npj | climate and atmospheric science Article

Published in partnership with CECCR at King Abdulaziz University

https://doi.org/10.1038/s41612-024-00832-w

Climate model trend errors are evident in

seasonal forecasts at short leads
Check for updates
1,2 2,3
Jonathan D. Beverley , Matthew Newman & Andrew Hoell2,3

Climate models exhibit errors in their simulation of historical trends of variables including sea surface
temperature, winds, and precipitation, with important implications for regional and global climate
projections. Here, we show that the same trend errors are also present in a suite of initialised seasonal re-
forecasts for the years 1993–2016. These re-forecasts are produced by operational models that are
1234567890():,;
1234567890():,;

similar to Coupled Model Intercomparison Project (CMIP)-class models and share their historical external
forcings (e.g. CO2/aerosols). The trend errors, which are often well-developed at very short lead times,
represent a roughly linear change in the model mean biases over the 1993–2016 re-forecast record. The
similarity of trend errors in both the re-forecasts and historical simulations suggests that climate model
trend errors likewise result from evolving mean biases, responding to changing external radiative forcings,
instead of being an erroneous long-term response to external forcing. Therefore, these trend errors may
be investigated by examining their short-lead development in initialised seasonal forecasts/re-forecasts,
which we suggest should also be made by all CMIP models.

Simulated global and regional trends in historical climate model simulations example, the tropical Pacific trend error has persisted for so long32 that more
often show differences from those found in observations for many variables, recent studies have concluded that incorrect model internal variability is
including sea surface temperature, sea level pressure, wind, and precipitation most likely not its dominant cause1,5. Other studies have argued that some
[e.g. refs. 1–10]. Arguably the most prominent and extensively studied of trend errors may be transient [e.g.33], though this does not reconcile the
these differences is found in the tropical Pacific Ocean. An overall discrepancy in the historical simulations. Errors from other regions [e.g. the
strengthening of the east-to-west SST gradient has been observed due to Southern Ocean,34–36], missing feedbacks37, and misrepresentation of
statistically significant warming of the west Pacific and no significant trends in atmosphere-ocean processes have also been suggested as possible causes14,38.
the central and eastern Pacific11–15, which has coincided with a strengthening In this study, we will diagnose climate model trend errors by comparing
of the Walker circulation16,17. By contrast, the majority of historical climate them to trend errors in seasonal re-forecasts, which we find develop at sur-
model simulations indicate the opposite, with a weakening of the east-to-west prisingly short leads. When analysing uninitialised climate model simulations
SST gradient across the tropical Pacific4,5, which models also project to with changing external forcings (i.e., radiative forcings including CO2 and
continue18–22. Understanding the reasons for this disparity is increasingly aerosols), it can be difficult to ascertain what fraction of a trend is forced
important, as tropical SST trends impact climate sensitivity23, projected versus what fraction is a residual of internal variability. One approach often
ENSO teleconnections24–27, and atmospheric trends worldwide28,29. Moreover, used to identify the forced component and its errors is to average across a
SST trend discrepancies have been noted in other regions including the large ensemble of uninitialised climate realisations39,40. In this study, we
Atlantic, Indian, and Southern Oceans1,7. Extratropical trend discrepancies instead identify trend errors using initialised seasonal re-forecasts (sometimes
have also been found in historical climate simulations, including in storm termed hindcasts or retrospective forecasts), which use physical models
track-related zonal winds and precipitation2,3,8,10, regional drought6, and similar to those used in uninitialised climate model simulations, are run over a
surface temperature [e.g. central Asia/southern Europe7]. historical period, and use the same external forcings (see Methods Section Re-
Many possible reasons have been proposed to explain differences forecasts and references therein). Re-forecasts are also initialised with realistic
between observed and uninitialised climate model trends. Perhaps the most atmospheric and oceanic states, meaning that any trend errors are relative to
straightforward argument is to note that an observed trend may fall within the initial conditions. In contrast, when analysing a free-running model, the
the range of plausible model outcomes once internal variability is climate model realisations do not have to match observations over the same
considered30,31. However, as the trend discrepancies have continued to period of time. Therefore, analysing trend errors in seasonal re-forecasts may
persist for decades, this possibility appears to be increasingly unlikely. For help us diagnose the initial trend error development as these trend errors are

1
Cooperative Institute for Research in Environmental Sciences (CIRES), University of Colorado Boulder, Boulder, CO, USA. 2NOAA Physical Sciences Laboratory,
Boulder, CO, USA. 3These authors contributed equally: Matthew Newman, Andrew Hoell. e-mail: [email protected]

npj Climate and Atmospheric Science | (2024)7:285 1

https://doi.org/10.1038/s41612-024-00832-w Article

Fig. 1 | Relating trend errors and changing mean biases. Difference in multi-model (1993–2016) for different indices (shown as boxes on a and b). All model time series are
mean (MMM) mean bias between the first half (1993–2004) and second half for two-season leads, verifying in JJA. Units for a are Kelvin per decade and for b are mm
(2005–2016) of the common re-forecast record for March initialisation, JJA verification per day per decade. Hatching indicates a mean bias difference that is statistically sig-
(two-season lead) for a SST and b precipitation. (c, d) SST and e, f precipitation raw time nificant at the 5% level.
series for observations (black) and the multi-model mean (red) for the common years

