acceptodds
Under review as a conference paper at ICLR 2027

EvoFIF: Evolutionary Discovery of Inductive Biases for Forecast-Oriented Time Series Imputation

Abstract

Time series imputation is widely used as a preprocessing step for forecasting, yet minimizing imputation error does not necessarily improve forecasting performance. Forecast-oriented time series imputation (FOTSI) therefore optimizes imputation for the downstream task, and existing methods are model-centric, since each of them fits a model to one dataset, missingness pattern or forecasting model and has to be retrained whenever the scenario changes. We propose EvoFIF (Evolutionary Forecast-oriented Imputation Framework), which instead discovers reusable configurations of imputation inductive biases by evolutionary search. EvoFIF extracts an initial pool of explicit biases from classical imputation algorithms with a large language model, evolves it by utility-weighted Thompson Sampling with LLM-based mutation and crossover, and verifies every candidate by downstream forecasting utility, objective alignment and computational cost. On nine multivariate datasets, four missingness patterns and four forecasting models, EvoFIF attains the lowest downstream loss, imputes a scenario orders of magnitude faster than every baseline, and its reusability dominates that of the existing frameworks, since one configuration transfers across datasets, patterns, forecasters and horizons without retraining.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series data, Time series imputation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8CCEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes EvoFIF, a framework for forecast-oriented time series imputation (FOTSI) that shifts the optimization object from a scenario-specific model to a reusable configuration of imputation inductive biases. It extracts an initial pool of explicit biases from seven classical imputation algorithms with an LLM, evolves the pool with utility-weighted Thompson Sampling and LLM-based mutation and crossover, and verifies candidates by downstream forecasting loss, an objective-alignment penalty, and computational cost. The paper formalizes FOTSI and claims that, if utility requires cross-scenario reuse without retraining, the FOTSI problem is equivalent to searching over bias configurations. Experiments use nine multivariate datasets, four missingness patterns, four forecasters, and seven imputation baselines. EvoFIF attains the lowest downstream loss on 21 of 36 dataset-pattern combinations, ranks in the top two on 33, imputes orders of magnitude faster than the baselines, and keeps one configuration across unseen forecasters and horizons without retraining.

Critical strengths
  1. The empirical scope is substantial. Nine datasets, four missingness patterns, four forecasters, and seven recent baselines. The result that EvoFIF is best on 21 of 36 combinations and top-two on 33 is a meaningful signal.
  2. The ablations are informative. Both the initial pool and the utility-weighted sampling carry most of the gain, and the utility terms contribute a smaller but consistent improvement. This supports the design choices.
  3. The case study adds interpretability. Reading the discovered configuration line by line and tracing it back to the initial pool is a feature that model-centric methods cannot offer.
Critical weaknesses
  1. No FOTSI baseline is compared. The paper's central claim is that EvoFIF is better than existing FOTSI frameworks, but no FOTSI method is in the comparison.
  2. The "across datasets" transfer claim is unsupported. §4.4 states that the held-out axes are the forecaster and the horizon, and it does not list the dataset. The nine datasets in the main comparison were likely all used in the search. The abstract and §5 nevertheless claim transfer across datasets, patterns, forecasters, and horizons.
  3. The search cost is not reported. The efficiency comparison excludes the one-off search. The number of LLM calls, tokens, and wall-clock time are not given. Without them, the amortization argument is incomplete and the efficiency claim is only about inference.
  4. The w/o Utility ablation is underspecified. Equation 9 removes the downstream-utility and cost terms. What remains is the conflict penalty. It is unclear what the variant optimizes. If it falls back to reconstruction error, the variant is no longer forecast-oriented, and the result that it is the least damaging variant needs explanation.
  5. The Table 2 outlier needs explanation. HyFAD on DLinear is 5.645 against 0.388 for EvoFIF, an order of magnitude above every other entry. The text does not discuss whether this is a method failure, a tuning issue, or a scaling artifact. A baseline that fails this badly weakens the comparison.
Questions for authors

see weakness

3 comments

Thank you for the careful reading and for the constructive report. We answer the five points below.

To W1, we agree that a FOTSI method should be in the comparison, and the difficulty is that none of them runs in our setting without becoming a different method: they target variable subset forecasting under a fixed forecaster and fit their imputation module jointly with it, so the downstream loss has to be back-propagated into the imputer, while our evaluation spans four forecasters, two of which admit no gradient for that joint fit, and every scenario would need its own fit.

