That depends on what VaR model was used. Most banks' models performed poorly which is not surprising given the popular use of historical simulation. While historical simulation provides stable VaR numbers, it has weak forecasting power and is entirely inappropriate during regime shifts (see Finger's "How Historical Simulation Made Me Lazy"). Responsive volatility estimators, such as EWMA and ARCH type models performed much better, and indeed provided early warning signals months before the full subprime meltdown in July 2007.
Download file Chart 1 illustrates the RM 2006 VaR forecast vs. realized log spread changes on the 2006-1 AAA ABX tranche. The first warning was a 300% vol increase from Dec 12 to 21 '06. The second was a 12 standard deviation / 350% vol spike on Feb 23 2007 (this was the day after HSBC announced that it fired the head of its US mortgage lending business as losses reached $10.5bn... the alarm bells were clearly ringing). Even though spreads almost tripled on that day from 11 to 30.8 bps, as seen in Download file Chart 2, it was not too late to hedge. In fact, spreads proceeded to tighten to a low of 14.08 bps on June 25 '07 before widening significantly in three major bear waves. In other words, risk managers had between two to six months lead time to execute hedges.
The main lessons are (1) pay attention to early warning signals, and (2) not all VaR models are created equally. To be useful, VaR should be dynamic and responsive to market conditions. After all, risk is dynamic. And while no model is perfect some models are certainly more useful than others.