acceptodds
Under review as a conference paper at ICLR 2027

Impute or Forecast Directly? A Systematic Benchmark for Time Series Forecasting with Missing Values

Abstract

Time series forecasting with missing values (TSFMV) is commonly addressed through either impute-then-forecast pipelines or end-to-end missingness-aware models. Recent studies increasingly advocate end-to-end modeling to avoid error propagation between separately optimized imputation and forecasting stages. Whether this design consistently improves forecasting performance, however, remains unclear because existing studies differ substantially in their datasets, missingness settings, and evaluation protocols. We introduce TSFMV-Bench, a systematic benchmark spanning three real-world datasets, seven missingness scenarios, and three missing rates. Beyond performance rankings, we investigate when the relative advantage between the two paradigms changes and how different missingness characteristics shape forecasting difficulty. Our evaluation yields three findings. First, strong impute-then-forecast pipelines remain competitive across a majority of evaluated conditions, challenging the common concern that stage-wise error propagation necessarily makes them inferior. Second, neither paradigm dominates uniformly, and their relative advantage changes with the type and severity of missingness. Third, paradigm preference is shaped by both the recoverability of missing observations and where reconstruction errors occur. Cross-variable information helps only when supporting observations remain available, while errors near the forecasting boundary can substantially disadvantage two-stage pipelines. These findings challenge a blanket preference for end-to-end modeling and motivate evaluating both paradigms across diverse missing-data settings. Code and benchmark resources are available in [anonymous repository](https://anonymous.4open.science/r/TSFMV-Bench-BE81).

Primary area:
datasets and benchmarks
Keywords:
Time Series Forecasting with Missing Values, Unified Benchmarking, Forecasting Paradigm Evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_L7WTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents TSFMV-Bench, a benchmark for comparing impute-then-forecast pipelines and end-to-end methods for time series forecasting with missing values. It evaluates multiple methods across three datasets, seven missingness scenarios, and three missing rates. The results show that the performance gap between the two approaches varies across datasets, missingness patterns, and missing rates.

Critical strengths

S1. The paper studies a practical and relevant problem: whether to impute missing values before forecasting or forecast directly from incomplete data. S2. The benchmark compares multiple methods across different datasets, missingness patterns, and missing rates, providing a useful basis for evaluation.

Critical weaknesses

W1. The authors construct two-stage baselines primarily by combining different imputers and forecasting backbones, without sufficiently comparing representative existing impute-then-forecast methods or explaining the rationale for this modular selection. Since the groups also differ in architecture, imputation method, and training supervision, the results primarily support comparisons among the selected methods rather than attributing performance differences to the paradigms themselves. W2. The notation is inconsistent. For example, the paper defines X=[x_{t,n}] but subsequently uses z_{t,n} without explaining the relationship between them. W3. The categorization does not sufficiently account for other missing-data forecasting strategies. For example, TOI-VSF [1] proposes task-oriented imputation through joint learning, explicitly imputing missing data while allowing the forecasting objective to guide the imputation module. W4. Some citations do not match the corresponding claims. W5. The discussion in Section 5.3 is inconsistent with Figure 6. W6. There are some typos in the paper. For example, “STFMV-Bench” should probably be “TSFMV-Bench.”

[1] Hao et al., “Is Precise Recovery Necessary? A Task-Oriented Imputation Approach for Time Series Forecasting on Variable Subset,” IEEE Transactions on Knowledge and Data Engineering, 2025.

Questions for authors

See above

3 comments
AuthorsOpenReview ↗

Paradigm Definition, Baseline Selection, and Evaluation Scope

W1 & W3

Regarding concerns about the representativeness of our baselines and the categorization of missing-data forecasting methods. We would like to clarify our paradigm definitions, the rationale for modular baseline selection, and the scope of our conclusions, as follows.