What we add instead is forecasting-guided selection (FGS), which optimizes the same quantity as EvoFIF without any search. Its candidate pool is the seven classical algorithms that seed our initial pool, each imputes the series of a scenario, the imputed series pass through the same forecaster, and the algorithm with the lowest forecasting MSE on the validation window is kept, the test window never entering the selection. We run it once for all 36 settings (FGS-global), which keeps the reusability EvoFIF has, and once per missingness pattern (FGS-pattern), which is the best a single classical algorithm can provide.

Table R1. Forecasting-guided selection against EvoFIF over the 36 settings.

ReferenceAlgorithm selectedMean L_predSettings with EvoFIF lowerMean gapImputation time (s)
FGS-globalTRMF, all settings0.799134 / 369.6%12.615
FGS-patternforward fill, KNN, TRMF0.770732 / 363.1%5.373
EvoFIFdiscovered configuration0.7508----0.085

EvoFIF attains the lower loss in 34 of 36 settings against the reusability-preserving reference and in 32 of 36 against the per-pattern one, with the widest margin under MNAR and the narrowest under BLACKOUT, and it imputes a scenario 63 to 148 times faster than the algorithm the reference selects. Since FGS-pattern already takes the best of the seven by the downstream loss for each pattern, what EvoFIF adds over the classical methods is the composition and the evolution of inductive biases rather than the choice of one of them. Section 4.1 defines the reference and Appendix F.5 reports every setting.

To W2, you are right that the nine datasets take part in the search, and that is a deliberate design rather than an oversight. They span hourly and minute-level transformer loads, sensor streams of 321 and 862 channels, meteorological and financial series, and a spiky count series of 966 steps, so the configuration is selected against different sampling frequencies, channel counts and statistical structures instead of one distribution. What the evaluation holds out is the forecaster and the horizon, as Section 4.4 states, and neither of them takes part in the search. The abstract and Section 5 state the result as reuse across scenarios, one configuration applied to every dataset, pattern, forecaster and horizon of the evaluation without retraining, and since the discovered object carries no fitted parameter, extending it beyond this coverage needs no retraining of an imputation model as long as the new series falls within what the bias library can express and combine, the condition Section 5 already states.

To W3, the cost of the search is now reported in Appendix F.3, and Section 4.2 points to it.

Table R2. Cost of the search behind the reported configuration.

ItemValue
LLM modeldeepseek-flash
Wall clock10 h 25 min 08 s, including the fitness evaluation
Input tokens0.8 M
Output tokens4.5 M
CostCNY 32, about USD 4.5

The search is one-off and returns a configuration rather than a fitted model, so none of these numbers is paid again when the configuration meets a dataset, forecaster or horizon it was not selected on. Deployment imputes a scenario in 0.085 s with no training, against 23 to 221 s for a single baseline imputation, and amortised over the 36 settings the search is about 17 minutes and CNY 0.9 per scenario. The efficiency claim therefore survives the amortisation, and the marginal cost of a further scenario stays at zero training plus 0.085 s.

To W4, the variant does fall back to reconstruction error, and Section 4.3 now says so in one sentence. It removes the downstream prediction term and the conflict penalty from the verification and keeps the cost term, so a candidate is ranked by its imputation error and the forecaster is never called during the search; what it searches is a purely reconstruction-oriented configuration over the same pool.

The relatively smaller average degradation of w/o Utility suggests that the initial pool and the evolutionary search already provide a reasonably strong starting point. Nevertheless, its substantial degradation under MCAR indicates that reconstruction-oriented selection cannot consistently preserve downstream forecasting utility. This highlights the role of the downstream objective in resolving cases where imputation accuracy and forecasting performance are misaligned. The distribution supports this reading, since only two of the 36 settings exceed 50 percent, weather and ETTm1 under MCAR, against four for each of the other two variants, and those two are precisely the settings in which a more accurate imputation and a worse forecast disagree most.

To W5, the HyFAD entry under DLinear is a method failure on one series rather than a tuning artefact. Table R4 gives the column per dataset.

Table R4. Per-dataset downstream loss under DLinear, averaged over the four missingness patterns. ILI is national_illness.

