This paper serves as one of the first studies that estimate the value at risk (VaR) via a Markov-switching (MS) model. Specifically, we use a two-regime MS specification, a MS setting with two sets of regime mean and regime variance, on TAIEX as well as Taiwan's major industrial group stock index returns. We demonstrate that MS effectively correct non-normality problems and outshine both GARCH and the mixing normal models, with the former (latter) alternative being subject to over- (under-estimating) the persistence of stock return volatility (hereafter volatility). As for estimating the 5% VaR, MS appears to be equally effective as Bayesian mixing normal and GARCH. In contrast, MS significantly outperforms the two nonlinear alternatives for estimating VaR with 1% or 2.5% tail probabilities. Furthermore, as for the window of learning period on rare events, we find that one need to go much farther back to effectively depict the left as opposed to the right tail.