Reorder raw simulation like scen in ExtremeValues #84
Conversation
|
Another approach to this could be to add tail_options in all bias adjustment techniques? This could be: |
|
!!!This was wrong, I keep it for reference, but don't pay too much attention!!! Some explorations. As I expected, working with non-synchronous sim/adjMBCn can lead to some problems mbcn is just mbcn without GEV adjustment mbcnOrdered reorders sim according to the reordering of mbcn, them performs the GEV adjustment. mcbnDisordered is the same without reordering sim.
It's nothing huge though |
|
!!!This was wrong, I keep it for reference, but don't pay too much attention!!! This might not make big differences in indicators at the end, but still, the difference is big enough IMO to justify changing algorithms
|
|
Actually I had reversed both datasets, so actually no reordering seems better in this case. I must have messed up the reordering implementation. With group="time", it seems that the MBCn has pretty much the same rank structure as the raw simulation for large values: However, for group="time.dayofyear, 31", the rank structure is more different, even for large values
|
|
Practically no difference between Disordered or Ordered for group="time" But we the difference for group=Grouper("time.dayofyear",31) |
|
That doesn't seem specific to MBCn though what I've just shown above. This is more about the fact that grouping with doy-31 has some effects on the rank structure. Consider the same plots as above with mbcn -> qdm
|
|
For reference, look at how the ranks are modified if we compare qdm-t and qdm to raw data (for tasmax here)
|
|
Interesting to see how much kind=* and kind=+ differ for the large precipitations. I think the justification of using was more about the variable being bounded by zero and not wanting to produce pr<0, but still, it has important effects for large values.
Oups, there should be cyan text saying: Max of QDM and QDM-GEV are the same points (for QDM-ordered) In any case, here is the main point: You can see that if you don't reorder the simulation, the largest value in QDM is not necessarily the same point as the largest value in QDM-GEV. I don't think that's something we want to do, we don't want to destroy the rank structure that was found in the MBCn (although, as underlined previously, the major factor of re-ranking was more about group="time" and group="time.dayofyear,31" rather than the reordering inherent to MBCn ) |
coxipi
changed the title
|
@RondeauG I just reordered I would still like to clear up whether the GEV was being performed on a day-of-year grouping in ClimateTools.jl |
Pull Request Checklist:
number) and pull request (:pull:number) has been added.What kind of change does this PR introduce?
reorder_simoption inExtremeValues.adjustextremes_adjustperforms a GEV adjustmentgev(ref,hist,sim) = adj_GEV, then it combines this withscen=adj_XYZcoming from a previous adjustment. The idea is simple to reordersimbefore thecomputation
sim_r = reordering(ref=scen, sim=sim), and use this instead:gev(ref,hist,sim_r) = adj_GEV_r. The idea is thatadj_GEV_rwill be more synchronous withscen=adj_XYZ. So the maximal value ofscenstays at the same place in time, but it get adjusted according to what is found from the GEV adjustment.Other information:
GEV is written all over the place, but as @RondeauG pointed out, we in fact use a Generalized Pareto. The approach is a bit different (peak over threshold) but my point on the usefulness of reordering
simstands. I will think more about it