Response to Academic Editor

Please ensure that your manuscript meets PLOS ONE’s style requirements, including those for file naming.

Thank you for this reminder. We have revised the manuscript and associated submission files to comply with PLOS ONE style requirements, including formatting, structure, figure/table presentation, and file naming conventions. We have carefully checked the revised submission to ensure consistency with the journal guidelines.

Please ensure that your code is shared in accordance with PLOS ONE guidelines and is available without restrictions upon publication.

Thank you for this important requirement. We have ensured that our code sharing fully complies with PLOS ONE’s guidelines for transparency and reproducibility.

All author-generated code has been made publicly available without restriction via an archived repository with a DOI:

Code repository: https://doi.org/10.5281/zenodo.19589534

The repository includes:

Full training scripts for all models (e.g., SARIMAX, TFT, and deep learning models)

Complete preprocessing pipeline

Evaluation and analysis scripts

Hyperparameter configurations

Environment specifications, including package versions and Python setup

To address data sharing constraints, meteorological data obtained from the Bangladesh Meteorological Department cannot be publicly distributed due to access restrictions. To ensure reproducibility despite this limitation, we have provided a synthetic dataset that preserves the structure and statistical characteristics of the original data, allowing all code components to be executed and validated. The shared materials therefore enable full reproducibility of the workflow, including preprocessing, model training, and evaluation, in accordance with best practices.

Please update your submission to use the PLOS LaTeX template.

Thank you for this instruction. We have updated the manuscript using the PLOS LaTeX template and revised the submission files accordingly to align with the journal’s LaTeX formatting requirements.

The repository noted in the Data Availability statement does not qualify as an acceptable repository. Please upload the minimal dataset necessary to replicate the study to a stable public repository.

Thank you for highlighting this requirement. We have revised our data sharing approach to comply with PLOS ONE’s data availability standards. We have deposited the materials in a public, stable repository with a DOI:

Repository: https://doi.org/10.5281/zenodo.19589534

Due to access restrictions, the meteorological data obtained from the Bangladesh Meteorological Department cannot be publicly shared. However, to ensure reproducibility, we have provided a synthetic dataset that preserves the structure and statistical properties of the original data, along with all necessary preprocessing and modeling scripts.

This dataset, together with the shared code and documentation, enables replication of the full analytical workflow, including preprocessing, model training, and evaluation.

Please revise the data availability statement and clarify your data sharing plan, noting that all data must be freely accessible upon publication unless an exemption applies.

Thank you for this clarification. We have revised our Data Availability Statement to provide a clear and complete data sharing plan in line with PLOS ONE requirements.

All code, documentation, and a reproducible dataset have already been made publicly available in a stable repository with a DOI:

Repository: https://doi.org/10.5281/zenodo.19589534

Due to third-party restrictions, the meteorological data obtained from the Bangladesh Meteorological Department cannot be made publicly available, as redistribution is not permitted under the data provider’s terms.

To ensure reproducibility, we have provided a synthetic dataset that preserves the structure and statistical properties of the original data, along with full preprocessing, modeling, and evaluation code. This allows the complete workflow to be replicated without access to the restricted data.

We respectfully request that this be considered as a valid case for restricted data exemption, and we have updated the Data Availability Statement accordingly to explain these constraints.

Please amend either the abstract on the online submission form or the abstract in the manuscript so that they are identical.

Thank you for noting this discrepancy. We have revised the abstract so that the version in the manuscript and the version entered in the online submission system are now identical.

Figure 1 may contain copyrighted map images. Please provide permission or replace the figure with CC BY–compatible content.

Thank you for raising this important point. We confirm that Figure 1 does not use proprietary sources such as Google Maps or other restricted platforms.

The map is based on administrative boundary data obtained from the Humanitarian Data Exchange (Bangladesh Subnational Administrative Boundaries dataset), which is derived from open and publicly available sources. We have verified that the dataset is compatible with open-use requirements.

