🏅 Top traders

The /pub/gistemp directory has been updated (using ghcnm labeled 20260407),

      ghcnm.tavg.v4.0.1.20260407.qcf.dat                    2026-04-08 01:40  165M  
      ghcnm.tavg.v4.0.1.20260407.qcu.dat                    2026-04-08 01:40  165M 

implying a value of 1.28 C (that they did the run yesterday).

@ChristopherRandles data is out: 1.28 official. This can resolve now

1.28 🎉☺️

What time does ERSST come out?

@zenarxy now.

what an odd distribution, wtf?

@ScottSupak I'm out of mana

@ScottSupak @zenarxy is very confident it will be biased high. I have less confidence overall

Final ERA5 point is in.

Same old ERA5->GISTEMP model suggests 1.221 C after own past forecast errors (1.210 C unadjusted).

I did do some experiments yesterday (some handcoded and some with Claude) to see if there was anything useful to learn from differences in ocean sst,skt,t2m from the ERA5 data (with an eye to the ERA5+ECMWF model as a proxy) as it relates to the GISTEMP ERSSTv5 data. The simplest way I thought of doing this was by interpolating the ERA5 grids (masked per the sst field) to the same ERSSTv5 grid and running GISTEMP using them instead of ERSSTv5; the idea would be this would give a better estimate of the actual potential improvements.

Generally the results were poor in trying to improve upon a baseline prediction (a method somewhat similar to what I use above for ERA5->GISTEMP) using purely ERA5 data, though the % improvement is extremely sensitive to the cutoff for the data (ie in comparing 1951- onwards, 1961-, 1971-, 1981- onwards) for more than one apparent reason.

The largest improvement is as expected from capturing the differences between the two datasets (ERSSTv5 vs the ERA5 data), so any future work should still focus on trying to capture any remaining spatial-temporal biases between the two datasets rather than between ocean/air differences within datasets.

That being said the handcoded separate notebook seems to confirm the below analysis, but the results seem similar. For March 2026 at least, using the best proxy I could produce for ERA5 while using a recent mean for the ersstv5-era5_t2m feature, and using a variety of cutoffs, a marginally cooler temperature relative to the baseline is indicated for a majority of the cutoff years, but generally in the range of 1.20-1.23, with most suggesting 1.21.

~

From the Claude based analysis:

(Year below is the cutoff year)

some sample data:

cutoff of 1951-

Month 1: ersstv5 - era5_t2m → Std Δ=+16.7%, Bias=+0.00001

Month 2: ersstv5 - era5_t2m → Std Δ=+12.7%, Bias=+0.00026

Month 3: ersstv5 - era5_t2m → Std Δ=+11.0%, Bias=+0.00009

Month 4: ersstv5 - era5_t2m → Std Δ=+16.8%, Bias=+0.00043

Month 5: ersstv5 - era5_t2m → Std Δ=+33.7%, Bias=+0.00011

Month 6: ersstv5 - era5_t2m → Std Δ=+29.5%, Bias=+0.00018

Month 7: ersstv5 - era5_t2m → Std Δ=+12.2%, Bias=+0.00029

Month 8: ersstv5 - era5_t2m → Std Δ=+11.1%, Bias=+0.00047

Month 9: ersstv5 - era5_t2m → Std Δ=+21.1%, Bias=+0.00000

Month 10: ersstv5 - era5_t2m → Std Δ=+21.6%, Bias=-0.00001

Month 11: ersstv5 - era5_t2m → Std Δ=+15.3%, Bias=-0.00006

Month 12: ersstv5 - era5_t2m → Std Δ=+19.9%, Bias=+0.00032

Month 3 - Method 1 Results (Minimum Variance Unbiased Estimator):

Difference Baseline_Std Method1_Std Std_Improvement_% Method1_Bias Is_Unbiased

ersstv5 - era5_sst 0.051559 0.048599 5.740658 0.000176 True

ersstv5 - era5_skt 0.051559 0.048012 6.880009 0.000185 True

ersstv5 - era5_t2m 0.051559 0.045907 10.962126 0.000087 True

era5_sst - era5_skt 0.051559 0.051181 0.733480 -0.000128 True

era5_sst - era5_t2m 0.051559 0.052015 -0.883356 -0.000093 True

era5_skt - era5_t2m 0.051559 0.052074 -0.997671 -0.000099 True

Month 6 - Method 1 Results (Minimum Variance Unbiased Estimator):

Difference Baseline_Std Method1_Std Std_Improvement_% Method1_Bias Is_Unbiased

ersstv5 - era5_sst 0.059448 0.052502 11.684761 -0.000020 True

ersstv5 - era5_skt 0.059448 0.050739 14.650156 0.000042 True

ersstv5 - era5_t2m 0.059448 0.041924 29.479058 0.000180 True

era5_sst - era5_skt 0.059448 0.053983 9.194186 -0.000025 True

era5_sst - era5_t2m 0.059448 0.058851 1.004960 -0.000116 True

era5_skt - era5_t2m 0.059448 0.059297 0.254855 -0.000101 True

cutoff of 1981-

Month 1: ersstv5 - era5_t2m → Std Δ=+10.4%, Bias=-0.00008

Month 2: ersstv5 - era5_t2m → Std Δ=+25.3%, Bias=-0.00012

Month 3: ersstv5 - era5_t2m → Std Δ=+23.2%, Bias=-0.00021

Month 4: ersstv5 - era5_t2m → Std Δ=+25.8%, Bias=-0.00027

Month 5: ersstv5 - era5_t2m → Std Δ=+33.4%, Bias=-0.00022

Month 6: ersstv5 - era5_t2m → Std Δ=+14.9%, Bias=-0.00014

Month 7: era5_sst - era5_skt → Std Δ=+4.4%, Bias=+0.00008

Month 8: ersstv5 - era5_t2m → Std Δ=+9.8%, Bias=+0.00019

Month 9: ersstv5 - era5_t2m → Std Δ=+24.6%, Bias=-0.00040

Month 10: ersstv5 - era5_t2m → Std Δ=+14.1%, Bias=-0.00008

Month 11: ersstv5 - era5_t2m → Std Δ=+11.4%, Bias=-0.00016

Month 12: ersstv5 - era5_t2m → Std Δ=+20.1%, Bias=-0.00039

Month 3 - Method 1 Results (Minimum Variance Unbiased Estimator):

Difference Baseline_Std Method1_Std Std_Improvement_% Method1_Bias Is_Unbiased

ersstv5 - era5_sst 0.04038 0.038648 4.289738 -0.000236 True

ersstv5 - era5_skt 0.04038 0.037926 6.077127 -0.000254 True

ersstv5 - era5_t2m 0.04038 0.030996 23.240164 -0.000209 True

era5_sst - era5_skt 0.04038 0.039001 3.415629 0.000061 True

era5_sst - era5_t2m 0.04038 0.038452 4.776187 -0.000087 True

era5_skt - era5_t2m 0.04038 0.038623 4.351906 -0.000121 True

Month 5 - Method 1 Results (Minimum Variance Unbiased Estimator):

Difference Baseline_Std Method1_Std Std_Improvement_% Method1_Bias Is_Unbiased

ersstv5 - era5_sst 0.048308 0.041973 13.113875 -0.000123 True

ersstv5 - era5_skt 0.048308 0.040123 16.943593 -0.000188 True

ersstv5 - era5_t2m 0.048308 0.032153 33.442118 -0.000223 True

era5_sst - era5_skt 0.048308 0.040373 16.426160 -0.000367 True

era5_sst - era5_t2m 0.048308 0.046826 3.067078 -0.000099 True

era5_skt - era5_t2m 0.048308 0.047664 1.333984 -0.000021 True

the very preliminary ghcnm data (yielding a gistemp loti of 135.5) is far from complete (and is currently biased warm with the data present) but with the spatial anomaly data I do have its speculatively worth looking at how much it could change...