relative to the observed realisation, thus minimising the uncertainty due to record (1993–2016; see Methods for a list of models). This is demonstrated by
internal variability. taking the difference in mean bias between the early (1993–2004) and late
The mean bias of a model represents the systematic error of the model (2005–2016) re-forecast years, as shown here for multi-model mean two-
relative to the observed state. For a seasonal prediction model, this is usually season lead re-forecasts verifying in JJA for SST (Fig. 1a) and precipitation (Fig.
calculated by averaging the errors relative to observations across all re-forecast 1b). This lead is chosen since the model mean biases have largely saturated by
years, for a given lead time and verification period. Note that for seasonal this time, but similar results are also seen for shorter leads, as is discussed
prediction models, mean biases do not represent loss of predictability, since further in Section Rapid development of re-forecast trend errors. There are
they are not random errors but rather represent systematic model errors that statistically significant differences in many areas for both SST and precipita-
occur across all forecasts. Since initialised re-forecasts and uninitialised model tion, which is also the case for other seasons (Supplementary Fig. 1).
simulations use similar physical models, many model errors are also common Another way of visualising the changing mean biases shown in Fig. 1a
to both, including mean biases in the tropical Pacific cold tongue and double and b is shown in the observed and multi-model mean (MMM) time series
Inter-Tropical Convergence Zone (ITCZ)41, the position and intensity of the for four locations (Fig. 1c–f). These time series suggest that the mean bias
Gulf Stream42,43, and the extratropical storm tracks44. Notably, some mean changes occur fairly continuously with time and in fact represent differ-
biases often develop rapidly as a function of forecast lead in seasonal pre- ences in observed and modelled trends across the re-forecast record. For
diction models [e.g.45,46], perhaps as a consequence of top-of-atmosphere some indices, such as Equatorial Pacific SST (Fig. 1c) and Southeast Asia
radiative imbalances47. These rapidly-developing mean biases in the atmo- precipitation (Fig. 1f), such differences may represent a reduction of the
sphere are distinct from the more slowly-evolving decadal oceanic drift that is mean bias by the end of the re-forecast record; for example, in the case of the
common among longer climate simulations [e.g.48]. Equatorial Pacific, an apparent reduction of the cold tongue mean bias. For
Here we will demonstrate that, even for re-forecast records of a few others, the mean bias may change sign (e.g. Southern Ocean SST, Fig. 1d), or
decades, many mean biases evident in initialised re-forecasts are not fixed but may grow over the re-forecast record (e.g. Equatorial West Africa pre-
rather change as the external forcings are changed. This leads to global trend cipitation, Fig. 1e). A similar picture of changing mean biases exists for both
errors in seasonal re-forecasts that also develop rapidly, often within months U200 and MSLP (not shown). For the remainder of this paper, we will
following initialisation. We will then show that these trend errors are very investigate these changing model mean biases by examining trend errors,
similar to those found in historical, uninitialised climate simulations. That is, defined as the differences in the slopes of the observed and modelled time
we suggest that model mean biases change as the imposed external forcings series best fit lines (e.g. Fig. 1c–f; see Methods Section Linear trends and
change, leading to similar trend errors in both re-forecasts and historical significance tests), in both re-forecasts and CMIP6 historical simulations for
climate simulations. the 1993–2016 period.

Results Rapid development of re-forecast trend errors

Changing re-forecast model mean biases As discussed in the introduction, many seasonal prediction model mean
First, we show how seasonal prediction model mean biases are not stationary, biases are known to develop rapidly, within the first few weeks following
but rather change roughly monotonically across their common re-forecast initialisation [e.g. refs. 45,46]. With this in mind, we next demonstrate how

npj Climate and Atmospheric Science | (2024)7:285 2

https://doi.org/10.1038/s41612-024-00832-w Article

Fig. 2 | Seasonal cycle and lead time dependence of re-forecast trend errors. Area- percentage of the observed (ERA5/GPCP) standard deviation for a given verification
averaged monthly a–d SST and e–h precipitation trend errors for a selection of different month. SST and precipitation regions are shown as boxes on Figs. 3 and 7. Statistical
regions and indices. Four different initialisation months are shown by the different significance is indicated by darker symbols with a black outline.
colours, with the x-axis indicating the verification month. Trend errors shown are as a

npj Climate and Atmospheric Science | (2024)7:285 3

https://doi.org/10.1038/s41612-024-00832-w Article

Fig. 3 | Seasonal variation of re-forecast sea surface temperature trend errors. averaged over two-season leads for verification in a DJF b MAM c JJA and d SON.
Multi-model mean sea surface temperature trend error relative to ERA5 Positive values indicate that the model trend is more positive (or less negative) than
(1993–2016) for four different initialisations (September, December, March, June) ERA5, and vice versa. Hatching indicates statistical significance at the 5% level.

rapidly the re-forecast trend errors develop, by comparing the dependence of are systematic errors suggestive of each forecast model going into its own
monthly trend errors upon the seasonal cycle and prediction lead time. state space. The stronger dependence of both precipitation and U200 trend
Multi-model mean re-forecast trend errors for a selection of different SST errors on the seasonal cycle compared to SST suggests that errors may be
and precipitation indices are shown in Fig. 2 for four different initialisation developing in the atmosphere first and reach something closer to saturation
months. For all variables, the trend errors strongly depend upon the seasonal there, whereas the SST errors take longer to develop and so have slightly
cycle and are mostly independent of the initialisation month and conse- more dependence on lead time. Errors relating to ENSO in re-forecasts have
quently the lead time, with significant trend errors often developing during likewise been suggested to have an atmospheric origin, showing similar
the first month following initialisation. For example, no matter the initi- levels of seasonal dependence, which is consistent with each model rapidly
alisation month, the Southern Ocean SST trend error (Fig. 2d) appears to transitioning to its own attractor50.
reach a maximum in July and a minimum in January, with roughly
monotonic evolution between these extrema. Similarly, for Atlantic Niño Trend errors in re-forecasts compared to CMIP6 simulations
(Fig. 2b), both the March and June initialisations show the same spike in For the remainder of the paper we will focus on seasonally-averaged trend
August, despite one having been initialised three months earlier. While we errors at two-season leads (see Methods). Significant SST trend errors are
have not investigated the development of trend errors for leads shorter than a found year-round and across the globe, including the tropical Pacific, the
month, we note that ref. 49 also found SST trend errors in Week 3 re- Indian Ocean, the Southern Ocean, and the Atlantic (Fig. 3). In the tropical
forecasts made by the European Centre for Medium-Range Weather Pacific, DJF, JJA, and SON trend errors resemble an El Niño-like pattern,
Forecasts Integrated Forecast System (their Fig. 5) in many regions, similar to the trend error pattern earlier found in the North American Multi-
including the central tropical Pacific, Atlantic Niño, and Gulf Stream, further model Ensemble (NMME) re-forecasts for the years 1982–202051,52. In the
demonstrating how rapidly these trend errors begin to develop. Still, for most Northwest Atlantic, the pattern is similar to the mean SST bias resulting
of the SST trend errors in Fig. 2 there remains some dependence on lead time. from errors in the Gulf Stream separation from the US eastern seaboard42,43.
The seasonality of the trend errors is even more striking for pre- These patterns of trend errors are largely consistent across the different
cipitation (Fig. 2e–h), where the errors are also almost independent of lead seasonal re-forecasts. Global pattern correlations for individual model trend
time. In some cases, such as Southeast Asia precipitation for May–January errors against the multi-model mean range between 0.55 and 0.85 (Sup-
verifications, the trend error has such similar magnitude for different plementary Fig. 3), thus indicating a systematic misrepresentation of
initialisation months that the lines sit almost on top of each other. This also observed SST trends in seasonal re-forecasts. The pattern of trend errors in
indicates that the errors are well developed at short lead times, in some cases Fig. 3c is also very similar to the difference in the mean bias between the two
within one month following initialisation. For example, in the Western US, halves of the re-forecast record shown in Fig. 1a (with the same also true for
the December verification trend error has almost identical magnitude for other seasons, Supplementary Fig. 1), with a global pattern correlation
both the December initialisation (re-forecast Month 1) and the September of 0.91.
initialisation (re-forecast Month 4). Similar levels of seasonal dependence of The SST trend error pattern in the tropical Pacific bears a strong
trend errors are also seen for 200 hPa zonal wind indices (Supplementary resemblance to the El Niño-like trend error exhibited by CMIP models over
Fig. 2). This decoupling of the re-forecast trend errors from the lead time the historical period1. To quantify this similarity both in the tropical Pacific
indicates that they are not simply due to a loss of predictability, but that they and globally, we define several indices based on the patterns of SST trend

npj Climate and Atmospheric Science | (2024)7:285 4

https://doi.org/10.1038/s41612-024-00832-w Article

Fig. 4 | Comparison of seasonal trend errors for SST indices for re-forecasts and trend error distribution and the third column shows two other observations-based
historical simulations. Seasonal trend error for a Equatorial Pacific b Atlantic datasets, relative to ERA5. As in Fig. 3, these are averaged for two-season leads for the
Niño c Gulf Stream d Southern Ocean e Niño Upwelling and f North Indian Ocean re-forecasts. Units are trend error per year as a percentage of the ERA5 index
SST (regions shown on Fig. 3). For each verification season, the first column shows standard deviation. Statistical significance is indicated by darker symbols with a
the seasonal re-forecast trend error for each individual model and the re-forecast black outline.
trend error for the multi-model mean, the second column shows the CMIP6 model

error in the re-forecast multi-model mean, which are shown as boxes on Fig. ERA5, as a measure of observational uncertainty. For comparison, we also
3 for SST. In Fig. 4 we show trend errors for these indices standardised by the reproduced this figure using the Months 1-3 re-forecast trend errors
observed standard deviation. The trend error in individual seasonal re- (Supplementary Fig. 4), and found generally similar results but with
forecasts is shown by the coloured scatter symbols, with the range of the somewhat weaker amplitudes.
CMIP6 ensemble illustrated by the box and whisker plots. We also include For all SST regions in Fig. 4, the CMIP6 model spread closely matches
the trend error of two additional observationally-based datasets, relative to the range of the seasonal re-forecasts, with the two different model sets

npj Climate and Atmospheric Science | (2024)7:285 5

https://doi.org/10.1038/s41612-024-00832-w Article

Fig. 5 | Seasonal variation of re-forecast 200 hPa zonal wind trend errors. Multi- season leads for verification in a DJF b MAM c JJA and d SON. Positive values
model mean 200 hPa zonal wind trend error relative to ERA5 (1993–2016) for four indicate that the model trend is more positive (or less negative) than ERA5, and vice
different initialisations (September, December, March, June) averaged over two- versa. Hatching indicates statistical significance at the 5% level.

tending to follow the same evolution across different seasons. The agree- between the re-forecasts and CMIP6 historical simulations (Fig. 6a). These
ment, at least in sign if not always magnitude, between the different seasonal errors are particularly large for the North Atlantic and Equatorial Africa jet
re-forecasts is also apparent, particularly for those indices and seasons where regions, with significant trend errors in all but one of the re-forecasts across
the multi-model mean is significant. Note also that the global mean SST the two regions.
trend overall is too warm in the seasonal re-forecasts and even more so in the Also included in Fig. 6 are two MSLP-based indices: the Arctic
CMIP6 historical simulations (Supplementary Fig. 5). Oscillation (AO) and North Pacific Indices (see Methods). As with the SST
There are also associated trend errors in mean sea level pressure and U200 indices, the CMIP6 ensemble closely matches the re-forecast
(MSLP, Supplementary Fig. 6) and 2 m temperature (Supplementary spread. This is particularly noticeable for the AO Index, which switches
Fig. 7). In the North Pacific for DJF, significant negative MSLP trend from a negative trend error in MAM to a positive one in JJA in both the
errors are indicative of an anomalous deepening and expansion of the seasonal re-forecasts and historical simulations, in agreement with the
Aleutian low (Supplementary Fig. 6a), with associated (though largely MSLP trend error maps (Supplementary Fig. 6). The agreement between the
not significant in the multi-model mean) 2 m temperature trend errors seasonal re-forecast trend errors from individual models is also striking,
over North America (Supplementary Fig. 7a). MSLP trend errors in the with all points clustered together for all verification seasons.
Arctic correspond to the negative phase of the Arctic Oscillation in Figure 7 shows that there are also significant precipitation trend errors
MAM (Supplementary Fig. 6b) and the positive phase in JJA (Sup- in many regions worldwide, which are largely in agreement with the SST and
plementary Fig. 6c), with similarly-matching 2 m temperature trend circulation (MSLP/U200) trend errors shown previously. These include
errors in the high latitudes that align with these phases53. MSLP trend precipitation trend errors that often look like the canonical El Niño signal
errors near Europe in DJF and SON also resemble the negative phase of (e.g. Southeast US in SON) or align with the climatological wet season (e.g.
the North Atlantic Oscillation, though these errors are not significant. Western US, Western/Eastern Equatorial Africa). Also of note is the north-
There is even greater similarity of the pattern of trend error across the south dipole extending from the far North Atlantic ( ~ 70∘ N) southwards to
different models for MSLP than for SST, with global pattern correla- Western Europe, which has been related to CMIP6 trend errors in the North
tions of individual models against the multi-model mean with a range Atlantic jet8, as seen in Fig. 5. These errors also have implications for sea-
of roughly 0.70–0.98 (Supplementary Fig. 8). sonal prediction skill, which will be the topic of a future study. The trend
Relatedly, models also exhibit significant trend errors in upper-level error patterns across the different models are again quite similar to each
(200 hPa) zonal wind (U200, Fig. 5). Many of these are related to well- other, with pattern correlations ranging from 0.75–0.95 against the multi-
known biases in the position and/or strength of the jet streams, such as the model mean (Supplementary Fig.10). The pattern of precipitation trend
North and South Pacific, North Atlantic, and tropical easterly jets. For errors in Fig. 7c also closely matches the mean bias difference between the
example, in the North Atlantic in JJA, trend errors are indicative of a greater first and second halves of the re-forecast record (Fig. 1b), with a global
northward displacement relative to the observed jet with time, a trend error pattern correlation of 0.89.
that has previously been found in SEAS554. Trend error patterns are similar The greater similarity of trend error patterns across the different
across all 9 re-forecast models, with pattern correlations between 0.71–0.98 models for precipitation, MSLP and U200 than for SST is suggestive of more
for individual models against the MMM (Supplementary Fig. 9). rapid atmospheric error development (Supplementary Figs. S3, S8, S9, S10).
Trend errors for selected 200 hPa zonal wind indices (three westerly As both mean biases and trend errors develop very rapidly (at short leads)
jets and one easterly jet; regions shown in Fig. 5) show close agreement within the re-forecasts, the similarity between the trend errors in the

npj Climate and Atmospheric Science | (2024)7:285 6

https://doi.org/10.1038/s41612-024-00832-w Article

Fig. 6 | Comparison of seasonal trend errors for 200 hPa zonal wind and MSLP shows the seasonal re-forecast trend error for each individual model and the re-
indices for re-forecasts and historical simulations. a Seasonal trend error for 200 forecast trend error for the multi-model mean, the second column shows the CMIP6
hPa zonal jet indices for the North Pacific jet (DJF), South Pacific jet (JJA), North model trend error distribution and the third column shows two other observations-
Atlantic jet (JJA), and Equatorial Africa easterly jet (SON) (regions shown on Fig. 5). based datasets, relative to ERA5. These are averaged for two-season leads for the re-
b and c show trend errors for two MSLP-based indices: b the Arctic Oscillation and forecasts. Units are trend error per year as a percentage of the ERA5 index standard
c North Pacific Index (see Methods). For each verification season, the first column deviation. Statistical significance is indicated by darker symbols with a black outline.

initialised seasonal re-forecasts and the uninitialised CMIP6 historical (Fig. 8e), with similar levels of consistency with the CMIP6 ensemble. We
simulations for SST, precipitation, MSLP and U200 is suggestive of all the again computed these regional trend errors for Months 1-3 and found very
models rapidly transitioning from their initialised states to their own little difference (not shown).
climates. On a global scale, both re-forecasts and CMIP6 historical simulations
Precipitation trend errors for selected indices (Fig. 8; see boxes on have trends towards becoming increasingly too wet (in agreement with
Fig. 7) show that both the seasonality of the errors and their similarity previous studies10), primarily during the extratropical cold seasons (Fig. 9),
between the re-forecasts and CMIP6 historical simulations are even more although observational uncertainty may also be an issue in the Southern
striking for precipitation than for SST (cf. Fig. 4). For example, in Southeast Hemisphere. These trend errors are linked to the storm track trend errors
Asia the switch between positive trend errors in JJA and negative trend shown previously (Fig. 5). For example, Southern Hemisphere precipitation
errors in SON is highly consistent between the two different model trend errors are largest in SON, coincident with the largest 200 hPa zonal
ensembles, with the re-forecast multi-model mean almost exactly matching wind errors.
the median CMIP6 value. Similar results are also seen for Western US, Calculating mean bias differences between the first 12 and last 12 years
Southeast US and Western Asia. The re-forecast multi-model mean values of the common re-forecast record for each region yields similar results to the
for individual seasons are significant for all but one of the African regions trend errors in Figs. 4 and 8, suggesting that the mean bias changes are

npj Climate and Atmospheric Science | (2024)7:285 7

https://doi.org/10.1038/s41612-024-00832-w Article

Fig. 7 | Seasonal variation of re-forecast precipitation trend errors. Multi-model trend is more positive (or less negative) than GPCP, and vice versa. Hatching
mean precipitation trend error relative to GPCP (1993–2016) for four different indicates statistical significance at the 5% level.
initialisations (September, December, March, June) for two-season leads for ver-
ification in a DJF b MAM c JJA and d SON. Positive values indicate that the model

largely linear for most regions and seasons (Supplementary Figs. climate simulations, because they are so similar to the re-forecast trend errors,
S11 and S12). Note also that for models with a longer re-forecast record, also likely represent the relatively rapid adjustment of the local model mean
including the three with re-forecasts of longer than 30 years (SEAS5, GEOS- bias to changing external forcing.
S2S, and SPEAR), we obtained similar SST and precipitation trend errors Our study cannot rule out the possibility that the imposed external
when using all years available for each model instead of only the common forcings themselves contain errors, which could drive similar trend
years (Supplementary Figs. S13 and S14). errors in both re-forecasts and historical simulations. This seems unli-
kely to be the dominant cause, as model mean biases are known to exist
Discussion independent of these radiative forcings, but it is entirely possible that it is
While several studies have examined trends, trend errors, and the statio- a contributory factor. This could be explored further using single forcing
narity of mean biases in either operational initialised seasonal re- runs, which is beyond the scope of the current study. Initialisation shock
forecasts51,52,55,56 or in historical climate model simulations5,7,57, to our is also unlikely to have a role here, as the seasonal prediction models
knowledge this is the first study that directly compares them, even though analysed include both first-of-the-month initialisations and lagged
both use generally similar models and are driven by similar CMIP external ensembles, as well as different initialisation datasets. These would all
forcings. We found that, over the 1993–2016 period, many common trend have to drive the same shock across much of the globe to result in the
errors in historical climate model simulations are also evident in seasonal re- trend error patterns that we find are so similar across different models.
forecasts, with a striking similarity in the amplitude and seasonal depen- Initialisation shock also cannot explain the trend errors in the unin-
dence of SST, precipitation, U200 and MSLP-based trend errors for many itialised models. All of the above points to a cause that is common across
different regions (Figs. 4, 6 and 8). This similarity is particularly noticeable different models, and is consistent with models transitioning to their
for trend errors of precipitation, U200 and MSLP-based indices, which own attractor, which also raises the possibility that there may be spurious
develop rapidly and are often fully developed within leads as short as one trends in model variability as well. Regardless of the cause, our results
month in the seasonal re-forecasts, suggesting that the errors may have an suggest that reducing or eliminating forecast model mean biases could
atmospheric origin. also help alleviate historical model trend errors.
Given how rapidly they develop, the re-forecast trend errors likely reflect Model mean biases can have various causes, including poorly repre-
the sensitivity of the model mean biases to the changing external forcing (e.g. sented atmosphere-ocean interactions, ocean processes and cloud radiative
CO2, aerosols). That is, the pattern and magnitude of trend errors for a given feedbacks58,59. Previous studies have argued that the rapid mean error
model, variable and lead time are very similar to the difference in mean bias development in forecasts may be used to help identify and correct defi-
between the first and second halves of the re-forecast record (cf. Fig. 1a, b and ciencies with model parameterisations that are used in both forecasts and
Figs. 3c and 7c). These mean biases are also known to develop rapidly (within climate model simulations60–62 or to help correct the mean biases
weeks following forecast initialisation) and we find that these re-forecast trend themselves63. Our results likewise suggest that diagnosis of climate model
errors develop on similar timescales. Moreover, since seasonal re-forecasts are trend errors should be made by examining the early development of the
initialised from observations, their trend errors are relative to a realisation mean bias in seasonal re-forecasts. We therefore suggest that all climate
drawn from nature and are unlikely to represent the uncertainty due to models that are used to produce CMIP7 climate projections should also be
internal variability. We therefore conclude that trend errors in the historical run as seasonal prediction models, with monthly start dates to capture the

npj Climate and Atmospheric Science | (2024)7:285 8

https://doi.org/10.1038/s41612-024-00832-w Article

Fig. 8 | Comparison of seasonal trend errors for precipitation indices for re- individual model and the re-forecast trend error for the multi-model mean, the
forecasts and historical simulations. Seasonal precipitation trend error for second column shows the CMIP6 model trend error distribution and the third
a Western US b Southeast US c Western Asia d Southeast Asia and e Equatorial East column shows another observations-based dataset, relative to GPCP. As in Fig. 7,
Africa (MAM), Southern Africa (MAM), Equatorial West Africa (JJA) Northeast these are averaged for two-season leads for the re-forecasts. Units are trend error per
Africa (JJA) and Equatorial Central Africa (SON) (regions shown on Fig. 7). For each year as a percentage of the GPCP index standard deviation. Statistical significance is
verification season, the first column shows the re-forecast trend error for each indicated by darker symbols with a black outline.

npj Climate and Atmospheric Science | (2024)7:285 9

https://doi.org/10.1038/s41612-024-00832-w Article

Fig. 9 | Comparison of global and hemispheric seasonal trend errors for pre- model distribution and the third column shows another observations-based
cipitation indices for re-forecasts and historical simulations. Seasonal pre- dataset, relative to GPCP. These are averaged over two-season leads for the re-
cipitation trend error for a the globe b northern hemisphere extratropics forecasts. Units are trend error per year as a percentage of the GPCP index
(30∘–90∘ N) c the tropics (30∘ S–30∘ N) and d southern hemisphere extratropics standard deviation. Statistical significance is indicated by darker symbols with a
(30∘–90∘ S). For each verification season, the first column shows the individual black outline.
re-forecasts and the multi-model mean, the second column shows the CMIP6

seasonal evolution of the errors. Since one uncertainty in our study is the and two for precipitation. For SST, U200 and MSLP, we use the Eur-
relatively short common re-forecast record, it would also be beneficial for opean Centre for Medium-Range Weather Forecasts (ECMWF) 5th
such re-forecasts to cover a longer period, to enable a more robust detection generation reanalysis [ERA5,65] as our primary reference dataset. We
of trends and trend errors. This would also allow for a direct comparison also include results based on the ECMWF Interim Reanalysis (ERA-
between trend errors from individual seasonal prediction and climate Interim,66, for SST, U200 and MSLP), the Hadley Centre Sea Ice and
models, which was not possible here due to the models available. SST (HadISST) dataset (for SST) and the Japanese 55-year Reanalysis
Given the importance of future trends in variables such as SST and [JRA55,67] for U200 and MSLP to give an indication of observational
precipitation to regional and global climate projections, diagnosing the causes uncertainty. For 2 m temperature (only used in the Supplementary
of model trend errors and working to alleviate them is essential. Since many of Information), we use ERA5. Our primary precipitation reference
the historical trend errors we have identified appear to have evolved roughly dataset is the Global Precipitation Climatology Project (GPCP)
linearly over the re-forecast period, it is possible that they will continue into analysis68. We also compare against the Climate Prediction Centre
the future and thereby impact projected trends as well. Until we understand Merged Analysis of Precipitation (CMAP).
more about the potential sensitivity of model mean biases to external forcing
and uncertainty in forcing fields, and are able to reduce these model mean Re-forecasts
biases so that their errors will not confuse the trends we are trying to discern64, We use multi-decade re-forecasts from 11 different operational seasonal
we have to examine any model trends in both historical and projected prediction models: the European Centre for Medium-Range Weather
simulations with great care. Forecasts (ECMWF) SEAS569, Deutscher Wetterdienst (DWD) GCFS2.170,
Environment and Climate Change Canada (ECCC) CanCM4i71, ECCC
Methods GEM5-NEMO71, Euro-Mediterranean Centre on Climate Change (CMCC)
Observations-based datasets SPS3.572, National Aeronautics and Space Administration (NASA) GEOS-
We use three observational estimates for sea surface temperature S2S73, Geophysical Fluid Dynamics Laboratory (GFDL) SPEAR74, United
(SST), 200 hPa zonal wind (U200) and mean sea level pressure (MSLP) Kingdom Met Office GloSea6-GC3.275, Meteo-France System 876, National

npj Climate and Atmospheric Science | (2024)7:285 10

https://doi.org/10.1038/s41612-024-00832-w Article

Centres for Environmental Prediction (NCEP) CFSv277 and Japan can be downloaded at http://iridl.ldeo.columbia.edu/SOURCES/.Models/.
Meteorological Agency (JMA) CPS378. Unless otherwise stated, analysis is NMME/.NASA-GEOSS2S/.HINDCAST/.MONTHLY/and http://iridl.ldeo.
performed for the common period across all 11 models (1993–2016). columbia.edu/SOURCES/.Models/.NMME/.GFDL-SPEAR/.HINDCAST/.
All of the seasonal prediction models listed above are initialised MONTHLY/, respectively. The CMIP6 data are available for download at
every year of their re-forecast period (Table S1) and each initialisation https://esgf-node.llnl.gov/search/cmip6/. ERA5 data can be downloaded at
is run out to approximately six months. In this study we only use four https://doi.org/10.24381/cds.f17050d7. ERA-Interim, HadISST, JRA55,
different monthly initialisations for each model (March, June, Sep- GPCP and CMAP data can be downloaded at https://psl.noaa.gov/data/
tember and December), and we primarily analyse seasonal averages gridded/.
over the fourth, fifth and sixth months of the re-forecasts, which we
term “two-season lead”. For example, for a March initialisation, Code availability
verification would be the June-July-August average. We use SST and The underlying code for this study is not publicly available but may be made
precipitation data from all models. MSLP, U200 and 2 m temperature available to qualified researchers on reasonable request from the corre-
data were not available for GEOS-S2S or SPEAR, so analysis of these sponding author.
variables excludes these models. Further model details can be found
in Supplementary Table S1. Received: 26 June 2024; Accepted: 4 November 2024;

Climate models
We also use model simulations from the Coupled Model Intercomparison References
Project Phase 6 [CMIP6,79] from 38 different models (listed in Supple- 1. Wills, R. C., Dong, Y., Proistosecu, C., Armour, K. C. & Battisti, D. S.
mentary Table S2). We use the historical simulations for the years Systematic climate model biases in the large-scale patterns of recent
1993–2014, and the Shared Socioeconomic Pathway (SSP)-245, a scenario sea-surface temperature and sea-level pressure change. Geophys.
with medium radiative forcing by the end of the century, for years Res. Lett. 49, e2022GL100011 (2022).
2015–2016, to align with the re-forecast record. We use surface temperature, 2. Vicente-Serrano, S. M. et al. Do CMIP models capture long-term
precipitation, MSLP and U200 data from these models. observed annual precipitation trends? Clim. Dyn. 58, 2825–2842 (2022).
All data are regridded onto a common 1-degree grid prior to analysis. 3. Donat, M. G. et al. How credibly do CMIP6 simulations capture
historical mean and extreme precipitation changes? Geophys. Res.
Linear trends and significance tests Lett. 50, e2022GL102466 (2023).
To calculate what we term “trend error” for each climate model or seasonal 4. Coats, S. & Karnauskas, K. Are simulated and observed twentieth century
prediction model, we take data for the relevant verification period from each tropical Pacific sea surface temperature trends significant relative to
climate model simulation or yearly re-forecast initialisation and calculate internal variability? Geophys. Res. Lett. 44, 9928–9937 (2017).
the error relative to observations. The trend error is then calculated by 5. Seager, R., Henderson, N. & Cane, M. Persistent discrepancies
computing the slope in the least squares estimator of this error over the between observed and modeled trends in the tropical Pacific Ocean.
relevant period, typically 1993–2016. For example, to determine the sea- J. Clim. 35, 4571–4584 (2022).
sonal prediction model DJF SST trend error shown in Fig. 4a, we calculate 6. Nasrollahi, N. et al. How well do CMIP5 climate simulations replicate
the area-averaged Equatorial Pacific DJF error for each yearly October historical trends and patterns of meteorological droughts? Water
initialisation from 1993–2016 to create a time series of the error, and then Resour. Res. 51, 2847–2864 (2015).
calculate the trend of that error. Note that this is the same as calculating the 7. Bhend, J. & Whetton, P. Consistency of simulated and observed
trend of the observations and models separately and calculating the dif- regional changes in temperature, sea level pressure and precipitation.
ference in the slopes. In the case of Fig. 4, this is then normalised by the Clim. Change 118, 799–810 (2013).
observed standard deviation of the relevant index. Significance is computed 8. Blackport, R. & Fyfe, J. C. Climate models fail to capture strengthening
using the Hamed and Rao modification to the Mann-Kendall trend test to wintertime North Atlantic jet and impacts on Europe. Sci. Adv. 8,
account for serial autocorrelation80, and significance is shown at the 5% level. eabn3112 (2022).
Significance for mean bias differences is calculated by comparing to a 9. Vautard, R. et al. Heat extremes in Western Europe increasing faster
distribution of 1000 randomly sampled mean bias differences, sampled with than simulated due to atmospheric circulation trends. Nat. Comm. 14,
replacement from the original time series. The actual mean bias difference is 6803 (2023).
labelled as significant if it falls above or below the 97.5th or 2.5th percentile of 10. Gu, G. & Adler, R. F. Observed variability and trends in global
the bootstrapped distribution, respectively. Note that this may give rise to precipitation during 1979–2020. Clim. Dyn. 61, 131–150 (2023).
slightly higher levels of significance overall than the Mann-Kendall trend 11. Cane, M. A. et al. Twentieth-century sea surface temperature trends.
test (cf. Fig. 1a and b), due to limitations of the Mann-Kendall test with Science 275, 957–960 (1997).
shorter time series81. 12. Karnauskas, K. B., Seager, R., Kaplan, A., Kushnir, Y. & Cane, M. A.
Observed strengthening of the zonal sea surface temperature
Indices gradient across the equatorial Pacific Ocean. J. Clim. 22, 4316–4321
The North Pacific Index, as defined by Trenberth and Hurrell82, is the MSLP (2009).
averaged over 30∘–65∘ N, 160∘ E–140∘ W. To calculate the Arctic Oscillation 13. Solomon, A. & Newman, M. Reconciling disparate twentieth-century
Index, the leading EOF of ERA5 1000 hPa geopotential height anomalies Indo-Pacific ocean temperature trends in the instrumental record.
poleward of 20∘ N, with seasonal cycle removed, is computed for the period Nat. Clim. Change 2, 691–699 (2012).
1979–2000. 1000 hPa geopotential height anomalies from individual models 14. Seager, R. et al. Strengthening tropical Pacific zonal sea surface
(computed from MSLP using the approximate relation Z = 8(MSLP−1000), temperature gradient consistent with rising greenhouse gases. Nat.
following Thompson and Wallace83) are then projected onto this loading Clim. Change 9, 517–522 (2019).
pattern. All other indices are calculated by averaging over the regions shown 15. Lee, S. et al. On the future zonal contrasts of equatorial Pacific climate:
as boxes on Figs. 3, 5 and 7. Perspectives from Observations, Simulations, and Theories. npj Clim.
Atmos. Sci. 5, 82 (2022).
Data availability 16. L’Heureux, M. L., Lee, S. & Lyon, B. Recent multidecadal
Monthly Copernicus seasonal forecast model data are available at https://doi. strengthening of the Walker circulation across the tropical Pacific.
org/10.24381/cds.68dd14c3. NASA GEOS-S2S and GFDL-SPEAR hindcasts Nat. Clim. Change 3, 571–576 (2013).

npj Climate and Atmospheric Science | (2024)7:285 11

https://doi.org/10.1038/s41612-024-00832-w Article

17. Sohn, B., Yeh, S.-W., Schmetz, J. & Song, H.-J. Observational 40. Kay, J. E. et al. The Community Earth System Model (CESM) large
evidences of Walker circulation change over the last 30 years ensemble project: a community resource for studying climate change
contrasting with GCM results. Clim. Dyn. 40, 1721–1732 (2013). in the presence of internal climate variability. Bull. Am. Meteor. Soc.
18. Liu, Z., Vavrus, S., He, F., Wen, N. & Zhong, Y. Rethinking tropical 96, 1333–1349 (2015).
ocean response to global warming: The enhanced equatorial 41. Li, G. & Xie, S.-P. Tropical biases in CMIP5 multimodel ensemble: The
warming. J. Clim. 18, 4684–4700 (2005). excessive equatorial Pacific cold tongue and double ITCZ problems.
19. Xie, S.-P. et al. Global warming pattern formation: Sea surface J. Clim. 27, 1765–1780 (2014).
temperature and rainfall. J. Clim. 23, 966–986 (2010). 42. Lee, R. W. et al. Impact of Gulf Stream SST biases on the global
20. Yeh, S.-W., Ham, Y.-G. & Lee, J.-Y. Changes in the tropical Pacific atmospheric circulation. Clim. Dyn. 51, 3369–3387 (2018).
SST trend from CMIP3 to CMIP5 and its implication of ENSO. J. Clim. 43. Chassignet, E. P. et al. Impact of horizontal resolution on global
25, 7764–7771 (2012). ocean-sea-ice model simulations based on the experimental
21. Huang, P. & Ying, J. A multimodel ensemble pattern regression protocols of the Ocean Model Intercomparison Project phase 2
method to correct the tropical Pacific SST change patterns under (OMIP-2). Geosci. Model Dev. Dis. 2020, 1–58 (2020).
global warming. J. Clim. 28, 4706–4723 (2015). 44. Priestley, M. D. K., Ackerley, D., Catto, J. L. & Hodges, K. I. Drivers of
22. Cai, W. et al. Changing El Niño–Southern oscillation in a warming biases in the CMIP6 extratropical storm tracks. Part I: Northern
climate. Nat. Rev. Earth Environ. 2, 628–644 (2021). Hemisphere. J. Clim. 36, 1451–1467 (2023).
23. Colman, R. & Power, S. B. What can decadal variability tell us about 45. Green, B. W., Sinsky, E., Sun, S., Tallapragada, V. & Grell, G. A.
climate feedbacks and sensitivity? Clim. Dyn. 51, 3815–3828 (2018). Sensitivities of subseasonal unified forecast system simulations to
24. Cai, W. et al. Increasing frequency of extreme El Niño events due to changes in parameterizations of convection, cloud microphysics, and
greenhouse warming. Nat. Clim. Change 4, 111–116 (2014). planetary boundary layer. Mon. Wea. Rev. 151, 2279–2294 (2023).
25. Yeh, S.-W. et al. ENSO atmospheric teleconnections and their 46. Song, S. & Mapes, B. Interpretations of systematic errors in the NCEP
response to greenhouse gas forcing. Rev. Geophys. 56, 185–206 Climate Forecast System at lead times of 2, 4, 8,..., 256 days. J. Adv.
(2018). Model. Earth Syst. 4 (2012).
26. Beverley, J. D., Collins, M., Lambert, F. H. & Chadwick, R. Future 47. Magnusson, L., Alonso-Balmaseda, M., Corti, S., Molteni, F. &
changes to El Niño teleconnections over the North Pacific and North Stockdale, T. Evaluation of forecast strategies for seasonal and
America. J. Clim. 34, 6191–6205 (2021). decadal forecasts in presence of systematic model errors. Clim. Dyn.
27. Beverley, J. D., Collins, M., Lambert, F. H. & Chadwick, R. Drivers of 41, 2393–2409 (2013).
changes to the ENSO–Europe Teleconnection under future warming. 48. Gupta, A. S., Jourdain, N. C., Brown, J. N. & Monselesan, D. Climate
Geophys. Res. Lett. 51, e2023GL107957 (2024). drift in the CMIP5 models. J. Clim. 26, 8597–8615 (2013).
28. Chung, C. E. & Ramanathan, V. Relationship between trends in land 49. Wulff, C. O., Vitart, F. & Domeisen, D. I. Influence of trends on
precipitation and tropical SST gradient. Geophys. Res. Lett. 34 (2007). subseasonal temperature prediction skill. Quart. J. Roy. Meteor. Soc.
29. Deser, C. & Phillips, A. S. Atmospheric circulation trends, 1950–2000: 148, 1280–1299 (2022).
The relative roles of sea surface temperature forcing and direct 50. Beverley, J. D., Newman, M. & Hoell, A. Rapid development of
atmospheric radiative forcing. J. Clim. 22, 396–413 (2009). systematic ENSO-related seasonal forecast errors. Geophys. Res.
30. Olonscheck, D., Rugenstein, M. & Marotzke, J. Broad consistency Lett. 50, e2022GL102249 (2023).
between observed and simulated trends in sea surface temperature 51. Shin, C.-S. & Huang, B. A spurious warming trend in the NMME
patterns. Geophys. Res. Lett. 47, e2019GL086773 (2020). equatorial Pacific SST hindcasts. Clim. Dyn. 53, 7287–7303 (2019).
31. Watanabe, M., Dufresne, J.-L., Kosaka, Y., Mauritsen, T. & Tatebe, H. 52. L’Heureux, M. L., Tippett, M. K. & Wang, W. Prediction challenges
Enhanced warming constrained by past trends in equatorial Pacific from errors in tropical Pacific sea surface temperature trends. Front.
sea surface temperature gradient. Nat. Clim. Change 11, 33–37 Clim. 4, 837483 (2022).
(2021). 53. Iles, C. & Hegerl, G. Role of the North Atlantic Oscillation in decadal
32. Kociuba, G. & Power, S. B. Inability of CMIP5 models to simulate temperature trends. Environ. Res. Lett. 12, 114010 (2017).
recent strengthening of the Walker circulation: Implications for 54. Patterson, M., Befort, D. J., O’Reilly, C. H. & Weisheimer, A. Drivers of
projections. J. Clim. 28, 20–35 (2015). the ECMWF SEAS5 seasonal forecast for the hot and dry European
33. Heede, U. K. & Fedorov, A. V. Eastern equatorial Pacific warming summer of 2022. Quart. J. Roy. Meteor. Soc. (2024).
delayed by aerosols and thermostat response to CO2 increase. Nat. 55. Shao, Y., Wang, Q. J., Schepen, A. & Ryu, D. Embedding trend into
Clim. Change 11, 696–703 (2021). seasonal temperature forecasts through statistical calibration of GCM
34. Hwang, Y.-T., Xie, S.-P., Deser, C. & Kang, S. M. Connecting tropical outputs. Int. J. Climatol. 41, E1553–E1565 (2021).
climate change with Southern Ocean heat uptake. Geophys. Res. Lett. 56. Balmaseda, M. A. et al. Skill assessment of seasonal forecasts of
44, 9449–9457 (2017). ocean variables. Front. Mar. Sci. 11, 1380545 (2024).
35. Andrews, T. et al. On the effect of historical SST patterns on radiative 57. Kerkhoff, C., Künsch, H. R. & Schär, C. Assessment of bias
feedback. J. Geophys. Res. Atmos. 127, e2022JD036675 (2022). assumptions for climate models. J. Clim. 27, 6799–6818 (2014).
36. Dong, Y., Armour, K. C., Battisti, D. S. & Blanchard-Wrigglesworth, E. 58. Richter, I. Climate model biases in the eastern tropical oceans: Causes,
Two-way teleconnections between the Southern Ocean and the impacts and ways forward. WIREs Clim. Change 6, 345–358 (2015).
tropical Pacific via a dynamic feedback. J. Clim. 35, 6267–6282 59. Zhang, Q., Liu, B., Li, S. & Zhou, T. Understanding models’ global sea
(2022). surface temperature bias in mean state: from CMIP5 to CMIP6.
37. Kang, S. M., Ceppi, P., Yu, Y. & Kang, I.-S. Recent global climate Geophys. Res. Lett. 50, e2022GL100888 (2023).
feedback controlled by Southern Ocean cooling. Nat. Geosci. 16, 60. Phillips, T. J. et al. Evaluating parameterizations in general circulation
775–780 (2023). models: Climate simulation meets weather prediction. Bull. Am.
38. Fu, M. & Fedorov, A. The role of Bjerknes and shortwave feedbacks in Meteor. Soc. 85, 1903–1916 (2004).
the tropical Pacific SST response to global warming. Geophys. Res. 61. Martin, G. et al. Analysis and reduction of systematic errors through a
Lett. 50, e2023GL105061 (2023). seamless approach to modeling weather and climate. J. Clim. 23,
39. Deser, C., Phillips, A., Bourdette, V. & Teng, H. Uncertainty in climate 5933–5957 (2010).
change projections: the role of internal variability. Clim. Dyn. 38, 62. Randall, D. A. & Emanuel, K. The Weather-Climate Schism. Bull. Amer.
527–546 (2012). Meteor. Soc. 105, E300–E305 (2023).

