In case bettors aren't aware, ERSSTv5 has been discontinued as of this month.

(see https://psl.noaa.gov/data/notices/)

So, next month necessarily I believe GISTEMP will have to shift to ERSSTv6 (some relatively small code changes).

I had to get the fortran version of the GISTEMP ERSST -> SBBX code working finally (using NCEI's monthly .nc files) rather than using my old hacky code to generate from the existing SBBX as a reference (its much faster this way). Also had to modify my fork of GISTEMP to update to use these changes (haven't pushed any of these changes to my repo since my tree is very messy atm).

For reference, here are the parquets of the GISTEMP global means (i.e. from result/mixedGLB.Ts.ERSSTV5.GHCN.CL.PA.csv), with runs both using ghcnm.v4.0.1.20260707 (which was used for the June GISTEMP run), and one with the ERSSTv5 and ERSSTv6.

https://jrpdata.free.nf/gistemp/loti_ersstv5.parquet

https://jrpdata.free.nf/gistemp/loti_ersstv6.parquet

Anyway here is what I have so far for July 2026 using this new ERSSTv6:

(edit: forgot to update pipeline for delta method (recalculating the gistemp boxes; i.e. switching to ERSSTv6 dropped the 1.2502 C to 1.2244 for that method !! this also dropped the weights as well for it slightly)

weights: old, delta2 method, loti analog3 method: 0.2281, 0.4684, 0.3034

Point estimate (old, delta2 method, loti3 method) adjusted by prediction error mean: 1.2062, 1.2244 1.1946

Point estimate (mix), pre-adjusted by member prediction error mean: 1.2112

Probs (mix) with super ens var

Std. dev.: 0.0388

Bin Probability (%)

<1.095 0.1

1.095-1.145 4.3

1.145-1.195 29.4

1.195-1.245 47.0

1.245-1.295 17.6

>1.295 1.5

@parhizj my July prediction has gone up by 0.006, all other previous data dropped by about 0.015

Thanks for sharing.

I still haven't finished reading the ERSSTv6 paper (only skimmed it) but my impression from previous work I have ongoing that I could visually analyze is that this shift is more likely contributions from the GISTEMP clim. period warming up a bit relative to afterwards. There's also some more HF signal now that they've dropped the 3-month persistence.

ERSSTv6 for last month you can see there is more detail, especially in the southern ocean (although you could interpret some of this HF as noise):

compared to the same run of ERSSTv5 for last month:

@zenarxy I've been looking at ersst to improve a different improvement to my pipeline (after the t2m superensemble) and I'm working on the ocean part now finally with OISST to improve over a simple climatology model (i.e. the sst portion's variables of the model would like: ERSST climatology + (mixed preliminary->final) OISST + super ensemble t2m.

The land analysis (a tiny portion of it below) on a separate factorial decomposition analysis I've been doing recently (only half finished since I haven't modeled forecasting the weights for the leading ghcnm month) using rolling corrections shows the land data variance from the specially modified ERA5 t2m is now finally lower than the ocean data (even with ERSSTv5, but along with the ERSSTv6 the variance also been reduced even further relative the t2m data has given me motivation for stopping and improving the ocean data further (it reduced quite a bit!)).

Two questions (if you have the time):

1) Have you done the analysis yet for the revised ERSSTv6 over v5, and if so, have you observed any improvement in reducing the hindcast variance (or MSE) for your SST forecast models? (I'd have guessed it improved with more HF signal in ERSSTv6).

2) Thanks to some sleuthing on the open directories for SST data, I did find the preliminary OISST data for it going back to Jan 1 2026 (the NCEI cmb folder has it) so I have enough statistical data to come up with a preliminary->final revision model as well.

I wonder if you've been doing so for a longer preliminary OISST dataset you've been accumulating?

I've come across some information relevant for you for that in such a case (i.e. STAR/NESDIS SST quality monitor):

(See NASA JPL SST L4 MUR Data Degradation - Earthdata Forum for the information on the event)

I'd imagine it would be good to mask/weight such events for the daily preliminary->final revision model's statistics.

~

Some plots I think you might find interesting from the analysis I've done so far:

For the specially weighted ERA5 land model (using ERSSTv6):

(compare these two below and to the GISTEMP one from ERSSTv6 in the above post) -- the weighting successfully works (at least in hindcast) much better at reproducing the GISTEMP methodology's spatial structure).

weighted (and masked) ERA5 land + ERSSTv6 for June 2026:

regular (interpolated) land ERA5 + ERSSTv6 for June 2026:

The warm and cold anomalies end up much better approximated to GISTEMP's land anomalies (with only purely weighting the ERA5 data). (lowering the stddev by something like half in earlymodern /modern epochs)

For the factorial decomposition I rerun my own reproduced version of step5 (completely accurate) that allows me to do the substitutions and various other data gathering.