  1. Clarification of paradigm definitions. Our benchmark distinguishes E2E and ITF primarily by whether imputation and forecasting are optimized independently or jointly. Specifically, ITF independently trains an imputation model before training a downstream forecasting model, whereas E2E directly optimizes forecasting from incomplete observations, potentially incorporating auxiliary reconstruction objectives. We acknowledge that task-oriented joint imputation methods, such as TOI-VSF, represent an important strategy that combines explicit imputation with forecasting-guided optimization. However, TOI-VSF primarily addresses variable-subset forecasting (i.e., complete data during the training phase and subset variable-missing during testing phase) and differs from the independently optimized ITF pipelines examined in our benchmark. We will clarify this distinction and discuss such methods in the revised manuscript.

  2. Rationale for modular ITF baseline selection. Our modular design is intentional and reflects the standard ITF workflow. We select Mean and ImputeFormer to represent statistical and deep spatiotemporal imputation, respectively, and STID (MLP), MTGNN (CNN-GNN), and PatchTST (Transformer) to cover different forecasting architectures and cross-variable modeling strategies. Their combinations allow us to assess ITF performance across multiple implementations instead of relying on a single pipeline. To further strengthen the diversity of imputation methods, we additionally evaluated SAITS [1] combined with STID and MTGNN, providing further evidence on how imputer selection affects downstream forecasting performance. Specifically, it shows that better imputation accuracy does not necessarily lead to better forecasting accuracy. Under AQI MNAR at a 75% missing rate, repairing recent missing values reduced forecasting MAE by 2.29%, compared with 0.57% for count-matched random repairs, despite similar aggregate reconstruction errors. Moreover, the SAITS-trained predictor retained its advantage when evaluated on IF-imputed histories, suggesting that the fitted predictor also contributes to the observed performance difference.

Dataset and conditionImputation MAE, SAITS / IFForecast MAE, SAITS / IF
AQI, MAR, 25%13.7610 / 10.457442.9765 / 45.9829
AQI, BLOCK-S, 75%20.3740 / 13.036849.9000 / 52.4792
ETTh1, MNAR, 50%0.5963 / 0.99112.1704 / 2.0486
  1. Scope of paradigm-level conclusions. Our objective is to systematically compare representative implementations of independently trained ITF pipelines and E2E forecasting methods under consistent experimental protocols. Our analyses identify cross-variable information availability and reconstruction errors near the forecasting boundary as important factors associated with their relative performance. We also agree that isolating paradigm effects is valuable, but disentangling architecture, imputation strategy, and training supervision presents methodological challenges. We consider more rigorous causal identification of paradigm effects an important direction for future work.

We will refine the paradigm definitions, explain our baseline selection criteria, and clarify the scope of our conclusions in the revised manuscript.

[1] Wenjie Du, et al. SAITS: Self-attention-based imputation for time series. Expert Systems with Applications. 2023

W2, W4 & W6: Notation, Citation Accuracy, and Presentation Corrections

We thank the reviewer for identifying these presentation and clarity issues. We will carefully revise the manuscript to improve notation consistency, citation accuracy, and overall presentation, as detailed below.

  1. Notation consistency (W2). The notation (Z) was used for the underlying reference time series in the missingness formulation, while (X) was used for the time-series observations in the forecasting formulation, without explicitly clarifying their relationship. We will unify the notation by using (X) for the reference time series and (X^{\mathrm{obs}}) for the incomplete observations, and revise the corresponding definitions, equations, and figures throughout the manuscript.

  2. Citation accuracy (W4). We will carefully cross-check all citations against their corresponding claims and correct any inaccuracies. We would also appreciate it if the reviewer could point out specific instances of mismatched citations, which would help us address these concerns more precisely.

  3. Typographical errors (W6). Thank you for pointing out this typo. We will correct “STFMV-Bench” to “TSFMV-Bench” and thoroughly proofread the manuscript to eliminate similar errors.

We appreciate the reviewer's careful reading and constructive suggestions, which will help improve the clarity and accuracy of our manuscript.

W5 Clarification of the consistency between Section 5.3 and Figure 6

In the paper, Section 5.3 investigates three factors associated with the relative performance of E2E and ITF: (1) cross-variable dependency structures and paradigm preferences (Figures 6(a)–(b)), (2) observation availability under different missingness patterns (Figure 6(c)), and (3) the influence of reconstruction errors near the forecasting boundary (Figure 6(d)). We will further refine Section 5.3 to explicitly discuss this contrast and clarify how each subplot supports the corresponding analysis.

Thanks for the clarification. The additional results show that the choice of imputer affects the downstream forecasting performance. However, my original concern remains. Since the ITF and E2E models differ in architecture and in whether they use imputed targets during training, it is unclear how much of the performance gap can be attributed to the training paradigm itself.

Reviewer_GKreOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper focused on benchmarking time-series forecasting under missing values. The authors claimed that previous benchmarks used unaligned missingness settings and failed to cover various missingness patterns and mechanisms. The authors therefore proposed appending synthetic missingness to datasets and applying systematic evaluation to time-series forecasting models. Experiments showed model performance under various missingness patterns and mechanisms.

Critical strengths
Critical weaknesses

1.The limitations of previous works are overclaimed. In paragraphs 2- 3, the authors summarized that existing methods carry their own missing rates, missing patterns, and missing mechanisms, which is a limitation. Yet, according to Table 1, STMMoE is already covering every missing pattern, while DMM (Cao et al., cited in line 81 but not appearing in Table 1) covered every missing mechanism. Even for missing rates, a lot of the methods in Table 1 are showing a wider variation than the established TSFMV benchmark. The paper should therefore frame its contribution as a unification of three independent scenarios where each of them has been thoroughly studied.

2.The intellectual contribution is unclear and not supported by experimental results. While the authors claimed that a systemic evaluation covering missingness mechanisms is needed, their missingness definition is taken from previous papers. The authors never strengthened why such a missingness definition is needed. It’s also unclear whether such a missingness definition actually aligns with real-world data, since the authors are appending synthetic missingness. Furthermore, Point seems to be the same thing as MCAR. And as also shown in Table 3, most of the models have nearly identical numbers and rankings under most missingness mechanisms and patterns, including MCAR and Point. The flips reported in Table 5 are also mostly marginal. The designed patterns and mechanisms are therefore not differentiating the models well, which weakens the necessity of the variations. Also, although the paper is positioning its contributions as building a benchmark, they’re not positioning themselves against any benchmarking papers. And their method selection is incomplete, missing a large number of methods in Table 1 and DMM.

Questions for authors

See the weaknesses

4 comments
AuthorsOpenReview ↗

Clarification of Novelty and Positioning Relative to Prior Work

W1

We agree that prior studies have investigated diverse missingness patterns, mechanisms, and rates. However, we would like to further point out that our work differs fundamentally in its research objective and the insights it aims to establish.

  1. Different research objectives and contributions. Existing studies primarily focus on improving specific methods under missing-data conditions. For example, STMMoE (Wu et al., 2025) evaluates all four structural missingness patterns to demonstrate the effectiveness of its proposed end-to-end forecasting method, while DMM (Cai et al., 2025) investigates MCAR, MAR, and MNAR mechanisms for time series imputation. In contrast, our objective is to systematically investigate when and why E2E and ITF paradigms differ in forecasting performance under diverse missingness conditions. To this end, we conduct a unified evaluation of five E2E methods and six ITF pipelines across three missingness mechanisms, four structural patterns, and three missing rates under consistent experimental protocols. Our contribution lies not in combining these settings, but in revealing the conditional advantages of different forecasting paradigms and analyzing the factors underlying their performance differences.

2.Clarification of prior work coverage. We appreciate the reviewer's suggestion and will revise Table 1 to explicitly include DMM and clarify the evaluation coverage and research objectives of existing studies, including STMMoE. We will also refine the Introduction to emphasize systematic paradigm-level comparison and empirical insights rather than presenting the coverage of missingness settings itself as a novelty.

AuthorsOpenReview ↗

Clarification and Supplementation of Contributions and Experimental Evidence

W2

We thank the reviewer for these valuable comments. In the paper, our main contributions lie in systematically investigating when and why E2E and ITF forecasting paradigms exhibit different advantages under diverse missing-data conditions. Specifically, paradigm preference depends on the availability of cross-variable information and the location of reconstruction errors.