MethodETTh1ETTh2ETTm1ETTm2Elec.Exch.ILITrafficWeatherMeanMean w/o ILI
EvoFIF0.3110.1570.2940.0670.3820.1491.7810.2900.0620.3880.214
SAITS0.3580.1670.3590.0930.3970.1512.0600.2980.0820.4410.238
FIM0.4460.1960.4310.1420.5270.1862.2460.3680.1110.5170.301
HyFAD0.4510.2110.4840.1600.5260.34248.1380.3710.1225.6450.333

National illness has 966 steps over 7 variables, the shortest and sparsest benchmark of the nine, and every method loses an order of magnitude more on it than on the others, from 1.781 for EvoFIF to 48.138 for HyFAD. A diffusion generator needs a long history to constrain the amplitude of its samples, so it is the method most exposed to a series this short, and the same weakness appears elsewhere, since HyFAD is the worst method on eight of the nine datasets. All methods, EvoFIF included, were run with their official defaults on the same GPU and the same budget, with no per-method hyperparameter search, so there is no setting tuned in our favour, and dropping national illness from the column leaves the ranking unchanged, 0.214 for EvoFIF, 0.238 for SAITS, 0.301 for FIM and 0.333 for HyFAD. Section 4.4 now explains this in the body of the paper.

The revised PDF marks these changes in the margin, labelled 8CCE.W1 to 8CCE.W5.

We would like to gently follow up on our revised manuscript. We sincerely appreciate your insightful feedback on the same-objective comparison, the scope of the reuse claim, the cost of the search, the w/o Utility ablation and the HyFAD outlier, all of which motivated additional experiments and a substantial revision of the paper.

Regarding your main concern, we have completed the comparison you asked for, under the same forecasting objective:

  • Direct comparison: forecasting-guided selection applies each of the seven classical algorithms that seed our initial pool and keeps the one with the lowest forecasting error on the validation window, run once for all settings and once per missingness pattern.
  • Results: EvoFIF attains the lower loss in 34 of the 36 settings against the first reference and in 32 against the second, by 9.6 and 3.1 percent, in 0.085 s against 12.6 s and 5.4 s. A single strong algorithm therefore does not reproduce the gain, which comes from evolving and composing the inductive biases.

Beyond your main concern, we have also made the following revisions:

  • Scope of the reuse claim: all nine datasets take part in the search on purpose, since they span minute-level loads, sensor streams of 321 and 862 channels and a 966-step count series, so the configuration is selected against heterogeneous distributions rather than one of them. What is held out, the forecaster and the horizon, is then served without retraining.
  • Cost: the whole search took 10 h 25 min 08 s, 0.8M and 4.5M tokens and CNY 32, about 17 minutes and CNY 0.9 per scenario, paid once, while deployment needs 0.085 s where a single baseline imputation needs 23 to 221 s.
  • w/o Utility: the variant falls back to reconstruction and degrades 28.5 percent under MCAR against 4.9, 2.6 and 10.2 elsewhere, with only 2 of the 36 settings above 50 percent, so the downstream objective matters exactly where imputation accuracy and forecasting disagree.
  • HyFAD outlier: the 5.645 of the DLinear column comes from one series, national_illness, 48.1 against 0.333 without it, and that method is the worst on eight of the nine datasets, so the entry is a method failure on a 966-step series rather than a tuning artefact, and the column ranking is unchanged.

We would be very grateful if you could take another look at our revised manuscript and reconsider your assessment in light of these additional experiments and revisions. We would also sincerely welcome any further comments or guidance you may have.

Thank you very much for your time and consideration.

Reviewer_XMBaOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript proposes EvoFIF, an LLM-assisted missing-value imputation method for time series forecasting. Specifically, the method extracts rules from classical imputation strategies, performs evolutionary search with LLMs, and applies the resulting rules to missing-value imputation for the task of time series forecasting. The authors also conduct experiments on multiple datasets under different missingness patterns, and with different time series forecasting methods to validate the effectiveness of the proposed method.

