Take the last 500 trading days, compute what the current book would have returned on each, sort the results, and read off the 5th percentile for a 95% one-day figure. No distributional assumption is required - the shape of the data is whatever it was.
The appeal is honesty about shape: fat tails and skew that actually occurred in the window are preserved. The weakness is that the window is the entire universe of possible events. A 500-day sample ending in a calm period contains no crash, so the model reports that crashes are impossible. Extend the window and you mix in regimes that no longer apply.
Use it with a deliberately long window and check the result against a short one. If the two disagree by a factor of three, the number you are reporting is a statement about the sample period, not about your portfolio.
Related: value-at-risk, parametric-var, monte-carlo-var, stress-testing