We would like to further clarify the motivation and necessity of our missingness settings, their ability to differentiate forecasting paradigms, the significance of our experimental findings, and the positioning and coverage of our benchmark, point by point.

  • Motivation. We adopt established missingness definitions rather than introducing new ones. These definitions capture two important aspects of incomplete observations: missingness mechanisms (MCAR, MAR, and MNAR) describe how observation availability depends on data values, while structural patterns (Point, Block-T, Block-S, and Block-ST) characterize how missing observations are distributed across time and variables. Our objective is to evaluate how these characteristics affect the relative forecasting performance of E2E and ITF methods.

  • Clarification of MCAR and Point. We appreciate the reviewer for highlighting their similarity. MCAR uses independent Bernoulli masking, whereas Point uniformly selects a fixed number of missing entries. Thus, Point can be viewed as a fixed-budget implementation of MCAR rather than a fundamentally different missingness process. We will clarify this distinction and avoid presenting them as independent evidence of distinct missingness effects.

  • Evidence that missingness conditions affect paradigm preference. We acknowledge that some configurations yield similar rankings and that some preference changes are small. However, the differences are not universally marginal, particularly under high missing rates. For example, on PEMS-BAY at 75% missingness, the relative MAE difference between ITF and E2E changes from −5.4% under MCAR (favoring ITF) to +11.0% under MAR (favoring E2E). Similarly, Block-S favors ITF by 3.1%, whereas Block-T favors E2E by 6.0%. These contrasts demonstrate that the nature of missingness, not merely its severity, can substantially affect paradigm preference.

  • Real-world missingness validation. To further address the realism of synthetic missingness, we additionally evaluated both paradigms on PhysioNet [1] clinical time series with naturally occurring missingness (80% missingness). BiTGraph achieves an MAE of 0.4337, outperforming the strongest ITF pipeline (0.4494). Moreover, its standardized MAE advantage increases from 0.0072 to 0.0771 when the target variable has no observations during the six hours preceding the forecast. These results provide additional evidence that recent observation availability is associated with the relative forecasting performance of E2E and ITF in a real clinical setting.

MethodMAE ↓RMSE ↓MRE ↓
Mean → STID0.51400.757675.27%
Mean → MTGNN0.50350.749573.73%
SAITS → STID0.45040.703165.96%
SAITS → MTGNN0.45070.703466.00%
ImputeFormer → STID0.45040.693265.95%
ImputeFormer → MTGNN0.44940.692365.81%
BiTGraph0.43370.680963.52%
GRU-D0.44640.693665.38%
  • Benchmark positioning and method coverage. We agree that the manuscript should better position itself relative to existing benchmarks. Our benchmark addresses a different question: the comparative effectiveness of E2E and ITF forecasting under incomplete observations. Our original model selection covers representative architectures from both paradigms, and we have additionally evaluated SAITS [2] as an alternative imputation method. Regarding DMM, its primary objective is time series imputation rather than direct forecasting; it could therefore serve as an imputation component within an ITF pipeline. We will expand the related-work discussion and clarify the scope of the evaluated methods.
Dataset and conditionImputation MAE, SAITS / IFForecast MAE, SAITS / IF
AQI, MAR, 25%13.7610 / 10.457442.9765 / 45.9829
AQI, BLOCK-S, 75%20.3740 / 13.036849.9000 / 52.4792
ETTh1, MNAR, 50%0.5963 / 0.99112.1704 / 2.0486

The results further show that better imputation accuracy does not necessarily lead to better forecasting accuracy.

We appreciate the reviewer's feedback and would welcome further discussion or clarification on any remaining concerns.

[1] Ruichu Cai, et al. Causal View of Time Series Imputation: Some Identification Results on Missing Mechanism. IJCAI. 2025

[2] Wenjie Du, et al. SAITS: Self-attention-based imputation for time series. Expert Systems with Applications. 2023

I have read the authors response and prefer to keep my score.

We thank the reviewer for taking the time to carefully read our responses. We also appreciate your feedback and respect your assessment.

We hope that our clarifications and additional experiments have addressed the main concerns raised in the review. We would be grateful for any further feedback on specific concerns that remain unresolved, as this would help us further improve the manuscript.

