Useful when the analytic sampling distribution is unknown or the normality assumption is doubtful - which covers most realistic statistics.
The time-series trap. Standard bootstrapping resamples observations independently, destroying autocorrelation and volatility clustering. Use a block bootstrap, resampling contiguous blocks, to preserve dependence structure.
The limitation. Bootstrapping cannot create information the sample does not contain. For tail quantiles it is unreliable, because resampling only reproduces extremes you already observed - it can never generate a worse event than the worst in your data.