Too much faith in black boxes? Although LTCM's problems have received most of the attention, the loss in equity value of large banks over the last two weeks of August 1998 make its losses look less dramatic. On August 21, when LTCM lost half a billion dollars, Citicorp's equity fell by more than Dollars 2bn. From August 26 to September 4, the market value of the equity of Bankers Trust, Chase Manhattan, Citicorp and J.P. Morgan fell by a combined Dollars 43bn or 29 per cent. Clearly, the risk management systems of LTCM were not the only ones that had problems on these days. It is, however, important to be clear about where the systems went wrong. Systems did not fail to predict that large losses were possible. A daily VaR that measures the maximum loss at the confidence level of 99 per cent will be exceeded one day out of 100 on average if it is estimated accurately. Risk managers knew that there was some possibility of devaluation in Russia or of capital outflows in Brazil.

However, the systems did not anticipate and were not prepared for the vicious circle of losses that developed as positions could not be unwound without creating further losses, which themselves forced further unwinding of positions. This vicious circle had two critical consequences. First, it made one-day VaR measures irrelevant because these measures are based on the presumption that positions can be undone rapidly at low cost. Second, as losses forced the unwinding of positions, the crisis spread to unrelated positions, thereby making the returns of unrelated positions highly correlated.

Any risk management system relies on forecasts of the distribution of returns of the portfolio or institution whose risk is managed. In normal times, forecasting the distribution of returns is much easier the world just keeps repeating itself with no dramatic surprises. Crisis periods are different the past becomes much less useful in forecasting the future, volatility often grows dramatically and correlations become much closer to one.

All this happened in August 1998. The previously uncorrelated positions of LTCM suddenly became highly correlated. By the time models based on the past adjust to these changes, it is no longer possible to reduce risk cheaply. The fundamental assumption of many risk management models is that bad draws occur randomly the fact that there was a large loss today means nothing about the probability of a large loss tomorrow. Unfortunately, in August 1998, events that risk models said had an infinitesimal probability of happening were happening several times a week.