Some images from today...

from ghcnm (0402), Anomaly matches what I expected, and we got data from Russia that was missing last month (the anomalies in Russia match up with what is expected from ERA5's t2m)

Best estimate of ERA5 t2m for month (filled in by super ensemble for last few days)

Subboxes from a run from today (notable is that GISTEMP's compositor appears to have notably weakened the cold anomaly in Russia (by eye looks warmer by something like 3-4 degrees, as the actual anomaly in the ghcnm shows stations with -7 to -8 C anomalies relative to 1991-2020 baseline); I don't place much luck in it getting colder with futur data if the end of the month data is complete for some stations since the forecasts from the end of the month were anomalously warm in that region.

(I've combined the ocean and land subboxes together now) (edit: the zorder for the ocean is too high -- will fix it tomorrow, so the boundaries show up properly)

I also plotted the (80) mixed boxes to get a better idea of how they are being filled in...

The two boxes over Canada (which are around 1C eyeballing it) will obviously become cooler (perhaps from +1 to about -6C) that should alone shift the anomaly down by about -0.175 C (-7 * 2 / 80), and am expecting the one over Greenland (which is around +3 to +4C) turning more neutral although its much harder to estimate that one since its mixed and on the coast (looks like its -1 to -3 C?, so halfway would be about 0 to +1C, so very roughly a reduction of about -2 C or so for that box, which would amount to a global reduction of -0.025C). So from those three boxes alone I do expect it to bring the global loti by about -0.19 C (to a global loti anomaly of 1.165 C, would seem to be a good lower bound).

Most of Africa is missing but from the sporadic cold that is present over ERA5 I don't expect the stations that do come in to actually bring down the anomaly given where they are located, actually it looks like it should warm across the Sahel a bit. A couple boxes are currently tilted cold across northern Africa (-0.7, -0.3C). Assuming the more western one that is around -0.3C flips to +0.7 C (+1C), that would be another +0.013 C change that I would expect though.

Similarly for South America, not expecting a huge change once its data comes in as its zone looks fairly warm already.

Australia is very hard for me to say but I don't know if it will get much cooler compared to how its being filled in from the ocean data, once its land data comes in if last month is any judge. If anything the 3 main boxes covering it might warm a bit despite the cold anomaly in the center on ERA5.

However the biggest problem seems the Antarctica data (we do have some but not everything yet), as ERA5 showing it very anomalously warm in certain regions. Comparing past stations from February at crunch time for instance it does look like it will warm up further, but I do not think it will be as nearly as bad as ERA5 suggests. If we get a lot of stations though this might be very wrong. Easily the bottom-left most box might warm up +1C (so +0.013 C).

So as far as I can guess, all of the above would indicate something along the lines of a change to around 1.19 C.

I know last month when I tried this analysis I wasn't good at it so this is another try. How I could go wrong: If I assume the cold anomaly over Canada is overdone in ERA5 in comparison with what we will get in GISTEMP (i.e. instead of a change of -7C change in those two boxes, we get a change of -3C or -4C), that would bring it up a further +0.075 to +0.1 C to the 1.265 to 1.29C). Numerous other ways I could go wrong in combination with the uncertainties I mentioned above with this, but I don't see a reason to change my base case with ERA5 as it stands.

Went back and reran gistemp for the previously mentioned model (subsituting the ocean data with ERA5 t2m,sst,skt, with the actual best prediction model relying on the loti differences from the ERSSTv5 vs loti ERA5 t2m runs with the ocean data substituted) and updated the predictions based on the claude model for a range of different cutoffs (1951- to 1991-).

This slightly raised the center point prediction from the ERA5 baseline now to 1.237 using the ERSSTv5 data and the initial LOTI data. I don't know how robust this will be given how preliminary the ghcnm is for this sole month (the boxes for SH over land will be ocean filled), so the increase may not be entirely justified. I'll only be able to tell and validate it until near/after the gistemp deadline.

Edit:

I do not know what zenarxy and polymarket data people are calculating on to get such a high center point prediction.

@zenarxy do you have an updated center point prediction?

~

The month over month anomaly trend would suggest from February you would expect the March anomaly to be slightly warmer but it appears it will be below the trend line.

Interesting to note is how negative the April trend is; I suppose winters are far less cold than they used to be, so March is warming up at a relatively faster rate comparable to April.