To ensure full compliance with PLOS ONE’s CC BY 4.0 licensing policy, we have:

Clearly cited the data source in the figure caption

Updated the caption to reflect that the map is created from open administrative boundary data

Confirmed that no copyrighted or proprietary basemaps are used

The revised figure and caption now comply with the journal’s licensing requirements.

Please ensure that you refer to Figures 2, 4 and 11 in your text as, if accepted, production will need this reference to link the reader to the figure..

Thank you for this observation. Following revisions to the analysis and figures, we have updated the manuscript to ensure that all figures are correctly numbered and consistently referenced in the text.

Specifically:

Figures previously labeled as 2 and 4 have been renumbered and updated, and are now properly cited in the revised manuscript

Figure 11 has been removed as part of the revised analysis

All figure references have been carefully checked to ensure consistency and correct linkage throughout the manuscript.

We note you have included tables that were not properly referred to in the text. Please ensure that Tables 1-4 are referred to in the manuscript.

Thank you for this observation. We have carefully revised the manuscript to ensure that all tables retained in the main text are explicitly referred to and discussed in the appropriate places.

In addition, to improve clarity and reduce length, we reorganized the tabular material during revision. Non-essential descriptive tables were moved to the Supporting Information or summarized in the text, while only the most relevant tables were retained in the main manuscript. All remaining tables are now consistently numbered and properly cited in the revised manuscript.

If the reviewer comments include a recommendation to cite specific previously published works, please evaluate whether they are relevant and should be cited.

Thank you for this clarification. We carefully reviewed the publications recommended by the reviewers and evaluated their relevance to the scope, methodology, and objectives of our study. Where appropriate, we have incorporated relevant citations into the revised manuscript. For works that were not directly aligned with the present study, we did not add citations.

Response to Reviewer #1

Thank you for handling and reading our manuscript. We have addressed all of your comments as follows:

##Minor Comments

Please clarify how the “dominant” label is determined for each year by defining what “type” refers to and stating the exact rule used to select the dominant category (e.g., the category with the highest value or frequency), with a brief example if possible.

Thank you for this suggestion. We have clarified the definition of the dominant serotype variable in the revised manuscript. Here, “type” refers to the dengue serotype category (DENV-1 to DENV-4). For each year, the dominant label was defined as the serotype with the highest annual value/frequency in the compiled serotype data, and that label was assigned to the monthly observations for that year. For example, if DENV-2 was the most frequent serotype in a given year, the dominant label for that year was recorded as DENV-2. We have added this explanation to the manuscript to make the construction of the variable explicit and reproducible.

There are inconsistencies in formatting, quotation usage, and punctuation around terms and names. Please correct this type of typographical inconsistency consistently throughout the manuscript.

Thank you for this careful observation. We have performed a thorough proofreading of the revised manuscript and corrected typographical inconsistencies throughout. In particular, we standardized formatting, punctuation, capitalization, quotation usage, and the presentation of terms, names, and figure/table references to improve consistency and readability across the manuscript.

In Table 2, please add adequate spacing between the Unit and Domain-Based Class columns to improve readability and prevent the entries from appearing merged.

Thank you for this suggestion. To improve readability and overall manuscript clarity, Table 2 has been moved to the Supporting Information as part of the revised manuscript structure. In the supplementary version, we have also improved the formatting and spacing, including clearer separation between the Unit and Domain-Based Class columns.

Please revise punctuation for figure citations and ensure consistent use of parentheses (e.g., write “…with excess zeros (Fig. 3b)” rather than leaving the figure reference outside the sentence structure).

Thank you for pointing this out. We have revised the manuscript to ensure that figure citations are punctuated consistently and integrated properly within sentence structure, including the consistent use of parentheses in the PLOS style (e.g., “(Fig. X)”). All figure references were reviewed and corrected throughout the manuscript for consistency.

Please include the missing figure numbers and cite them consistently in parentheses throughout the manuscript.

Thank you for this observation. We have carefully revised the manuscript to resolve the missing figure numbers and to ensure that all figures are now cited consistently in parentheses throughout the text. We also checked figure numbering globally after revision so that all in-text references match the final figure order.

Please add proper spacing for consistency, e.g., AvgTemp_lag_3 (r = 0.372).

Thank you for this suggestion. We have revised the manuscript to improve formatting and readability by standardizing the presentation of variable names and statistical notation. For example, terms such as AvgTemp_lag_3 (r = 0.372) have been reformatted to a clearer style (e.g., Avg. Temp (lag 3)), and spacing has been made consistent throughout the manuscript.

Use a semicolon when listing definitions so each item is clearly separated (e.g.,$\tau$: forecast horizon (steps ahead), with $\tau\in\{1,\ldots,6\}$.).

Thank you for this suggestion. We revised the presentation of notation and symbol definitions to improve clarity and consistency. In the revised manuscript, notation is now presented more clearly in the nomenclature table and related mathematical descriptions, so that individual definitions are separated and easier to read. We also standardized punctuation and formatting in notation-related text throughout the manuscript.

Quotation marks are used inconsistently; please replace them with proper LaTeX quotes (e.g., ``Normal'' instead of "Normal") throughout the manuscript.

Thank you for this observation. We have revised the manuscript to standardize quotation mark usage throughout and replaced inconsistent quotation formatting with proper LaTeX quotation style where applicable. This was checked across the full manuscript to improve consistency and presentation.

##Major Comments:

In the figures, key text elements (e.g., axis tick labels, legend entries, and titles) are too small and difficult to read. Please increase the font sizes across all plots to ensure clear visibility and consistent readability.

Thank you for this important observation. We have revised all figures in the manuscript to improve readability by increasing font sizes for axis labels, tick labels, legends, and titles. We also ensured consistent formatting across all plots so that text elements are clearly visible both on screen and in print.

Figures 5 and 6 are currently presented in grayscale; using distinct colors would make the plotted elements easier to distinguish and improve overall readability.

Thank you for this suggestion. We revised the figure set to improve visual clarity and distinguishability. In the revised manuscript, the relevant figures were updated to use distinct colors instead of grayscale where appropriate. Because the figure set was reorganized during revision, the original numbering also changed: the figure previously cited as Figure 5 was renumbered, and the original Figure 6 was removed as part of the revised analysis. More broadly, we applied improved color formatting across the figures to enhance readability throughout the manuscript.

Add uncertainty and calibration results for the quantile-based TFT using coverage and pinball loss, ideally reported separately for Normal and Outbreak periods.

Thank you for this valuable suggestion. We have expanded the revised manuscript to include an explicit uncertainty and calibration analysis for the quantile-based TFT model. Specifically, we now report interval coverage, mean interval width, and quantile pinball loss, and we present these results separately for Normal and Outbreak periods across forecast horizons. This addition is included in the revised Results section(Uncertainty and calibration analysis) and provides a clearer assessment of TFT’s probabilistic performance and calibration under both routine and epidemic conditions. (Table 8: TFT regime-wise 80% interval coverage, mean interval width, and pinball loss by forecast horizon.)

Using a time-series aware significance test, such as a block bootstrap or rolling-window paired tests, would better account for autocorrelation in forecast errors.

Thank you for this thoughtful suggestion. We agree that statistical comparison of forecasting models should account for temporal dependence in forecast errors. In the revised manuscript, we addressed this concern by adopting a time-series-aware Diebold-Mariano testing framework with the Harvey-Leybourne-Newbold small-sample correction and Newey-West long-run variance estimation, with lag selected as a function of forecast horizon, presented in the result section(Pairwise significance testing); Table 10. Pairwise Diebold–Mariano significance testing comparing Prophet with shortlisted competing models across horizons.

We therefore revised the significance-analysis section to use a temporally appropriate comparison procedure, even though we did not implement the specific alternatives suggested by the reviewer (e.g., block bootstrap or rolling-window paired tests). This updated procedure is now described in the manuscript and the corresponding results are reported in the revised Results section.

Presenting feature-group ablation experiments (climate-only, serotype-only, climate plus serotype, then adding demography and case-lags) would clarify which inputs drive performance at each horizon.

Thank you for this constructive suggestion. We agree that feature-group comparisons can help clarify which inputs drive forecasting performance across horizons. In the revised manuscript, we addressed this issue through a feature-group sensitivity analysis based on one-factor-at-a-time ablation, using the same preprocessed panel, split structure, and horizon definitions as in the main experiments. Specifically, we compared the retained full specification against variants with climate variables, serotype variables, temporal variables, or population density removed, depending on model class. This allowed us to assess how the

Attachments

Attachment

Submitted filename: Response to Reviewers.pdf

Response Letter

Manuscript ID PONE-D-26-03317

Tactical vs. strategic: An adaptable framework for horizon-dependent dengue forecasting using data-driven approaches with serotype and climate covariates in Bangladesh

PLOS ONE

Response to Academic Editor

Thank you for the opportunity to revise our manuscript. We have carefully addressed the remaining comments and prepared a revised manuscript, a marked-up version, and a clean version for submission. We also reviewed the reference list, added the relevant recommended studies where appropriate, and checked the manuscript for clarity, consistency, and moderated claims. The specific changes are described point-by-point below.

Response to Reviewer #1

Thank you for handling and reading our manuscript. We have addressed all of your comments as follows:

1. Please clarify how the “dominant” label is determined for each year by defining what “type” refers to and stating the exact rule used to select the dominant category (e.g., the category with the highest value or frequency), with a brief example if possible.

Thank you for pointing this out. We have added an explanation in the Sensitivity and Uncertainty Analysis section. We now clarify that the zero outbreak-period coverage in Table 8 means that none of the observed outbreak counts fell within TFT’s predicted 80% intervals. The model still produced prediction intervals, but these intervals were systematically below the observed outbreak magnitudes, indicating poor calibration and underestimation of the upper tail during rare epidemic surges.

2. It is recommended to cite the following studies:

• https://doi.org/10.1371/journal.pone.0342764

• https://doi.org/10.26855/abr.2022.12.001

• https://doi.org/10.1038/s41598-025-28135-x

• https://doi.org/10.4236/am.2022.1310053

Thank you for the recommendation. We have cited the suggested studies in the Introduction to strengthen the manuscript’s positioning within the broader infectious-disease modelling literature. These studies are used as methodological background for predictive, stochastic, and mathematical modelling approaches relevant to epidemic assessment, resource allocation, and intervention planning.

Response to Reviewer #2

Thank you for handling and reading our manuscript. We have addressed all of your comments as follows:

1. The manuscript is now generally well structured, but the claim of a “generalizable framework” in the title and discussion should still be stated a little more cautiously. The framework is promising and well-motivated, but the present study demonstrates it on one country case study with one specific panel structure and one epidemic context. It may be better to describe it as a transferable or adaptable framework unless broader validation is provided. This point is most relevant in the Title, Methodological Framework & Case Study Setup, Discussion, and Conclusion.

Thank you for this helpful suggestion. We agree that “generalizable framework” overstated the scope of validation, as the present study is demonstrated using one country-level case study, one district-panel structure, and one epidemic context. We have therefore revised the title and relevant wording throughout the manuscript to describe the approach as an “adaptable framework” rather than a “generalizable framework.” We also added cautious wording in the Methodological Framework, Discussion, and Conclusion to clarify that broader transferability would require external validation in other geographical and epidemiological settings.

2. The treatment of serotype information is interesting and potentially novel, but the manuscript would benefit from slightly clearer explanation of how the annual “dominant serotype” variable was constructed and how this annual information aligns with monthly forecasting targets. Since the final selected feature set includes lagged serotype variables and the discussion emphasizes serotype importance, a short clarification in Dataset Development for Comprehensive Analysis and Feature Engineering would improve reproducibility.

Thank you for this helpful comment. We have clarified how the annual serotype variables were aligned with monthly forecasting targets. Specifically, we now state that annual DENV-I--DENV-IV percentage values were assigned to all monthly observations within the corresponding year, and that the dominant serotype variable was defined as the serotype with the highest annual percentage value for that year. We also clarified in the Feature Engineering section that lagged serotype variables were generated after this annual-to-monthly alignment. This revision improves reproducibility and clarifies how annual serotype surveillance information was incorporated into the monthly forecasting framework.

3. The treatment of serotype information is interesting and potentially novel, but the manuscript would benefit from slightly clearer explanation of how the annual “dominant serotype” variable was constructed and how this annual information aligns with monthly forecasting targets. Since the final selected feature set includes lagged serotype variables and the discussion emphasizes serotype importance, a short clarification in Dataset Development for Comprehensive Analysis and Feature Engineering would improve reproducibility.

Thank you for this helpful suggestion. We have revised the Feature Selection section to clarify the final manual selection step. We now explain that SelectKBest was used as a screening step, and that final predictors were selected from the screened set by prioritizing target relevance, low redundancy after correlation and VIF filtering, temporal availability at the source forecasting month, and representation of key epidemiological, climatic, temporal, and demographic domains

4. The evaluation design is one of the strong parts of the paper, especially the separation of overall, per-horizon, regime-wise, and outbreak-detection performance. However, because the outbreak thresholds are model-specific in Table 9, the comparison of alert metrics across models becomes a little harder to interpret directly. I suggest adding one sentence in the Evaluation Protocol and Practical Implications: From Magnitude Failure to Actionable Alerts sections to explain this trade-off more explicitly.

Thank you for this important observation. We have added clarifying sentences in both the Evaluation Protocol and Practical Implications sections. We now explain that the model-specific outbreak thresholds support within-model operational calibration, but make direct cross-model comparison of threshold-dependent metrics such as precision, recall, and F1 less straightforward. We also note that threshold-independent metrics such as ROC-AUC should therefore be interpreted alongside the alert metrics.

5. The statistical-significance analysis is useful and much clearer now. Still, the manuscript should make sure the reader understands that the differences between Prophet and the shortlisted competitors are generally modest, with significance appearing mainly for one-step absolute error comparisons. This is already partly stated, but a slightly more explicit caution in the Pairwise significance testing section and Discussion would help avoid over-interpretation.

Thank you for this helpful suggestion. We have added more explicit caution in the Pairwise significance testing section and Discussion. The revised text now clarifies that Prophet’s advantage over the shortlisted competitors was generally modest, and that statistically significant separation appeared mainly for one-step absolute-error comparisons rather than consistently across horizons or loss functions.

6. The sensitivity-analysis results are interesting, especially the finding that removal of serotype variables improved some models and that a simpler non-seasonal SARIMAX outperformed the main seasonal specification. These findings are important and deserve a bit more interpretation in the Sensitivity and Uncertainty Analysis section and Discussion, because they suggest that some feature groups or baseline assumptions may not contribute uniformly across architectures.

Thank you for this helpful suggestion. We have added further interpretation in both the Sensitivity and Uncertainty Analysis section and the Discussion. We now clarify that the contribution of serotype variables was architecture-dependent, as their removal slightly improved MLR and more clearly improved attention-LSTM. We also discuss that the simpler non-seasonal SARIMAX specification outperformed the seasonal baseline, suggesting that fixed seasonal ARIMA terms may be restrictive when outbreak timing shifts across years. These revisions emphasize that feature groups and baseline assumptions should be evaluated by model architecture and forecast objective rather than assumed to improve all models uniformly.