Reviewer_PRc7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TSFMV-Bench, a systematic benchmark for comparing E2E forecasting with ITF methods under diverse missing-data settings. The experiments show that neither paradigm consistently dominates; instead, performance depends on the missingness pattern, available cross-variable information, and distortion of recent history.

Critical strengths

S1: It investigates the timely and controversial question of whether E2E methods are necessarily superior to ITF approaches,covering diverse synthetic missingness mechanisms, patterns, and rates.

Critical weaknesses

W1: The E2E–ITF comparison may be confounded by unequal candidate model sets. In Table 3,since ImputeFormer (IF) is the strongest baseline, the results may simply indicate that IF outperforms the other baselines, rather than demonstrating that the ITF paradigm as a whole is superior. W2: The benchmark evaluates only three datasets with missingness patterns largely imposed synthetically, which may not represent real-world missing-data processes. In addition, the paper does not sufficiently report results across multiple random seeds and independent mask realizations with uncertainty estimates. Consequently, aggregate claims about paradigm preference may be sensitive to the selected datasets and random samples.

Questions for authors

Q1: How many independent training seeds and missing-mask realizations were used for the main results? Please report variability estimates to assess whether the reported win counts are robust. Q2: How were the three datasets selected, and do they adequately represent broader applications? Experiments on more domains or real missingness patterns would strengthen the generality of the conclusions. Q3: Could the ITF advantage be mainly driven by ImputeFormer (IF)? Please clarify whether the result reflects the ITF paradigm or the performance of a single imputer.

3 comments
AuthorsOpenReview ↗

The ITF Advantage Beyond ImputeFormer

W1 & Q3

We thank the reviewer for highlighting the potential influence of the limited set of imputation methods. To examine whether the observed ITF advantage is driven solely by ImputeFormer (IF), we additionally evaluated SAITS [1] across all datasets and missingness conditions. SAITS-based ITF also outperforms the corresponding E2E methods, showing that the observed ITF advantage is not limited to IF.

The results further show that better imputation accuracy does not necessarily lead to better forecasting accuracy. For example, under AQI MAR at a 25% missing rate, SAITS has a 31.59% higher imputation MAE than IF but achieves a 6.54% lower forecasting MAE. The table below illustrates this difference across representative settings.

Dataset and conditionImputation MAE, SAITS / IFForecast MAE, SAITS / IF
AQI, MAR, 25%13.7610 / 10.457442.9765 / 45.9829
AQI, BLOCK-S, 75%20.3740 / 13.036849.9000 / 52.4792
ETTh1, MNAR, 50%0.5963 / 0.99112.1704 / 2.0486

To further investigate this distinction, we conducted a fixed-predictor diagnostic. Under AQI MNAR at a 75% missing rate, repairing recent missing values reduced forecasting MAE by 2.29%, compared with 0.57% for count-matched random repairs, despite similar aggregate reconstruction errors. Moreover, the SAITS-trained predictor retained its advantage when evaluated on IF-imputed histories, suggesting that the fitted predictor also contributes to the observed performance difference.

Finally, we would like to highlight that our study examines the conditions under which each paradigm has an advantage, while existing studies on E2E methods often report better performance than ITF approaches. Table 3 reports results averaged across missing rates, which may cover differences under individual conditions. As shown in Figure 5, E2E methods perform better in some settings. Our findings therefore emphasize the conditional advantages of the two paradigms, rather than the overall superiority of either one.

AuthorsOpenReview ↗

Dataset Election Clarification and New Dataset Inclusion

W2 & Q2

Our original benchmark follows established evaluation protocols to systematically compare forecasting paradigms under controlled missingness conditions. To further address generalizability, we have added PhysioNet, a fourth application domain with naturally occurring missingness. The analysis of new results (shown in the following table) reinforce our central finding that neither E2E nor ITF is universally superior, and their relative advantages depend on data and observation characteristics.

