DISCUSSION: seasonal timescale: persistence or climatology

Problem description

Does a persistence forecast for leads at a monthly spacing make sense? Lets assume your hindcast is initialized on 1st Jan. so lead 6 would be the month of Jul. Who would want to compare a December value with a July value? makes no sense to me.

Or would the point be removing the annual cycle before the analysis and then checking whether the persistence sustained at lead 6?

probably a climatology forecast is much for useful. how? taking the control/uninitialized run and create monthly climatology hist.groupby('time.month') and prescribe to the following years.

[judithberner]

I think the answer is highly state- and regionally dependent. For metrics for which SST or ocean heat content is relevant (e.g. Western United states precip, NAO -related metrics, e.g. wind gusts over Northern Germany, …), you can memory up to three months easily (that is statistically significantly, not necessarily socially relevantly). Same for S2S, if you stratify by the MJO state or Sudden Stratospheric warmings, you might get state-dependent predictability beyond the typical (i.e. average) predictability horizon. It’s not clear to me why- in general- persistence should be worse than climatology.

[aaronspring]

Sorry I meant to close the issue because there is nothing we should change in climpred about persistence or climatology (once #566 is merged).

I did not mean to say that persistence is never good. I raised the issue to discuss where persistence takes a constant values (it does) or adapts to seasons (that’s climatology).