@parhizj Polymarket is just reacting to what gistemp is producing at the moment, which seems gullible but is not necessarily wrong. LOTI on release date is normally +-0.1°C to what GISTEMP produces a few days before release.

It will drop, but nowhere near 1.2. I think the only place you're going wrong is how aggressively you expect the subboxes to change.

My final midpoint is 1.3, my model produced a figure close to that consistently since 26th now.

I also have an improvement in development which I should be able to finish up today so next month I'm hoping to be more accurate.

@zenarxy Thanks. 1.3 seems high, but the current ghcnm is ridiculously high still.

Yeah I very well could be overestimating the drop from Canada/Greenland -- its hard to tell since we have the compounding effects of the ghcnm homogenization on the stations and the gistemp compositor that may both reduce the anomaly further.

updated ERA5:

~

with latest ghncm (0403) gistemp increased to 140.9776

More of Antarctica came in and all of the boxes over it raised substantially:

Its possible Antarctica now though that could change either way (up or down) though with more stations; I don't have a further prediction since its very hard to tell now. Typically the extreme heat anomaly is less well represented compared to ERA5 from what I've observed in the past given the sparseness of the observations.

South America's missing box got filled in, but I think the Brazil box is presently overdone by +1 degree C comparing to ERA5. So that should bring down the anomaly by about -0.0125C (or even more), but as I mentioned yesterday the Africa box looks like they would warm a comparable amount so that effect is canceled out.

That just leaves still the Canada/Greenland boxes as I mentioned yesterday, 1.40976 - 0.19 ~= 1.22 C, in line with the ERA5 prediction.

Vitally, still missing Canada/Greenland:

ghcnm.v4.0.1.20260404:

139.46

Updated ERA5:

~

Net expectations haven't changed (still waiting on Canada/Greenland to drop it by around -0.19 C):

We got a few more widely dispersed Antarctica stations so now I am more confident saying there won't be much further change there warm/cold. 🤞 (I've increased my bet slightly because of this)

We got all of Australia and NZ it looks like, but not some of the island stations surrounding it, and net doesn't look like much change.

The Brazil box still looks warm (some of Paraguay's stations now infilling it on the hot side).

The Peru box looks mostly appropriately infilled considering the ocean anomaly, but even if we get the stations in time I don't know if it it will cool (a few stations near the Brazil border yes, but its hard to tell from the ERA5 chart whether the coast will cool definitively), so not expecting much change there.

Even though we are still missing China, India, and large parts of southern Asia, still not anticipating too much net change buts its hard to calculate it visually (the box for Indonesia that is neutral might marginally warm while India's box should cool).

Looking at how the subboxes for SW Mexico are being infilled, it looks a bit too warm, but even if it cools it is only a tiny portion of the greater box for the Ocean, so maybe 0.5C cooler should be expected, so less than <0.01C cooling globally once Mexico comes in.

For Africa, my net expectations is I don't expect much change from Africa with further stations:

We got some Algerian stations so now the box for Africa that was supposed to warm is now in.

Mali might warm the western box a bit, but Senegal should balance it out.

Assuming we get stations from Kenya and not just Rwanda, the boxes for the central African stations for them should slightly cool.

If we get Mozambique in time and not just Madagascar and South Africa in time it looks like those boxes should only marginally warm, but if we don't get Mozambique it will warm since the Madagascar anomalies look warmer than how it is currently being filled, so possibly 0.5 C warmer across a couple boxes at least, which would be about 0.01 C warmer globally.

Still missing the crucial Canada/Greenland data...

ghcnm.v4.0.1.20260405:

137.306

~

New data from Iceland and Brazil had the largest effect in cooling, with the global temps moving by -0.02 C, fairly inline with predictions so far:

Even though the Greenland Box has not yet fully cooled down yet absent its data Iceland did bring it down a bit

The main Brazil box did drop on the order of ~ 1C (marginally countered by one of the NE Brazil stations, Macapa, warming up the box above very slightly)

SE Asia data slightly cooled multiple boxes (India, Thailand, Malaysia, etc.); this in particular removed quite a bit of smaller scale aggregate uncertainty from these boxes, so I increased my bet again based on this.

Largest unknowns left are still Canada+Greenland that will cool it substantially, with Africa and Peru being slightly possible factors for warming.

Edit:

Last updated ERA5, since all data is now public

@parhizj I still think you're insane to think it will drop to 1.2.

I'm excited for tomorrow.

edit// on another note, era5 preliminary data can be obtained in grib format the same day the global mean t2m's come out on the era5 csv if you dig around

@zenarxy Ok, I finally altered the plotting to see what the actual values were for both (had claude quickly interpolate ERA5 to the boxes) (to verify the values are roughly correct.. taking the difference from the mean of the boxes matches what I'd expect, which is actually a -0.16 C difference globally down to ~1.21)

>>> df_loti['anomaly_change'].mean() - df_era5['anomaly'].mean()

np.float64(0.15907094353876716)

Then a separate program to toggle the differences in the boxes (GISTEMP - ERA5) to see the effect so I wouldn't have to hand compute everything, finally yielding ...

Below, yellow is gistemp is warmer than era5 for the same box (the actual difference), while all the labels above the chart at the top are the calculations /80 (for the 80 boxes)

It does appear I was a bit optimistic (-0.13 instead of -0.19) on the size of the drop from those particular 3 boxes in Canada/Greenland as parts of the contintental US anomaly impinge on it and its not nearly as dramatic as it appears from ERA5 (this is why I probably failed visually for the most part).

Something I overlooked, but may not effect any change, is the far upper left box from Alaska and Canada which could are now infilled and biased cold as is from mostly Alaska stations, but its theoretically possible parts of NW Canada end up warming it depending on what stations we get (looking at last month and 2025 March it doesn't appear any stations will end up warming it significantly though -- it was hard to tell from ERA5 from the mollweide projection but the platee carree allows us to see it more clearly that some of them are neutral and even warm in one of the islands). That being said I don't think this is a factor based on what stations I think are expected/available, so I think the far upper left box may warm slightly but no where near what ERA5 shows.

On the other hand, I may have been too pessimistic about the chances elsewhere of COOLING as Africa shows far more opportunities for cooling than warming even if it is biased -- i.e. perhaps I am wrong about South Africa -- even if we don't get Mozambique, but do get Kenya, could get enough net cooling if the South African stations are cooler than I expect to bring it down a bin (any of the boxes would do). So if it does end up not being in the highest bin I think this is the reason so now...

In summary, based on this it looks like almost coin toss now (assuming Canada cools as expected, it will depend on what Africa data we get, all other things being equal)... I've updated my bets in line with it slightly favoring the higher bin.

~

per the ERA5 preliminary data, as far as I know they aren't available public (at least not from CDS). I am aware there are multiple different ERA5 versions, but none of them are available early, and recall the copernicus website explicitly says the preliminary data isn't public?

I don't think we have enough data as it stands for a release (Africa/Greenland/Mexico/Peru/etc...).

ghcnm.v4.0.1.20260406:

127.9895

The release was scheduled for Thursday, but it may be delayed as we don't even have most of Africa.

It appears we got most of Canada but I don't think we got all of it (should still cool further with more stations closer to the border and more from homogenization). Still no Greenland (although that is being infilled on the cold side).

Still a chance to reduce further I think with more stations in Canada (as ERA5 shows the 2 boxes as COLDER than the Alaska box, but that's not the case at the moment); even if we don't get all the cooling, the Alaskan box shows the two boxes should drop at least a further -1C each, if not more, but the Greeland box should warm slightly; but with this alone it will still be in the 1.25 to 1.26 C range.

If we were to get the rest of Africa stations in the next run then maybe it won't be delayed in which case the higher bin will be favored more heavily, but if its delayed that should only increase the chances over time of it dropping further. I'm increasing my bet on the higher bin further based on the higher bin is more likely at the moment, but the uncertainty from a delayed release and whatever tonight's data brings is enough to be not too confident.

ghcnm.v4.0.1.20260407:

127.9364

Aside from more of Europe stations coming in (which generally cooled their boxes, especially Turkey), we got a couple of South Africa stations (that cooled its box, and were on the cool side) and some of the smaller islands surrounding Madagascar came in, but still missing large parts of Africa. Also still no Greenland, Mexico, Peru, etc. so missing significant parts of Central and South America. Despite only missing about 9% of infilled data there is still a fairly significant coverage gap.