  • Dataset selection and diversity. We selected three widely used datasets spanning electricity transformer monitoring (ETTh1), air quality (ChinaAQI), and traffic (PEMS-BAY). They differ in the number of variables (7, 36, and 325), sampling frequencies (hourly and five-minute), temporal coverage (approximately two years, one year, and six months), and cross-variable dependency structures. This diversity enables us to investigate how data characteristics interact with missingness.
  • Motivation and precedent for synthetic missingness. Controlled synthetic missingness is an established evaluation approach in prior research, including E2E forecasting studies such as BiTGraph (ICLR 2024) and GinAR (KDD 2024), as well as the missing-mechanism study Causal View of Time Series Imputation (IJCAI 2025) [2]. These studies introduce controlled missingness into real-world time series to assess model behavior under specified conditions.
  • Additional experiments with naturally occurring missingness. We additionally evaluated E2E and ITF methods on PhysioNet [1] clinical time series using their native missingness patterns (missing ratio is 80%). Unlike the original controlled experiments, this evaluation examines forecasting with naturally incomplete observations. The results show that BiTGraph achieves the best forecasting performance, reducing MAE by 3.49% and RMSE by 1.65% relative to the strongest ITF pipeline. We further examined the influence of recent observation availability. When the target variable was observed within six hours before the forecasting origin, BiTGraph's standardized MAE advantage over the best ITF method was 0.0072; without such recent observations, the advantage increased to 0.0771. This result reinforces the importance of recent historical information, consistent with our PEMS-BAY analysis showing that reconstruction errors near the forecasting boundary can substantially influence prediction accuracy.
MethodMAE ↓RMSE ↓MRE ↓
Mean → STID0.51400.757675.27%
Mean → MTGNN0.50350.749573.73%
SAITS → STID0.45040.703165.96%
SAITS → MTGNN0.45070.703466.00%
ImputeFormer → STID0.45040.693265.95%
ImputeFormer → MTGNN0.44940.692365.81%
BiTGraph0.43370.680963.52%
GRU-D0.44640.693665.38%
AuthorsOpenReview ↗

Robustness Across Random Seeds and Missing-Mask Realizations

W2 & Q1

The original benchmark comprised 693 experiments across three datasets, seven missingness scenarios, three missing rates, and eleven methods. Considering the high computational workloads, we use one training seed and one missing-mask realization per condition. To assess reproducibility, we conducted two additional independent repetitions on ETTh1, resulting in three paired training/mask seeds, and report the mean and standard deviation (SD) across repetitions.

The main finding on ETTh1 remains consistent across all three repetitions. Specifically, STID+ImputeFormer and PatchTST+ImputeFormer consistently outperform all five E2E methods in both MAE and RMSE across all 21 missingness conditions under every seed. This confirms that the observed ITF advantage on ETTh1 is reproducible across different training seeds and missing-mask realizations, rather than being driven by a single random configuration.

The following tables report the mean ± SD across three repetitions, with results aggregated over the three missing rates (25%, 50%, and 75%).