npj Climate and Atmospheric Science | (2024)7:285 12

https://doi.org/10.1038/s41612-024-00832-w Article

63. Ma, H.-Y. et al. On the correspondence between seasonal forecast 83. Thompson, D. W. & Wallace, J. M. The Arctic Oscillation signature in
biases and long-term climate biases in sea surface temperature. J. the wintertime geopotential height and temperature fields. Geophys.
Clim. 34, 427–446 (2021). Res. Lett. 25, 1297–1300 (1998).
64. Palmer, T. & Stevens, B. The scientific challenge of understanding and
estimating climate change. Proc. Natl Acad. Sci. 116, 24390–24395 Acknowledgements
(2019). The scientific results and conclusions, as well as any views or opinions
65. Hersbach, H. et al. The ERA5 global reanalysis. Quart. J. Roy. Meteor. expressed herein, are those of the author(s) and do not necessarily reflect the
Soc. 146, 1999–2049 (2020). views of NOAA or the Department of Commerce. This work was supported
66. Dee, D. P. et al. The ERA-Interim reanalysis: Configuration and by NOAA cooperative agreement NA22OAR4320151 and the Famine Early
performance of the data assimilation system. Quart. J. Roy. Meteor. Warning Systems Network. The authors wish to thank Maria Gehne and the
Soc. 137, 553–597 (2011). three reviewers for their helpful feedback.
67. Kobayashi, S. et al. The JRA-55 reanalysis: General specifications and
basic characteristics. J. Meteor. Soc. Jpn. 93, 5–48 (2015). Author contributions
68. Adler, R. F. et al. The Global Precipitation Climatology Project (GPCP) J.D.B., M.N. and A.H. all contributed to the design and implementation of the
monthly analysis (new version 2.3) and a review of 2017 global research, to the analysis of the results and to the writing of the manuscript. All
precipitation. Atmosphere 9, 138 (2018). authors read and approved the manuscript.
69. Johnson, S. J. et al. SEAS5: the new ECMWF seasonal forecast
system. Geosci. Model Dev. 12, 1087–1117 (2019). Competing interests
70. Fröhlich, K. et al. The German climate forecast system: GCFS. J. Adv. The authors declare no competing interests.
Model. Earth Syst. 13, e2020MS002101 (2021).
71. Lin, H. et al. The Canadian seasonal to interannual prediction system Additional information
version 2 (CanSIPSv2). Weather. Forecast. 35, 1317–1343 (2020). Supplementary information The online version contains
72. Gualdi, S. et al. The new CMCC operational seasonal prediction supplementary material available at
system. CMCC Tech. Note. TN0288, 1–34 (2020). https://doi.org/10.1038/s41612-024-00832-w.
73. Molod, A. et al. GEOS-S2S version 2: The GMAO high-resolution
coupled model and assimilation system for seasonal prediction. J. Correspondence and requests for materials should be addressed to
Geophys. Res. Atmos. 125, e2019JD031767 (2020). Jonathan D. Beverley.
74. Delworth, T. L. et al. SPEAR: The next generation GFDL modeling
system for seasonal to multidecadal prediction and projection. J. Adv. Reprints and permissions information is available at
Model. Earth Syst. 12, e2019MS001895 (2020). http://www.nature.com/reprints
75. MacLachlan, C. et al. Global Seasonal forecast system version 5
(GloSea5): A high-resolution seasonal forecast system. Quart. J. Roy. Publisher’s note Springer Nature remains neutral with regard to
Meteor. Soc. 141, 1072–1084 (2015). jurisdictional claims in published maps and institutional affiliations.
76. Batté, L., Dorel, L., Ardilouze, C. & Guérémy, J. Documentation of the
METEO-FRANCE Seasonal Forecasting System 8. Technical Report Open Access This article is licensed under a Creative Commons
(2021). Attribution-NonCommercial-NoDerivatives 4.0 International License,
77. Saha, S. et al. The NCEP climate forecast system version 2. J. Clim. which permits any non-commercial use, sharing, distribution and
27, 2185–2208 (2014). reproduction in any medium or format, as long as you give appropriate
78. Hirahara, S. et al. Japan Meteorological Agency/Meteorological credit to the original author(s) and the source, provide a link to the Creative
Research Institute Coupled Prediction System Version 3 (JMA/MRI- Commons licence, and indicate if you modified the licensed material. You
CPS3). J. Meteor. Soc. Jpn. advpub, 2023–009 (2023). do not have permission under this licence to share adapted material
79. Eyring, V. et al. Overview of the Coupled Model Intercomparison derived from this article or parts of it. The images or other third party
Project Phase 6 (CMIP6) experimental design and organization. material in this article are included in the article’s Creative Commons
Geosci. Model Dev. 9, 1937–1958 (2016). licence, unless indicated otherwise in a credit line to the material. If material
80. Hamed, K. H. & Rao, A. R. A modified Mann-Kendall trend test for is not included in the article’s Creative Commons licence and your intended
autocorrelated data. J. Hydrol. 204, 182–196 (1998). use is not permitted by statutory regulation or exceeds the permitted use,
81. Wang, F. et al. Re-evaluation of the power of the Mann-Kendall test for you will need to obtain permission directly from the copyright holder. To
detecting monotonic trends in hydrometeorological time series. Front. view a copy of this licence, visit http://creativecommons.org/licenses/by-
Earth. Sci. 8, 14 (2020). nc-nd/4.0/.
82. Trenberth, K. E. & Hurrell, J. W. Decadal atmosphere-ocean variations
in the Pacific. Clim. Dyn. 9, 303–319 (1994). © The Author(s) 2024

npj Climate and Atmospheric Science | (2024)7:285 13