A Methodological Comparison of Forecasting Models Using KZ Decomposition and Walk-Forward Validation
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsComments:
1. The additive framework used to recombine components could benefit from a rationale, but why not multiplicative or other nonlinear combinations?
2. The manuscript does not mention how missing values were handled. This is crucial for time series integrity.
3. Some references lack full citation details. In some references, DOI links are added, whereas it is not added in others.
4. The manuscript mentions early warning systems. Could the model be adapted for real-time deployment?
5. Discuss whether the framework could be applied to other environmental variables (e.g., PM2.5, rainfall).
Author Response
We attached a pdf file.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for Authors1. Abstract:
- Does not mention limitations or limitations of the approach used (e.g., sensitivity to KZ parameters, reliance on long historical data for walk-forward validation).
2. Introduction:
- While mentioning previous research, it may not explicitly emphasize the specific research gaps that this study fills (e.g., direct comparison between many classical and ML models on the components of the KZ decomposition results for T2M with walk-forward validation).
- Does not explicitly state why T2M was chosen as the target variable over other climate variables in the context of multiscale challenges.
3. Methods:
- The selection of KZ parameters (m=365, k=3) for all variables and components may not be universally optimal; it is not clear whether these parameters were extensively tested or validated for all different variables and components.
- Walk-forward validation was used, but there is no mention of confidence interval or prediction analysis, which are important for assessing model uncertainty.
- Only one geographic location and time span are used, which may limit the generalizability of the results.
4. Results:
- The results show that the short-term component is very difficult to predict with standard models (low R²), suggesting that the decomposition approach may be less effective for portions of the data that are inherently random or heavily influenced by unmeasured short-term processes.
- For the short-term component, univariate models (LSTM, AutoReg) performed better than multivariate models using external explanatory variables. This may indicate that the explanatory variables used were less relevant or not powerful enough to capture short-term dynamics.
5. Conclusions:
- The conclusions generally state the superiority of the decomposition approach and ensemble models, but perhaps underemphasize the implications of the models' failure to predict the short-term component.
Author Response
We attached a pdf file.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors used the Kolmogorov-Zurbenko (KZ) filter to decompose the two predictor variables and the target variable into short-term, seasonal, and long-term components. Each component was independently modeled using three classic regression methods (linear regression, ridge regression, and Lasso regression) and two machine learning algorithms (random forest and XGBoost). The predicted components were then recombined using an additive framework, and the results validated the effectiveness of this method.
The number of references is too small to effectively illustrate the application value and context of this research.
The authors need to increase the number of strictly relevant references and also include references covering the past five years, particularly those covering 2024 and 2025.
Why did the authors only use data collected through December 31, 2020, and not data from 2021 to 2025? Does this have a necessary impact on the results?
Why did the authors choose a window size of m = 365 and a number of iterations of k = 3 when using the KZ filter? What was the rationale behind these choices? How do the specific values ​​of the window size and number of iterations affect the results?
What is the process for standardizing variables? Are there missing values ​​or outliers? How are they identified and corrected?
The paper's language needs professional assistance to improve readability. The comma is followed by a capitalized word.
For the results corresponding to Model 2, the authors should conduct a residual analysis of the residual terms, including analysis of their mean, variance, identical distribution, independence, and whether they meet the original figure's definition and Gaussian assumptions.
The innovation of this study is not prominent and clear. The authors should organize and expand on the original content and present it point by point.
For the results corresponding to Model 6, the authors should also conduct a residual analysis of the residual terms, including analysis of their mean, variance, identical distribution, independence, and whether they meet the original figure's definition and Gaussian assumptions.
Comments on the Quality of English LanguageThe paper's language needs professional assistance to improve readability. The comma is followed by a capitalized word.
Author Response
We attached a pdf file.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsManuscript can be accepted.
Author Response
The response is provided in the attached PDF file.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsSlight improvement in lines 53-64, the author should add several sentences about AI. the following recent references can be added:
- https://journal.unesa.ac.id/index.php/vubeta/article/view/40135
- https://journal.unesa.ac.id/index.php/vubeta/article/view/39360
Author Response
The response is provided in the attached PDF file.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for Authorsaccept
Comments on the Quality of English Languageaccept
Author Response
The response is provided in the attached PDF file.
Author Response File: Author Response.pdf