Critical strengths
  1. The research goal and motivation are reasonable in principle: The manuscript highlights a gap between the objectives of missing-value imputation and time series forecasting, and addresses this issue with an LLM-assisted search for imputation strategies for time series forecasting.
  2. The experimental coverage is broad: The authors conduct experiments on multiple datasets under different missingness patterns, and with different time series forecasting methods, providing support for the effectiveness of the proposed method.
Critical weaknesses
  1. The overall cost of the proposed method is not reported sufficiently: The total time cost of LLM-assisted search is not reported clearly. Although the authors demonstrate the effectiveness of the proposed method, they also mention the high cost of refitting models for each dataset, so this issue needs to be clarified.
  2. The comparison between the proposed method and classical imputation methods is not sufficient: Although the proposed method extracts initial strategies from classical imputation methods, the manuscript does not provide direct comparisons with these classical methods. Consequently, it remains difficult to evaluate the additional performance gains provided by the proposed method over these classical methods.
Questions for authors
  1. Could the authors clarify, based on the existing experimental results, whether there is any relationship between the gains in forecasting performance and the capabilities or modeling principles of the forecasting methods?
  2. Could the authors provide direct comparisons with the classical imputation methods to clarify the additional forecasting performance gains achieved by the proposed method?
3 comments

Thank you for the detailed reading. The four points bear on two things, the cost of the search and the comparison with the classical imputers the search starts from, and both are reported in the revision.

To W1, we report the cost of the search in Appendix F.3 (Table 10).

Table R1. Cost of the search behind the reported configuration.

ItemValue
LLM modeldeepseek-flash
Wall clock10 h 25 min 08 s, including the fitness evaluation
Input tokens0.8 M
Output tokens4.5 M
CostCNY 32, about USD 4.5

The search runs once and returns a configuration rather than a fitted model, so nothing in this table is paid again when the configuration meets a dataset, missingness pattern, forecaster or horizon it was not selected on. Deployed, the configuration imputes a scenario in 0.085 s with no training, against 23 to 221 s for a single baseline imputation of the same scenario, and amortised over the 36 scenarios of the main comparison the search amounts to about 17 minutes and CNY 0.9 per scenario. The marginal cost of extending the coverage is therefore essentially zero, while a model-centric method pays a full refit for every new scenario.

To Q1, Table 11 in Appendix F.4 gives for each of the four forecasters the loss of EvoFIF, the strongest baseline under that forecaster, the absolute and the relative gain, and the mean rank. The gain is positive under all four, which cover statistical modelling, machine learning and deep learning, and its absolute size is of the same order everywhere, 0.026 to 0.052, with the mean rank between 1.17 and 1.58. It is therefore not the capability of the forecaster that produces the gain: the forecaster with the lowest error does not give the largest absolute gain, and only the relative gain rises as that error falls, since the denominator shrinks. What the gain does track is the missingness pattern, largest under MNAR and smallest under BLACKOUT. The strongest baseline, by contrast, changes with the forecaster, Glocal-IB under TimesNet and ARIMA, SAITS under DLinear and PSW-I under XGBoost, so which baseline wins is a property of the pairing while which method wins is not.

To W2 and Q2, we now compare EvoFIF directly with the classical imputers it starts from. Section 4.1 defines the reference, forecasting-guided selection (FGS): the seven classical algorithms of the initial pool impute the series of a scenario, the imputed series pass through the same forecasting model, and the algorithm with the lowest forecasting MSE on the validation window is kept, the test window never entering the selection. We run it once for all 36 settings (FGS-global), which keeps the reusability EvoFIF has, and once per missingness pattern (FGS-pattern), which is the best a single classical algorithm can provide.

Table R2. Forecasting-guided selection against EvoFIF over the 36 settings.

ReferenceAlgorithm selectedMean L_predSettings with EvoFIF lowerMean gapImputation time (s)
FGS-globalTRMF, all settings0.799134 / 369.6%12.615
FGS-patternforward fill, KNN, TRMF0.770732 / 363.1%5.373
EvoFIFdiscovered configuration0.7508----0.085

EvoFIF attains the lower loss in 34 of 36 settings against the reusability-preserving reference and in 32 of 36 against the per-pattern one, with the widest margin under MNAR and the narrowest under BLACKOUT, and it imputes a scenario 63 to 148 times faster than the algorithm the reference selects. Since FGS-pattern already takes the best of the seven by the downstream loss for each pattern, the additional gain over the classical methods comes from composing and evolving inductive biases rather than from choosing one of them. Appendix F.5 reports both references and every setting.

The revised PDF marks these changes in the margin, labelled XMBa.W1, XMBa.Q1 and XMBa.W2&Q2.

Thank you for reporting the search cost and adding comparisons with classical imputation methods. These address my main concerns, so I have raised my rating from 2 to 3.