For ERSSTv6 and the weighted ERA5 t2m land version (decomposition analysis over 1940-) with deltas of the global (zone 16, annzon substituion means you find in LOTI)

Delta total (all ERA5 t2m subboxes replacing GISTEMP (1940-), put through an equivalent GISTEMP step5);

Delta land (ERA5 t2m land subboxes replace GISTEMP land subboxes);

Delta ocean (ERA5 t2m ocean subboxes replacing the ERSSTv5 subboxes).

Rolling 10 yr analysis of the residuals for the substitions (epochs based on satellite coverage to relate subjectively what I think the ERA5 roughly, structurally changed):

(prior to 1957 is especially poor for disagreement between the datasets)

After putting them through a simple (monthly) model with a rolling correction for the residuals, a (rolling 5 year mean of the) variance is decent (with the train MSE for delta land comparable to my best t2m delta analog model's train MSE; probably can be improved better with a ridge regression instead, etc). the delta total is still signicantly worse though than the analog models I use. (Hindcast residual analysis below based on the decomposition)

(the total variance is decent in the modern epoch and the latter half of the early modern epoch). For instance, delta land shows how much the variance would improve if the ocean was perfect (it would be comparable to the delta analog model I use), but in reality it would be somewhere in between).

~

For reference Gistemp 1880<-> GIstemp with a 1940- cutoff also has marginal disagreement that needs to be modeled but can be improved if you take care to fix the subbox types (there are 39 total subboxes, of which up to ~30 can change in a single month from ocean to land if you don't fix the subbox type (i.e. fix it as ocean) -- this has the effect of effectively masking those subboxes for those sea ice minimum months (where they get masked due not having 20 total years (240 months) of data) but it is absolutely worth it since the land-ocean variance introduces greater disagreement of course i..e fixing it takes the disagreement stddev from ~ 0.007 C to 0.003 C which is worth it even correcting for later on since it should be reducable even further with statistical methods).

(this was the same for ESSTv5 as well for below; but below is ERSSTv6)

subbox histograms (sorted by increasing total valid months), showing the sea ice months and the # total valid subboxes in the right annotated column for the different cutoffs; Bottom samples are beginning, middle, end of the rest of the sort. This is a tiny bit hard to interpret since you have to pay attention to the vertical scale of the histogram

Time series of the above (pink == invalid / masked periods):

(Edit fixed title)

NASA JPL SST L4 MUR Data Degradation - Earthdata Forum

@parhizj I haven't done a full controlled ERSSTv6/v5 analysis specifically for the SST forecast component. Most of what I changed was the input handling, post-processing, and some small adjustments to the OISST->ERSST bias correction. These produced a slight improvement, but nothing especially noticeable, which is probably unsurprising given how little global SST changes on a day to day

However... I have noticed that the daily LOTI output has been slightly more stable since the changes, although it has only been four days, so it is likel ytoo early to draw any conclusions

Pre-changes:
13/07 – 1.213
14/07 – 1.230
15/07 – 1.235
16/07 – 1.218

Post-changes:
17/07 – 1.223
18/07 – 1.221
19/07 – 1.220
20/07 – 1.216

For the preliminary OISST data, I haven't been accumulating a separate archive. My script replaces the preliminary files as soon as the final versions become available. I didn't previously see much benefit in retaining both, given the low daily global SST variance and the amount of post-processing and bias correction already applied

Your point about known degradation events is a good one, I just don't know if it's worth the time and effort put into building a model that takes that data into consideration given how little of a change it would contribute to the LOTI after all the post-processing

@zenarxy Thanks for writing that up, I appreciate it!

Bettors may be interested in this AUGUST vs JULY market for ERA5 below (default probability set to the prior):

Some of the EC46 runs have been shooting up significant beyond climo, but today's was extreme. Normally climo is superior except we are also well on our way to a very strong / potentially historical El Nino....

Paris getting some more warmth second week of July (July 10-11)

mean +7.5 C anomaly 1991-2020 (absolute 26.4 C) in my super ens gridded data... (max 99F, 37.2C from on accuweather).

France and Antarctic are warmer for first half of month in today's run:

(from 06Z) LOTI: 1.192 +- 0.094 C (dropped a bit with the extra day, and end of medium range dropping a bit (European heat wave a bit milder than yesterday at end of week 2)

(using old method, delta analog3, and loti3 methods (MTD doy models on super ensemble+statistical_1harm + Prophet for beyond medium range)

Ensembles dropped again today, and it appears the change is mainly from Russia, though Antarctica is cooler as well compared to yesterday.

EPS/00Z shows greater troughing over northern Russia in second week of July (more meridional jet stream), showing it persisting for most of the week, while ECM diagnostically shows it as a cutoff low.

Yesterday's run and today's superensemble run for July 11:

ECM, EPS (via tropicaltidbits) for July 10, 00Z: