Looking only slightly biased up this month, +0.03

@zenarxy April anomaly has been cooler than March anomaly in last couple decades but this month looks like a repeat of last month at least spatially ... last 5 days missing so it will be warmer than shown, but the spatially warm/cold anomalies look somewhat similar.

The medium range ensembles generally following the very basic climatological rise you'd get from a statistical model...

I've been on the cool side repeatedly for last several months, so I've updated my error adjustment likewise to include all my own past forecast errors now, rather than excluding a couple tails (this raises the adjustment to +0.02 whatever the adjusted superensemble gives rather than a further +0.01)...

I don't think I will make any improvements in the ERA5->GISTEMP model as it stands for a long while -- until I can do a MCS type simulation essentially over station availability to calculate subbox/box biases/variances -- this means getting the weights for each station for each subbox & box for each year and then figuring out how to map ERA5/ECM onto those weights, in order to produce some spatial model that outputs bias, variances for the gridded data and incorporating that into my already complicated super ensemble pipeline sounds like a nightmare).

Forecast has really shifted downwards for mid-April.

@zenarxy bias shifted down to basically nothing

edit: Compared to forecast 4 days ago, notably more modest warm anomaly across central/eastern US and also generally elsewhere in the western hemisphere, but more anomalously warm poles (Greenland and Antarctica); However, looks like generally, northern Africa and Asia has warmed up a bit.

@parhizj and bias up to +0.05. so much variance this month

@zenarxy your bias numbers you have been putting out are relative to some objective ERA5 t2m ->GISTEMP model?

Based on my superensemble, the ERA5 t2m should be around 14.92 C for April 2026 (there is a slight adjustment upward based on each forecast day's long term super ensemble bias (relative to what validates in ERA5) that moves it up +0.02C from 14.90) in the current run).

For comparison (two closest years for ERA5 t2m for April are 2025 and 2020)...

April 2025: ERA5 t2m was 14.96 C

GISTEMP LOTI was 1.24C

April 2020: ERA5 t2m was 14.88 C

GISTEMP LOTI was 1.12 C

Currently the forecast puts it right in the middle of these in terms of ERA5 t2m, imlying a value around ~ 1.18 C just using these closest two data points.

~ This seems the simplest analysis one could do... ~

~~~~~~~~~~~~~~~~~~~~~~~

rant...

However in the usual ERA5->GISTEMP model I do use I adjust for ERA5 temp tries to detrend based on the absolute value of the ERA5 temp over time when applicable. I did a sensitivity type analysis and re-enabling the scorer picker (rather than just choosing the monthly model, let it choose the yearly one, which I disabled sometime last year, based on p-values) over ERA5 cutoffs starting from 1940,1951,1961,1971,1991 (1981 yielded no valid model), yielded center points (not finally adjusted by own forecast errors) of: 1.138, 1.177, 1.176, 1.146, 1.151 (the ~1.177 pair use the yearly model which has almost no downward adjustment); the mean is then 1.1576 (this is not far from the 1.163 I get after adjusting the 1.138 upwards by own forecast errors).

~

The motivation for this was that I did discover an error today in the same ERA5->GISTEMP model in the usual notebook I've been using for a long time that would have been biasing it slightly high actually for forecasts prior to the final ERA5 data (not actually effecting the final predictions at the end of the month, only further out, that would tend to make the forecast slightly high, i.e. 1.148 instead of 1.138 C; it was a weird bit of interpolation code that did not calculate the weights properly for the penultimate adjustment from the super ensemble bias on the forecast data only); fortunately I keep all of this separate from the actual ERA5 forecast data so none of the data I collected is effected.

That said, given the normal monthly model I've been using looks fine now, for the moment I will reference 1.163 C (the value after adjusting for past forecast errors using the usual 1940 cutoff).

~

Looking at the spatial data (below)....

Assuming a prior of a positive bias seems warranted given the last few months as I have been forecasting under, although this is from the extreme warm anomalies this winter in the NH, however the ERA5 t2m picture paints a modest warm anomaly in the 1991-2020 baseline compared to the last few months.

Breaking it down by usual suspects: I do expect the two boxes for Canada overlapping the continental US to be warmer than ERA5 shown below. The top left Alaska-Canada Arctic box is hard to tell since the subboxes from Russia near the pole will be very anomalously warm and influence it good amount. So, net warmer for these couple boxes.

For the Antarctica cold box its hard to tell (the single station should return a cold anomaly and make the box cold, but is not in the location of the most severe cold anomaly for the region, however it will infill alot of what ERA5 sees as warm anomalies as cool) so eyeballing that particular box from ERAE5 t2m and the station temp, net it looks close enough not to be a factor relative to the t2m. However, the other 3 Antarctica boxes look like they generally will be cooler in GISTEMP than what's suggested by ERA5's t2m based on what stations I expect to be available. So net cooler for Antarctica than suggested by ERA5.

For Namibia, it looks like it will be infilled on the warm side from the two South Africa stations we got last time (even though the t2m box anomaly is warm below from the ocean temps, it should become slightly warmer, but can't say how much). So net slightly warmer for boxes overlapping SW Africa.

Similarly, I recall ocean ssts was biased on the warm side compared to t2m last month. It seems like this should continue this month also. This should be the greatest net warmer contributor.

Subjectively, the usual suspects together point towards a moderate cool bias over land, and a warm bias for the ocean over ERA5 t2m relative to the combined subboxes.

~

Going back and looking at the subbox anomalies, separated by land/ocean for last month (to compare it against the other two closest years), March 2026 (these are all 1991-2020 baseline):

[LAND] Mean_diff=0.953, contrib_missing=-0.025, contrib_shared=0.316, green_area_1e3km2=15812.0, missing_area=8.3%

[OCEAN] Mean_diff=0.454, contrib_missing=0.000, contrib_shared=0.280, green_area_1e3km2=0.0, missing_area=0.0%

the land anomaly for the subboxes was +0.95 C and the ocean was +0.45 C relative to the 1991-2020 baseline for gistemp loti... (this will be different than the actual final gistemp global loti computations, due to the compositor which would weight the subboxes differently, and also finally apply different debiasing for certain bands):

for April 2025.. (1991-2020 baseline)

[LAND] Mean_diff=1.014, contrib_missing=0.000, contrib_shared=0.375, green_area_1e3km2=637.6, missing_area=0.3%

[OCEAN] Mean_diff=0.375, contrib_missing=0.000, contrib_shared=0.230, green_area_1e3km2=0.0, missing_area=0.0%

for April 2020... (1991-2020 baseline)

[LAND] Mean_diff=0.784, contrib_missing=-0.000, contrib_shared=0.290, green_area_1e3km2=637.6, missing_area=0.3%

[OCEAN] Mean_diff=0.318, contrib_missing=0.000, contrib_shared=0.195, green_area_1e3km2=0.0, missing_area=0.0%

I went back and plotted the data I extracted from the subbox anomalies from the ocean subboxes and the land subboxes, and the inertia and trends in the ocean anomalies is much clearer..

For reference... The April-March change is usually neutral, with a stddev is ~ 0.11 C, but given the situation I would expect the April-March change to be neutral-to-positive than negative given the situation, which implies a ocean anomaly warmer than 2025.

So, it looks over land 2020 might be a better guide as far as biases as the GISTEMP boxes as there are more scattered cold boxes in the NH (although the Alaska box will likely be infilled on the cool side), while for the ocean 2025 should be closer to 2026 of course.

So for April 2026 I think based on the above and below images, the land anomaly will be less than 2025 and possibly also 2020, but as for the ocean anomaly it seems from various sources (the EPAC in particular) that 2026 will be warmer than April 2025..

Interestingly when I compare the ERA5+forecast box interpolations to the GISTEMP Box values (i.e. the mixedBX npz) for April 2020 and 2025, on the differences between each box (i.e. taking that there is zero bias between the anomalies for the ERA5 interpolated boxes and the GISTEMP boxes, which definitely isn't the case), however, both data suggest a value of ~1.22 C (1.216 and 1.217 C respectively). This is a partly separate data computation (still relies on superensemble biases calculated, but then adjusts a gridded ERA5+superensemble and interpolates directly). I am not sure if its a coincidence that they line up so well, or whether this is superior to the other method I've been using -- it is only 2 data points: (GISTEMP 2025 - (ERA5&superensemble) 2026 April boxes, and for GISTEMP 2020 - ""; i.e +# means colder now )

I then looked at the same type of differencing for the actual GISTEMP loti values for April 2025 and April 2020 compared to ERA5 for the same months, and the differences were +0.017 and -0.025 respectively for those suggested values (so it doesn't seem the ERA5 interpolated to these boxes is all that terrible from these two samples).

Compare this to (including own forecast error adjustment of ~ +0.02 C) the ERA5->GISTEMP model that I am using has an actual (hindcast) error for 2025 and 2020 respectively of +0.038 C and -0.001 C.

Intuitively, from these two samples suggests that 1.163 C from the ERA5->GISTEMP might be low (suggesting nudging up to 1.18 C), while the suggested value of 1.216 C might be a tiny bit high, nudging it downwards to 1.21 C; the latter is close to the yearly model that doesn't have the absolute ERA5 temp monthly adjustment).

It is interesting since it provides a simpler, alternate method I haven't tried before (since I didn't have the gridded super-ensemble forecast until recently) that might be more productive than a global mean -- i.e. rather than developing a model that tries to simulate the biases at each step based on station availbility, sum the differences of the anomalies at the box level for the gistemp-box interpolated ERA5 target year against gistemp box anomalies for each reference year (each difference being a sample), and then take each sum against the actual LOTI values to get a final value. This seems more doable than my previous idea, but don't know if this will pan out; at the very least I would like to try to get trends over time for each box to see if that reveals any interesting patterns.

~~~~~~

Global temps for the month's forecast have only increased by about 0.01 C in the last day, but whatever apparent bias there might be has also probably gone up now a bit: more of the west coast of US is warmer than a couple days ago (which should bias more of Canada's fewer stations even warmer with homogenization & compositing), Kazakhstan and Russia have warmed up a bit, while parts of Europe have gotten slightly cooler.

The amount of differences in the gistemp boxes is far less compared to last month (especially as the land anomalies are milder), but last month we did have a 0.04C warmer ocean if I recall compared to ERA5 t2m over the same subboxes), and that probably should continue this month as we transition to El Nino ...

~

GISTEMP style Boxes

April 2026 (ERA5 and forecast), t2m:

April 2025 GISTEMP boxes:

April 2020 GISTEMP boxes:

Following up from a run I did much earlier today (up to 17/06Z), when the ERA5 preliminary from copernicus was a couple days behind what it is presently now at this moment (13th vs 15th), so tomorrow there may be a large jump when I update.

~

I followed up on this delta analog method using era5 - gistemp box anomalies using a varying model type/cutoffs per month to try to improve it.

Although it does seem to have a comparable (in terms of the validation data from either base models), or even slightly worse variance (in terms of my own predictions for the last year in hindcast), it looks robust enough for all months to use now and seems like a good idea since it diverges significantly from the other ERA5->GISTEMP model for some months in hindcast (sometimes better, sometimes worse for particular months).

For April, val_std=0.0572, pred_std=0.0054; where pred_std confirms my speculation yesterday that using many independent analogs seems a valid method, since the gistemp box anomalies from all the years stay strongly correlated with the actual gistemp loti value even if you change the baseline to 1991-2020 in each; I recalled incorrectly previously that there was some band/zonal weighting debiasing in GISTEMP after the boxes were created from when I went through the code some time ago, but this recollection was not the right interpretation (there is weighting of the bands but its only to preserve equal area since there are a different number of boxes in the different bands.)

Hindcast from the last 12 months shows that the variance from weighting the two predictions (by inv. variance on residuals of own predictions and hindcast for the other model) is only slightly worse (previous std dev of 0.048 from last 12 samples, becomes a std dev of 0.051 in the mixed), so I expect should make my own predictions more robust even if the combined variance appears marginally worse in the last 12 limited samples of hindcast (this results right now in weighting the old ERA5->GISTEMP model as ~ 60% and the new delta analog by about 40%).

After adjusting based on recent errors this yields, 1.162 and 1.209 from ERA5->GISTEMP old model, and the delta box anomaly analogs, and then weighting them (~0.6 and 0.4), yields a point estimate of 1.181 C.

This is the common sense value again based on the simple analysis I did yesterday.

@parhizj close to mine, I got 1.176 yesterday, 1.173 today