Here is an optional suggestion. Table 10 lists "deepseek-flash" as the model used for the search. If you used DeepSeek’s official API, it would be helpful to report the exact model version or versions used, since the version served under this identifier can change over time. If the exact model version is unavailable, the date of the search would still be helpful.

Thank you so much for reading our response and for raising your rating. We are truly grateful for your encouragement.

Regarding your suggestion, we used DeepSeek-V4-Flash for the evolutionary search. We also sincerely appreciate your constructive feedback, which helped us improve the paper substantially. In particular, your comments motivated us to report the total search cost more clearly and to add comparisons with classical imputation methods, making the effectiveness and computational cost of EvoFIF easier to assess. These additions have made our experimental evaluation more comprehensive and our presentation clearer.

Thank you again for your time and your careful review!

Reviewer_1sTXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a time series imputation method that transforms imputation inductive biases from structural assumptions implicit in trained models into explicit, composable search objects with executable instantiations, and shifts the optimization objective of FOTSI from a scenario-specific model to a transferable configuration of inductive biases. The method realizes forecast-oriented time series imputation through several components, including Thompson sampling and LLM-based mutation and crossover. In addition, the paper conducts experimental validation on nine datasets, four missingness patterns, and four forecasters, and confirms the superiority of the method.

Critical strengths

S1: This paper provides a detailed analysis and argumentation of the proposed EvoFIF method. It first gives a formalized description of the basic definitions and problem in the form of equations, clearly presenting the foundation of the method's principles. It then elaborates on the proposed method in detail from three parts: the extraction of inductive biases, the evolution of inductive biases, and the verification of inductive biases. Moreover, it provides extensive and sufficient descriptions of the core principles, including Thompson sampling and LLM-based mutation and crossover. S2: This paper conducts sufficient experiments. It thoroughly validates the proposed method from four perspectives—comparative experiments, ablation studies, generalization studies, and case studies—demonstrating the effectiveness of the method.

Critical weaknesses

W1: Insufficient introduction of related work. The EvoFIF method proposed in this paper mainly targets the problem of forecast-oriented time series imputation, but the paper does not introduce frontier research in this direction. As a result, readers cannot clearly understand the current state of research in this area or the paper's deep understanding of the problem. W2: Insufficient experimental evidence for effectiveness. The abstract and introduction mention three requirements for forecast-oriented time series imputation: effectiveness, computational efficiency, and reusability. In the experiments in Section 4, the comparative experiments and generalization study provide sufficient experimental evidence for computational efficiency and reusability, but there is a lack of same-objective methods and key reference groups for effectiveness, so effectiveness is not fully demonstrated.

Questions for authors

Q1: The introduction mentions prior research related to the EvoFIF method, but it lacks an introduction to related work on the target problem of this paper—forecast-oriented time series imputation. This makes it difficult for readers to fully understand forecast-oriented time series imputation, and the authors are expected to supplement this. Q2: This paper conducts sufficient experiments on downstream forecasting effectiveness, but it lacks methods with the same objective and key reference groups. The authors are expected to supplement these.

1 comment

Thank you for the careful reading and for pointing out where the paper still leaves the reader without context. The two weaknesses and the two questions ask the same two things, so we answer them together below.

To W1 and Q1, we added a paragraph to the introduction that places FOTSI in its literature. It groups the work that precedes it into three families, evaluation-guided methods, joint optimization methods and representation learning methods, and then states what all three leave open, that each is designed for one scenario and has to be retrained when the scenario changes, which is the gap EvoFIF addresses. The paragraph closes with a pointer to Appendix A.1, which already carried the method-by-method survey. It is 75 words, so the state of the art is now visible in the body of the paper without displacing the argument.

To W2 and Q2, we added a reference that optimizes the same objective as EvoFIF. We report forecasting-guided selection (FGS), which targets the downstream prediction loss with neither a search over inductive biases nor any retraining of the forecaster. The seven classical algorithms of our initial pool impute the series of a scenario, the imputed series pass through the same forecasting model, and the algorithm with the lowest forecasting MSE on the validation window is kept, the test window never entering the selection. We run it at two granularities: FGS-global selects one algorithm once and uses it in all 36 settings, so it keeps exactly the reusability EvoFIF has, while FGS-pattern selects one algorithm per missingness pattern, which is the strongest a single algorithm of the pool can provide.