1. ETTh1 forecasting MAE (mean ± SD, lower is better).

MethodMARMCARMNARPOINTBLOCK-SBLOCK-TBLOCK-ST
E2E forecasting
BRITS2.1086 ± 0.13671.8810 ± 0.02652.0635 ± 0.14071.8702 ± 0.02001.8493 ± 0.02001.9020 ± 0.02652.1048 ± 0.0387
GRU-D1.9252 ± 0.09001.8392 ± 0.02242.0456 ± 0.10151.8182 ± 0.01411.8060 ± 0.01411.8218 ± 0.00302.0479 ± 0.0316
BiTGraph1.8839 ± 0.15941.7029 ± 0.00701.9424 ± 0.05661.7153 ± 0.00681.7041 ± 0.01001.7240 ± 0.01411.9332 ± 0.0091
Hi-Patch2.1026 ± 0.15071.9035 ± 0.01732.1623 ± 0.11491.8974 ± 0.05001.8777 ± 0.02241.9708 ± 0.07142.2601 ± 0.0520
S4M2.1543 ± 0.28502.0157 ± 0.00992.1756 ± 0.07682.0302 ± 0.02832.0078 ± 0.02652.0883 ± 0.09642.2448 ± 0.0245
ITF forecasting
MTGNN+Mean3.1575 ± 0.12883.0532 ± 0.02003.1581 ± 0.71683.0579 ± 0.02243.0882 ± 0.02653.0719 ± 0.01003.0607 ± 0.0283
STID+Mean3.0988 ± 0.13153.0382 ± 0.00803.1080 ± 0.66843.0425 ± 0.01733.0774 ± 0.02243.0481 ± 0.00643.0237 ± 0.0141
PatchTST+Mean2.8095 ± 0.11452.5290 ± 0.00292.8214 ± 0.58682.5541 ± 0.00402.6122 ± 0.00442.5445 ± 0.01412.6289 ± 0.0141
MTGNN+IF1.8062 ± 0.09751.6733 ± 0.02451.8263 ± 0.09951.6637 ± 0.01001.6936 ± 0.02451.7588 ± 0.03321.9686 ± 0.0173
STID+IF1.6391 ± 0.08941.5300 ± 0.01731.6491 ± 0.05201.5162 ± 0.00681.5315 ± 0.01411.6109 ± 0.02001.7395 ± 0.0100
PatchTST+IF1.6654 ± 0.08661.5536 ± 0.01411.6677 ± 0.04241.5478 ± 0.00501.5611 ± 0.00251.6125 ± 0.01731.7444 ± 0.0069

2. ETTh1 forecasting RMSE (mean ± SD, lower is better).

MethodMARMCARMNARPOINTBLOCK-SBLOCK-TBLOCK-ST
E2E forecasting
BRITS3.7077 ± 0.33933.3066 ± 0.02833.8853 ± 0.47013.2697 ± 0.03003.2371 ± 0.00803.3605 ± 0.01733.8191 ± 0.0678
GRU-D3.4933 ± 0.27353.2977 ± 0.02833.9367 ± 0.43833.2830 ± 0.01003.2592 ± 0.02243.3120 ± 0.00943.8305 ± 0.0742
BiTGraph3.5097 ± 0.36973.2046 ± 0.01003.9883 ± 0.25463.2444 ± 0.02833.1970 ± 0.00423.2511 ± 0.03463.6901 ± 0.0387
Hi-Patch3.9180 ± 0.40363.5266 ± 0.03164.2691 ± 0.43163.5343 ± 0.06003.4902 ± 0.07423.7262 ± 0.16464.3687 ± 0.1034
S4M3.8664 ± 0.55713.6161 ± 0.00894.1112 ± 0.29383.6511 ± 0.10153.5898 ± 0.04003.7482 ± 0.15594.1373 ± 0.0265
ITF forecasting
MTGNN+Mean5.2911 ± 0.35005.1142 ± 0.04125.5278 ± 0.87165.0993 ± 0.04475.1503 ± 0.03465.1356 ± 0.02455.1800 ± 0.0557
STID+Mean5.2004 ± 0.26005.0899 ± 0.01735.4498 ± 0.77885.1011 ± 0.03615.1495 ± 0.03165.1016 ± 0.02005.1175 ± 0.0332
PatchTST+Mean4.7337 ± 0.29854.3732 ± 0.01415.1136 ± 0.65764.4069 ± 0.01734.4868 ± 0.01004.3749 ± 0.02004.5949 ± 0.0361
MTGNN+IF3.2996 ± 0.26913.0192 ± 0.03163.4191 ± 0.17553.0100 ± 0.02833.0963 ± 0.04473.2156 ± 0.11583.6771 ± 0.0300
STID+IF3.1607 ± 0.22832.9491 ± 0.03463.2875 ± 0.14872.9075 ± 0.00972.9527 ± 0.01413.0617 ± 0.02453.4367 ± 0.0332
PatchTST+IF3.2365 ± 0.23223.0218 ± 0.03463.3778 ± 0.17582.9952 ± 0.01733.0338 ± 0.00433.1129 ± 0.00463.4802 ± 0.0374

[1] Wenjie Du, et al. SAITS: Self-attention-based imputation for time series. Expert Systems with Applications. 2023

[2] Ruichu Cai, et al. Causal View of Time Series Imputation: Some Identification Results on Missing Mechanism. IJCAI. 2025

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.