None of the above. You should use 20-day returns calculated off the rolling 20-day window. The produced data will be correlated this way because neighboring windows will overlap by 18 days. This is not a problem though. Even for correlated observations barrier frequencies are unbiased estimates of the true barrier probabilities.
Note that the number of produced 20-day returns will be the number of 1-day returns minus 19. Almost the same sample size. Our discussion is based on the assumption that this sample size will be sufficient. If not, you will have to use parametric methods to extrapolate data into the tails. For example, to estimate the tails of a single variable you can use extreme value theory
. To quickly postulate how various pieces move together, you can use copulas
with non-zero tail dependence. I am not saying that you should model various constituents of your portfolio separately and throw them into a copula. I am saying that you may notice that your portfolio is sensitive to, say, oil and there has not been enough variation in oil over your trading history. Then you model oil separately and enforce some dependency structure onto oil and your portfolio.