Table R1. Forecasting-guided selection against EvoFIF over the 36 settings.

ReferenceAlgorithm selectedMean L_predSettings with EvoFIF lowerMean gapImputation time (s)
FGS-globalTRMF, all settings0.799134 / 369.6%12.615
FGS-patternforward fill, KNN, TRMF0.770732 / 363.1%5.373
EvoFIFdiscovered configuration0.7508----0.085

EvoFIF attains the lower loss in 34 of 36 settings against the reusability-preserving reference and in 32 of 36 against the per-pattern one, with the widest margin under MNAR, 0.9476 and 0.8496 against 0.8243, and the narrowest under BLACKOUT, 0.7289, 0.7289 and 0.7250. It is also 63 to 148 times faster than the algorithm the reference selects. Since FGS-pattern already takes the best of the seven by the downstream loss for each pattern, what EvoFIF adds is the composition and the evolution of inductive biases rather than the choice of one algorithm. Section 4.1 defines the reference, and Appendix F.5 reports every setting.

We did not run the VSF methods as baselines, and Section 4.1 states the reason: they target variable subset forecasting under a fixed forecaster and fit their imputation module jointly with it, while our evaluation spans four forecasters, two of which admit no gradient for that joint fit, and every scenario would need a separate fit, which removes the reusability the paper demonstrates.

The revised PDF marks both changes in the margin, labelled 1sTX.W1&Q1 and 1sTX.W2&Q2.

Comments on the paper

1 comment

We thank all reviewers for the time they spent on the paper and on our responses, and for the specific suggestions and follow-up questions raised during the discussion. The feedback made clear which conclusions needed more support, and gave us the chance to state what our contribution is more precisely. Based on it, we added the following to the revised manuscript.

  1. Positioning against forecast-oriented imputation methods. We added a related-work paragraph to the introduction and extended Appendix A.1, stating why the methods of Table 1 are not baselines and where our contribution sits, a configuration of inductive biases discovered by search. (Section 1, Appendix A.1)

  2. A same-objective reference. We added forecasting-guided selection over the seven classical algorithms that seed our pool, keeping the algorithm with the lowest validation error of the same forecasting model, at two granularities. EvoFIF attains the lower loss in 34 of the 36 settings against the first and in 32 against the second, by 9.6 and 3.1 percent, so the gain is not reproduced by picking a single strong algorithm. (Section 4.1, Appendix F.5)

  3. Cost of the search and of deployment. The search is one off, 10 h 25 min 08 s and CNY 32 in total, about 17 minutes and CNY 0.9 per scenario, while deployment takes 0.085 s against 23 to 221 s for one baseline imputation. (Appendix F.3)

  4. Role of the forecaster and the horizon. Across four forecasters in three paradigms the gain is 3.4 to 14.5 percent at a mean rank of 1.17 to 1.58, and it tracks the missingness structure rather than the forecaster, so the strongest baseline changes with the forecaster while the first place does not. (Appendix F.4)

  5. Scope of the reuse claim. All nine datasets take part in the search on purpose, since they span minute-level loads, sensor streams of 321 and 862 channels and a 966-step count series, so the configuration is selected against heterogeneous distributions, while the forecaster and the horizon are held out and served without retraining. (Abstract, Sections 4.4 and 5)

  6. The w/o Utility ablation. We state that the variant falls back to reconstruction and read its degradation, 28.5 percent under MCAR against 4.9, 2.6 and 10.2 elsewhere with 2 of the 36 settings above 50 percent, which locates the downstream objective where accuracy and forecasting diverge. (Section 4.3)

  7. The HyFAD entry of Table 2. The 5.645 is the DLinear-column mean and comes from national_illness, where that method reaches 48.1 against 0.333 without it. It is the worst on eight of the nine datasets at official defaults, so the entry is a method failure rather than a tuning artefact. (Section 4.4)

Taken together, the revisions leave the claims of the paper where they were and make them harder to misread. Effectiveness is backed by a reference that rules out the strongest single algorithm under the same objective, efficiency by a one-off search amortised to 17 minutes and CNY 0.9 per scenario with sub-second deployment, and reusability by one configuration applied without retraining to every dataset, pattern, forecaster and horizon of the evaluation.

The revised manuscript summarizes the new results in the main text and gives the complete settings, results and analysis in the relevant sections and appendices. We again thank all reviewers for the feedback they provided during the review and the discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.