warning models, crisis dates are modelled as binary variables (0 if the country is not in crisis at that quarter and 1 otherwise). Alternatively they are modelled as regime changes when the economy switches from calm to crisis. In these types of models the variable of interest is the predicted probability of a crisis. How well the model performs is whether the predicted probability (phat) can actually predict the crisis states. In some papers the plots of phat shows sudden changes which is actually built into the model by assumption. In this sense then the collapse is "unpredicted". In early warning models, phat can either perform well or not depending on the covariates but in general I think that both approaches are not very informative if we focus on phat.

The underlying covariates in early warning models tell us something about the economy - e.g. stress. Construct an indicator (e.g. using factor analysis possibly) and use that as a predictor of a crisis. What we might possibly want is the answer to the following question: "If an economy is under stress, how much stress can it withstand before the economy falls into crisis?" In this case we are looking for some cutoff value of the stress indicator (just as we are looking for a cutoff value of phat in early warning models) and is more informative than a phat.

Unfortunately, the value of a stress indicator are only slightly more informative since it really doesn't tell us much about what is causing the stress. A decomposition of the index would be required. However, this approach may lead to better information on how an economy collapses. If an economy collapses without and change in the indicator then something is missing from the indicator. Alternatively, perhaps the factor weights are changing over time. All this gets into the black box of a crisis in a way that early warning models do not.

Alternatively, one can argue that phat from early warning models already constitute such a stress indicator. The logistic regression approach using phat does not allow the analyst to explore with changing weights of factors - although in fairness neither does factor analysis.

The idea here is that whether the economy is under stress depends on a bunch of factors. Over time as regulators focus on one aspect it's importance as a possible contributory cause decreases but others may rise over time and this is the aspect that is not well captured by current models.

No comments: