MAJOR

L21: Given the lower skill of the analogs (e.g. Fig.1 and 2) but that they are potentially very useful as a tool for making seamless predictions, I think the abstract should make it clear that the skill is lower rather than ‘competititve’.

L120: This presumably results in all members having the same trend? If so, this needs a little discussion in the text with pros and cons as you are losing the individual model response to forcing and replacing it with the multimodel mean. Does this also reduce the variance in the ensemble?

L146: Also on trends. The reference forecast R is stated to be a trivial climatological forecast but what does this mean? Is it a constant climatological value for each variable? Why not use a linear trend for Ts? This would seem like a fairer test.

L160: Is it fair to compare ensembles of different sizes? There is plenty of literature on this point and all scores should either be calculated for the same ensemble size or corrected for ensemble size to make them equivalent. Even if large ensembles of analog forecasts are easy to generate this is important for the comparison and understanding the relative merits of the methods.

L170-175, Fig2 and 3, L375: While I am sure readers will be open-minded to this method of forecasting this passage feels somewhat biased in favour of the analog method. The dynamical seasonal forecasts have a better correlation. This discussion needs to be rephrased and a panel of the difference in correlation scores is also needed, perhaps in place of the current panel 1b and panel 3b.

Fig.4 and Fig.11: I think it is important that these metrics are changed to the average correlation skill over land where it is significant, rather than just the area that is significant because the current metric does not reflect the higher skill of SEAS5 in many regions and this is important for the value of the forecasts.

L250: in fact all the indices are of weak amplitude (even Nino3.4) so this needs to be stated with some comments about the ability to recalibrate the amplitude.

Fig.6: The analogs are clearly more competitive on this longer timescale and the striking similarity with the dynamical model is impressive, at least with EC-EARTH. However, I am not convinced EC EARTH is the best decadal prediction system. Does this result hold for other models? Either way, I think the abstract should reflect the benefit of analogs may be greater for the longer timescales.

Fig.8e: Presumably this result comes from the fact that the analogs can be selected from any year? Does it improve if the analogues have to be selected e.g. from the same decade as the target? Or is this already accounted for by the removal and replacement of the forced trend?

MINOR

L52: ‘is meant to constitute a pool…’ of course it does not always achieve this

L55: the number is not very small as it is now over 10 on subseasonal, seasonal and decadal scales. See for example Kumar et al, 2024, BAMS. Suggest to say “limited number”

L64: ‘…of a more sophisticated’

L64: it is stated earlier that models drift to their own climatology and that this reduces skill. However L64 states that the analog method is not subject to drift because the model is in its own climate. This seems very one sided in favour of the analog approach and so it needs to be rephrased.

L70 Kushnir et al., 2019, Nat. C.C. is an important missing reference on the operationalisation of decadal predictions.

L80-85: please state the total sample size (in years), is it really greater than the decadal hindcast size?

L104: constraint

L215: there is a long literature on Sahel forecasts so please add some references here.

Fig.10: please reduce the vertical scale to better show the variability.

L340: Smith et al 2018 specifically examined the ability of GCMs to predict global temperature: Smith et al, 2018. Predicted chance that global warming will temporarily exceed 1.5C. Geophys. Res. Lett.

In the manuscript entitled “Seamless seasonal to multi-annual predictions of temperature and standardized precipitation index by constraining transient climate model simulations”, Acosta Navarro and colleagues used an analog method to provide seamless climate forecasts of temperature and SPI across seasonal to multiannual timescales. This method has the advantage of providing forecasts that are not impacted by initialization shocks or drift and that can easily be updated monthly. I found that the method proposed by the authors, as well as the evaluation carried out, is interesting and will be of interest to the readers of the journal. However, I have some major and minor issues and comments that I hope are constructive, especially regarding the way the skill of the analog method is presented, as well as some methodological aspects.

Abstract: 

l.17-18: Although the analog method generally provides better skill than the unconstrained CMIP6 ensemble mean, this is not always the case, with some regions consistently showing non-significant improvements. For example, this applies to large parts of the Northern Hemisphere continent in the seasonal prediction of surface temperature (Figs. 1 and 2e). For multiannual prediction, we can clearly see some regions where the analog method is less skillful than the unconstrained CMIP6 ensemble (Figs. 7 and 8e). Therefore, I believe the statement 'consistently outperforming the unconstrained CMIP6 ensemble' is somewhat biased and should be more nuanced.

l.20-22: The skill of the analog method is generally lower than that of seasonal prediction (e.g., Fig. 4). For multiannual prediction, the results are more mixed: the analog method appears to be slightly more effective than the EC-Earth3 prediction system for 12- and 24-month predictions but clearly shows lower skill at 48-month predictions (e.g., Fig. 11). Therefore, I believe the statement 'competitive compared to state-of-the-art initialized climate prediction systems' should again be more nuanced, emphasizing the method’s potential at annual to biennial timescales, where it seems more competitive with state-of-the-art initialized climate prediction systems. Additionally, this does not detract from the fact that this method could be a highly useful tool for seamless predictions. 

Method: 

l.93: How was the 1960–2030 period chosen? What is the added value of selecting analogs over a near-future period (i.e., selecting analogs from 2030 onward for a 2024 forecast, if I understand the method correctly)?

Table S4-5: Why use 4-month tests instead of 3-month tests, as in the other table, for the 24-month prediction?

l.105-106: Why the period used is smaller for longer prediction ?

l.114: I am a bit confused about Method 3. I understand that the authors aim to maintain a similar ensemble size while maximizing the number of models, but I am unsure how to interpret the results, especially in cases where the same member is chosen for the majority of the five analogs.

l.119-123: The method used to remove the trend may need further clarification. Are you referring to removing a linear trend from the analog-method predictions and observations (this is not specified), as indicated in Figure 5 of Smith et al. (2019) ? If so, the potential implications should be clarified, as this approach may not effectively remove the forced signal compared to using the ensemble mean for each model. Additionally, the implications of this methodological choice—where all members will have the same trend—should be discussed further.

l.121-122: Does this mean that the analog-based method failed to capture the forced response in surface temperature ? 

l.137-138: I am a bit confused here. For surface temperature predictions, you removed the external forcing, added the CMIP6 ensemble mean, and then used linear regression at each grid point to remove the CMIP6 ensemble mean in order to estimate internal variability. Why not directly estimate internal variability as the residual after removing the external forcing, before adding the CMIP6 ensemble mean?

l158: Why did you choose the EC-Earth3 prediction system, given that several centers now provide such predictions? It would be interesting to see whether the regions where the analog method performs better or worse than the EC-Earth3 prediction system remain the same for another prediction system. Additionally, I’m curious whether there is any known bias in the EC-Earth3 predictions that the analog-based method might improve.

Results: 

l.175-178: Although the spatial pattern between the analog method and the prediction system is quite similar, the prediction system seems to have an overall larger correlation. A map of the difference between both would help clarify this, perhaps instead of Fig. 1c? This also seems to be the case for Fig. 2.

l.215: “in which the skill seems to result from the external forcing.” It would be nice to add some reference to support this point. 

l.230: I would not say that the skill is 'comparable' for TAS, as there are some seasons for which the SEAS51 predictions have a global land fraction significantly more correlated with the observations, with values more than 10% higher than those for the analog method.

l.271: Are you talking about model bias that influence the analog-method or bias in the analog-method results ? If it is the first one, references would be welcome here. 

l.272: As for Fig 1 and 2, it would be nice to see the map of the difference. 

Fig 11: It is interesting to see that the analog-based method is very close to EC-Earth3, or even slightly better, in terms of the fraction of global land area that is statistically significant with the observations for 12- and 24-month predictions of residual temperature and SPI. However, for 48-month predictions, the EC-Earth3 prediction system appears to be better. Do you have any thoughts on why the analog-based method might perform worse than the EC-Earth3 prediction system for long-term predictions?

l.366-367: I think this needs a bit more clarification here or in the Fig. 11 legend to make it easier for the reader to follow the analysis. In Fig. 11, if I understand correctly, the months correspond to the predictions for each month relative to the forecast time. For example, June for TAS 12-months corresponds to the temperature prediction for the first month of June in the forecast, and June for TAS 24-months corresponds to the temperature prediction for the second month of June in the forecast, with the prediction starting in November. Does the light green line represent the same thing, but starting in May instead of November? In that case, we can expect this result, as the prediction time is shorter for the light green curve. 

l.375-376: Same comments as for the abstract. 

l.403-405: This is a strong added value of the analog method, I think it should be emphasized more

l.408-416: I would just add a point to remind that the analog-method does not induce any drift due to the shock of the initialization, which is also an important added value of the method. 

Small correction: 

l.99: The reference is not complete

l265-266: “The residual correlation of the analog forecasts is illustrated in Figure 6b.” → This information is already in the legend, so I'm not sure how useful it is here. 

l302: I think it is Fig 8 instead of Fig 9