GARCH models capture positive serial correlation in volatility. Sometimes events occur "out of the blue", causing volatility that a GARCH model cannot be expected to predict. One example is the S&P 500 falling 4.37% on December 8, 1941, the day after Pearl Harbor. Another is the U.S. market falling about 10% at the reopening 1 week after the 9/11 attacks.
How should the GARCH estimation procedure be modified, if at all, in the presence of exogenous events? Just remove the "shock" days from the likelihood function? Doing so is plausible but could bias the unconditional variance downward, since such shocks will occur in the future.