We also got the rest of Canada stations it seems, but did not get any reduction as I expected, and in fact actually got some warming in all 5 of the boxes. Even if Greenland stations comes in then I think we shouldn't expect any significant cooling with its contribution (and it may in fact warm).

I did look at the plots I did in August 2025 coverage when it got delayed, and we do have less infilled coverage now then when they did a preliminary release, so I think its still possible it gets delayed a couple days (not sure how likely though at 9% missing -- the 9% though is what's missing as far as INFILLED subboxes, not overall, which should be significantly larger, since all of Mexico, Central America, Greenland, etc for instance are infilled.).

However, considering the data that is missing, using the same tool as previously and its differences from ERA5 (i.e. adding up all the notable boxes from Africa, Central/South America), this does not bring it down below 1.25 (although somewhat close). It seems we will need a significant downward revision to the existing station data (i.e. we would have to assume that the Canada data would have to be incomplete as it is now) to easily bring the temp down a bin (or Peru / Africa stations would have to be quite a bit colder than ERA5 forecasted which seems unlikely).

I do not think that or a very large delay is likely (at latest a release by early next week) or any of the other scenarios is likely either, so it seems the 1.20-1.25 is increasingly unlikely.

ghcnm.v4.0.1.20260408:

129.8653

1.25+ seems all but assured now with data from Greenland, Mexico, and some more of Central America stations in, despite still missing quite a bit of Africa, especially since they haven't posted an update yet as of 10 am and the release is in an hour. So I guess we are going with the run from a couple days ago despite missing quite a bit of Africa (8% of infilled subboxes missing).

~

Diagnosing what went wrong this month, all signs still point towards a Canada warm bias (the two boxes highlighted below):

It's clear some of the warm biasing in Canada is from the relatively high station density in northern US compared to Canada: when you examine the isotherms of neutral anomalies and compare the station anomalies (and ERA5 which matches) to the GISTEMP subbox anomalies, the 0 deg. anomaly is far north (~500 km for some of the provinces) in the subbox compared to the station data (much worse around Ontario, Manitoba, Quebec where Hudson Bay additionally creates a bias towards the US stations).

Zooming in on the labeled ERA5 plot that is interpolated to the GISTEMP subboxes, and examining some of the stations from GHCNm it doesn't appear that ERA5 is too far off from GHCNm station anomalies at the subbox level (at least not the -2 C we see in the boxes!)

Compared to the zoomed in version of the (labeled) gistemp subboxes you can see the difference more clearly:

The 1200 km inclusion limit for a subbox doesn't seem hurt enough by the distance weighting:

But even the stations farther to the north are quite a bit warmer than you'd expect, so it's not just long distance stations and density creating the bias, some of it is that the temps that somewhat less extreme in ghcnm than the station data suggests so this has to be from the ghcn homogenization and simple ERA5 differences/biases compared to station data. I'd guess of the 2C gap in the boxes, 1C from actual station differences to ERA5 and 1C from some long distance station density bias.

The persistence of bmax here is something to behold!

@ScottSupak bmax?

@parhizj 1.24+ (maximum bracket)

April Market

April looking warm...

@ScottSupak Well, if you look at the general mean t2m it doesn't look alarming. What is alarming is how hot the ocean already is considering we are only just about to come out of La Nina. It does look like we're going to be hitting some records

@zenarxy indeed, there will be a lag between the ssts and air temp rising. But I was also dropping a subtle hint that there is no April market yet. ;)

@ScottSupak What values would you suggest for the 0.05C ranges? I will set it up soon.

@ChristopherRandles I'd say the midpoint should be 1.2

@zenarxy Agree. 1.2-1.25 bin, with 3 bins above & below should be plenty.

@parhizj perfect.