| 11 Aug 2025
Seamless climate information for the next months to multiple years: merging of seasonal and decadal predictions, and their comparison to multi-annual predictions
Abstract. Stakeholders across climate-sensitive sectors often require climate information that spans multiple timescales, e.g. from months to several years, to inform planning and decision-making. To satisfy this information request, climate services are typically developed by separately using seasonal predictions for the first few months, and decadal predictions for subsequent years. This shift in information source can introduce inconsistencies. To ensure the information is consistent across forecast time scales, some centres have produced initialised multi-annual predictions, run twice a year and covering 2–3 years ahead, with increased ensemble sizes. An alternative methodology to provide coherent climate information across timescales involves temporal merging, where seamless predictions are created by postprocessing seasonal and decadal forecasts in combination. One approach selects members from large ensembles of decadal predictions or climate projections that closely align with seasonal predictions or past observations, transferring short-term predictability into longer timescales.
This study evaluates the skill of seamless forecasts using different constraints (e.g. variables, regions, temporal aggregations), and compares them with initialised multi-annual predictions. The analysis focuses on predictions of the Niño3.4 index and spatial fields of surface temperature, precipitation, and sea level pressure for the first three forecast years. Results show that while initialised multi-annual predictions achieve the highest overall skill, temporally merged forecasts offer a computationally efficient alternative that still performs well and can be produced regularly as monthly updates of observations or seasonal predictions become available. Besides, both sets of predictions outperform the unconstrained ensembles of decadal predictions and climate projections over large regions. During the period where the seasonal predictions and seamless predictions overlap, their skill is comparable. These findings illustrate the potential of temporal merging as a cost-effective strategy for extending climate information across timescales and enhancing coherence for operational climate services provision.
In the manuscript entitled ‘Seamless climate information for the next months to multiple years: merging of seasonal and decadal predictions, and their comparison to multi-annual predictions’, Delgado-Torres and colleagues evaluate the added value of seamless forecasts from multi-annual predictions and from several methods based on constraining large ensembles of simulations, in comparison with seasonal and decadal prediction systems, as well as with ‘non-initialized’ large ensembles of historical simulations, to predict the Niño 3.4 index and spatial fields of temperature, precipitation, and sea-level pressure. Overall, I found that the authors carried out interesting analyses that highlight the relevance of both multi-annual predictions and constraining methods, the latter being a cost-effective alternative that can be updated much more frequently. I have some minor issues and comments, especially regarding the evaluations.
Title :
I found the title not very clear. I am not sure that the term ‘merging’ is appropriate here, as there is no actual merging of seasonal and decadal predictions in the study, but rather a constraint of decadal predictions and historical simulations based on seasonal predictions. If you used the term ‘merging’ in the sense of combining different data, then ‘blending’ may be more suitable.
Introduction :
l.38-41: I wouldn’t describe the methods cited here as ‘temporal merging’ methods, since they do not merge time series (this approach is not used in these studies). Indeed, they use observations or decadal predictions to constrain large ensembles of non-initialized historical simulations. The term ‘temporal merging’ is more consistent with the study of Befort et al. (2022), cited on line 50, where historical simulations and decadal predictions are concatenated.
Data :
l.96 : The term ‘climate projection’ with HIST as a reference is misleading, especially since there are not only climate projections but also historical simulations.
Method:
Fig S1 : What does « accum » mean ?
l.110 : Can you provide more explanation on the « bias adjustments (correcting both the mean and variance) »
Results:
Fig S3b : It is confusing for the November initialization that the skill from DP just after initialization (dark green), which starts in January as indicated in the legend, is shown as starting at the same month (0) as the other dataset that begins in November. Shouldn’t it instead start at month 2 to be consistent with the other dataset?
l.217-219 : Indeed, this is not a very fair comparison with decadal predictions. It would be preferable to use the same representation as in Fig. S3B, based on the DP system initialized in November.
Fig S4 and S5 : As in my previous comment, it would be preferable to also include the DP system initialized in November for the November forecast in the Figures.
l.223-224 : If I understand the method correctly, the selected members from the DP ensemble are also initialized 5–7 months prior for the May forecast and 10–12 months prior for the November forecast. It would be interesting to see whether selecting members from the DP system initialized in November of the same year of the forecast could increase the skill in Fig. S4b.
l. 224-227 : The fact that some methods using only HIST show such poor skill suggests that the predictor used for the constraint provides no information on the evolution of El Niño. Conversely, methods with skill comparable to SP and MP in the first forecast months appear to rely on more informative predictors. Are these best methods based solely on the Niño 3.4 index? And is there a common predictor among the worst methods as well?
l. 239-240 It seems from these figures that many members are selected from two models (MIROC6 and CESM1). Do you have any thoughts on why this is the case ? Are these models better in their representation of El Niño?
Fig 6 : It would be helpful to clarify the choice of constraints for the different tests. For example, in panels 6d, e, f, is it based on HIST+DP? Similarly, for panels 6g, h, i, is it based on OBS or SP?
Fig 6 : The legend for the fifth row is unclear and quite confusing. In the legend, you describe the mean absolute error for Nino3.4 or NAO (two scores), the spatial ACC, the spatial centered-RMSE, and the spatial uncentered-RMSE with respect to TOS or PSL (is this two scores, or four if TOS and PSL are tested for both centered- and uncentered-RMSE?). However, only four distributions are highlighted in the figure to test the scores, with, for example, only one as the error index, which I assume corresponds to the mean absolute error — but is it for Nino3.4 or NAO? Is one missing?
small correction:
Fig 1 : It is hard to see the brown HIST line over the purple lines. For the legend, it would be more convenient for the reader to indicate the period over which the skill is calculated, so that this information is available directly in the legend.
Fig S3b : the dark green line is missing in the Figure legend below the x-axis.
l.210: remove the tilde over the « 1 »
Fig S6 : It would be easy for the reader to have directly _DP at the end of the models that correspond to the DP ensemble.