IK86, that's not even the question (which is where I part company from Yves' plaint). Large firms used direct historical sampling (more than the 2 years indicated in NYT) across thousands of data series to estimate VaR (so higher moments are baked in as well), forgetting that past performance is no guarantee ... not to mention the problems of classification for mixed and leveraged assets. For individual asset classes, there have been numerous attempts to generalize distributions e.g. for option pricing, and Taqqu & Samorodnitsky's Stable Non-Gaussian Random Processes lays the basis for stochastic analysis, but multivariate approaches remain lacking. Tail estimation is another matter (see e.g. Embrechts' Extreme Value Theory). Deficiencies in multivariate and marginal techniques are further compromised by sparse observational data and by asynchronous behaviors (GARCH notwithstanding). All of these alternatives were established by the late 90s, but none of the attendent problems overcome.
[clarification: it's multivariant estimation that remains lacking, not the formalism itself]