As it is often presented, Extreme Value Theory (EVT) is applied to financial losses (and gains as well, but always focusing in the tails). By EVT, I especially refer to:
1) Mean excess plot to find an appropriate threshold 2) Generalized Pareto Distribution (GPD) to fit to the data in the tail, beyond the estimated threshold in 1).
My question is: does it make sense to apply steps 1) and 2) to daily (arithmetic) returns, and not to PnL?
I have also seen some cases where steps 1) and 2) are applied to the residuals of a fitted model, e.g. ARMA or GARCH.
Which approach is the best?