7. The interpretability and causal-analysis sections add value to the manuscript, but the distinction between predictive usefulness and causal importance should be stated very carefully. The current text is already moving in the right direction, especially when discussing serotype centrality versus forecasting gain, and I encourage the authors to preserve this cautious wording. A small amount of polishing in Model Interpretability and Feature Importance and Causal Inference Analysis would further strengthen the paper.

Thank you for this helpful comment. We have polished the Model Interpretability and Feature Importance and Causal Inference Analysis sections to distinguish predictive usefulness from causal importance more carefully. We now clarify that interpretability results describe model reliance and predictive contribution rather than causal effects, and that PCMCI network centrality reflects statistical dependency within the fitted causal-discovery framework rather than definitive biological causation or guaranteed forecasting improvement.

8. The discussion is thoughtful and one of the stronger parts of the manuscript. The tactical-versus strategic framing is useful and helps explain why different models perform best for different objectives. I would only suggest some light language polishing in the Discussion and Conclusion, especially in places where the phrasing is slightly awkward or grammatically rough.

Thank you for this positive and helpful comment. We have lightly polished the Discussion and Conclusion to improve clarity, grammar, and flow while preserving the tactical-versus-strategic framing. We revised awkward phrasing, clarified model-specific interpretation, and made the limitation regarding external validation more precise.

9. The authors may consider citing a few recent related works to further strengthen the positioning of the manuscript within the current forecasting and mathematical-modeling literature, for example:

• https://doi.org/10.1016/j.cnsns.2025.109246

• https://doi.org/10.3329/ganit.v43i1.67858

• https://doi.org/10.48550/arXiv.2604.18642

• https://doi.org/10.1016/j.aej.2025.12.039

where relevant to the revised discussion and methodological context.

Thank you for this useful suggestion. We have added the suggested recent studies where relevant to strengthen the manuscript’s positioning within the current forecasting and mathematical-modeling literature. The Bangladesh-focused dengue forecasting study was cited in the methodological context, the broader infectious-disease modelling studies were cited in the modelling motivation, and the dengue control/decision-support study was cited in the Discussion to distinguish forecasting from intervention-focused modelling.

10. Overall, this is a solid and improved manuscript with a relevant case study, a well-motivated framework, and a practically useful evaluation perspective. The remaining issues are mostly about wording, clarification, and slight moderation of some claims.

Response to Overall Comment:

Thank you for the positive assessment of the revised manuscript. We appreciate the reviewer’s recognition of the case study, framework, and evaluation perspective. In response to the remaining comments, we have made targeted revisions to clarify the methodology, moderate generalization claims, improve interpretation of sensitivity and significance results, and polish the wording in the Discussion and Conclusion.

Attachments

Attachment

Submitted filename: Response_to_Reviewers_auresp_2.pdf

Response to Academic Editor

Thank you for the opportunity to revise our manuscript. We have carefully addressed the comments and prepared a revised manuscript, a marked-up version, and a clean version for submission.

Response to Reviewer #1

Thank you for handling and reading our manuscript. We have addressed all of your comments as follows:

1. Could you please recheck Comment 1 carefully? It appears that this comment was not part of the current revision round, and there may have been a mismatch in the included comment. Please address the appropriate comment accordingly.

Thank you for your meticulous observation. Actually, it was mistakenly placed. However, the question and the answer is included below.

‘‘The interval coverage during the outbreak period appears to be consistently zero across all forecast horizons (Table 8). This requires an explanation.’’

Thank you for pointing this out. We have added an explanation in the Sensitivity and Uncertainty Analysis section. We now clarify that the zero outbreak-period coverage in Table 8 means that none of the observed outbreak counts fell within TFT’s predicted 80% intervals. The model still produced prediction intervals, but these intervals were below the observed outbreak magnitudes and shallow learning from dominant low dengue count period.

Attachments

Attachment

Submitted filename: Response_to_Reviewers_auresp_3.pdf