Risk Professional - December 2009

MODELING

The US Housing Crash:

Model or Management Failure?

Executives who were quick to blame “faulty” risk models for their inability
to predict the real estate downturn need to take a closer look in the mirror.

any factors have contributed to the recent financial crisis. The surge and subsequent decline in US real estate prices is one of the most important ones. By itself, however, a housing crash is not enough to spark a banking crisis. If the estimated probability of a crash is large enough, financial institutions that are concerned about their future solvency should either reduce their exposures or set aside enough capital to weather a storm. Surely, many institutions failed to do this.

But who takes the blame? Some observers point to fundamental deficiencies in standard risk modeling that also affected the reliability of forecasts for the housing market — e.g., reliance on historical data runs the risk of missing extreme events that have not occurred in the past; distributional assumptions often do not include fat tails; and responsiveness to structural breaks is typically slow. The ubiquity of such problems can lead to a distrust in the reliability of risk models. An extreme example of such a negative perspective is given by Nassim N. Taleb in his book The Black Swan. 1

To assess whether risk models really failed
to warn of the housing crash, one can put one-
self in the shoes of a risk manager at the height
of the housing boom (2005) and do what many
banks claimed they did as part of their risk

M

management: generate stress scenarios. The usual practice is to derive stress scenarios based on a time-series analysis of historical data, and this means that one could consider using standard autoregressive integrated moving average (Arima) models (probably with time-varying volatility or other additional features).

Sometimes, model choice can be decisive for the results. For the problem studied here, it turns out that it does not matter much whether one uses a very simple model (such as an autoregressive process of order one) or a more complex one that optimizes the lag structure and adds time-varying volatility. 2

In 2005, a risk manager’s first pick of an index that tracks the housing market might have been the Case-Shiller national home price index (or another index from the Case-Shiller family). Applying time-series analysis to this index would have produced very mild stress scenarios. Even when considering a 0.1% worst-case scenario

(i.e., only 0.1% of all scenarios are worse), house prices would not have declined; rather, they would have stayed more or less flat over a horizon of three years

— from 2005 to 2008.

A vigilant risk manager, however, should have been very hesitant to trust an analysis based on the Case-Shiller index. The index is only available from 1987. Given that slow-moving nature of house prices, this is a small timespan.