acceptodds
Under review as a conference paper at ICLR 2027

Are Neural Models Always Necessary For Irregular Time Series Forecasting?

Abstract

Irregular multivariate time-series (IMTS) forecasting is commonly assumed to require specialized neural architectures for their asynchronous nature. We audit this assumption by asking how much standard benchmark performance is explained by simple structure already present in the observed history. We introduce the Anchor family, a set of transparent, gradient-free diagnostic controls based on persistence, smoothing, sparse repetition, recurrence, and non-test empirical rule selection. Across global IMTS benchmarks and novel neural baselines under matched protocols, our Anchor variants achieve an optimal standard-window error in every dataset-metric setting, while AutoAnchor remains competitive across lookback-horizon sweeps and paired entity-level tests. Perturbation, thinning, bootstrap, and ablation analyses show that different datasets reward distinctive low-complexity historical structures, and that neural models regain advantages when history is insufficient or temporal configuration becomes demanding. These results suggest that current IMTS benchmarks can overstate the necessity of complex neural representations, thereby motivating stronger non-neural diagnostic controls first. The source code is available at: https://anonymous.4open.science/r/Anchor-Family/.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Irregular Time Series, Diagnostic Controls, Temporal Representation Learning

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

This manuscript challenges existing public benchmarks for irregular multivariate time-series (IMTS) forecasting. By constructing and evaluating naive methods within a specific anchor family on these benchmarks, the paper points out that the complex designs of modern neural networks for IMTS may be redundant. Furthermore, the authors conclude that neural modeling is mainly advantageous when the input context is short and less informative.

Critical strengths
  • The problem is clear, and the motivation to question the existing irregular time series benchmark is reasonable and meaningful. As corresponding neural models become more and more complex, my own intuitions also indicate that the patterns under these irregular series may be sufficient enough for imputation and therefore easy to model. The authors verified this point in a relatively competitive benchmark.
  • The verifiers used in the study are simple and reasonable, and thus are robust and trustworthy in reflecting the conclusion.
Critical weaknesses

Although the motivation is valid and the empirical results support the authors' main claim, the content of the manuscript is rather thin, whether in the methodology or the experimental section. It highlights a critical issue—that the community should be more mindful of the necessity and validity of complex neural modeling for irregular time series—which is valuable. However, the authors merely report experimental results showing that naive methods can achieve competitive or even superior performance on the selected benchmarks. While this is a meaningful empirical finding, it does not meet the bar for a full scientific paper in the ICLR main track. Besides, the main figure and tables need substantial improvements; the AI-generated evidence is strong.

In conclusion, I lean toward a 'reject' rather than a 'strong reject,' as the paper raises a valid challenge to the community, even though its content and depth are far from sufficient for a rigorous scientific publication. Since that specific intermediate option is unavailable in this year's scoring standard, I am assigning a score of 2.

Questions for authors

Could you polish the main figure and include more in-depth analyses? For example, how should future neural networks for irregular time series be designed based on your findings, and what are the theoretical intuitions behind this redundancy? Furthermore, would this conclusion still hold if the patterns in the irregular time series were sufficiently complex?

2 comments
AuthorsOpenReview ↗

[Request for Acceptance Appraisal] Strengthened Scientific Depth, Mechanistic Evidence, and Boundaries of Neural Modeling Necessity

We appreciate the reviewer’s recognition of the central question and take the concern about scientific depth seriously. The revised paper is not merely about simple models winning: it develops a benchmark audit of which historical signals explain IMTS performance, when they suffice, and when neural models add value, aiming to enforce a basic but challenging guardrail that all future IMTS studies must surpass before claiming the SOTA title.

Q1. Is there sufficient methodological and scientific depth?

Section 3 and new Figure 2 define ANCHOR as a five-level diagnostic framework, not one naïve predictor. A shared library contains persistence, robust summaries, EMA, bounded trends, and phase lookup. Naive-, Expo-, and SparseAnchor isolate structures; ERMAnchor uses non-test empirical selection; AutoAnchor adds structural diagnostics. Per-variable choices are frozen before testing, without sample-wise test routing. The evaluation now spans 18 latest neural baselines, four datasets, and matched splits, scaling, masks, and metrics.

The framework can in principle operate on regular or irregular series, without implying untested regular-time superiority. It controls for predictability already available from history. Our Anchor family serves as a diagnostic tool for future studies in irregularity.

Q2. Are the underlying irregular series genuinely complex?

The new Table 1 documents measurable heterogeneity in the sampling process:

DatasetGap CVAvg. observed variables/time
P120.147.1
MIMIC-III0.935.6
HumanActivity0.383.0
USHCN0.911.1

USHCN combines highly variable gaps with sparse asynchronous variable coverage. Sampling irregularity is real, but it does not imply that the target-relevant information is irreducibly complex.

Q3. What explains the apparent redundancy of neural complexity?

The revised analyses move beyond aggregate error:

EvidenceInterpretation
Temporal perturbations, Table 4: shuffling timestamps changes AutoAnchor MSE by +0.0007 (P12), +0.0102 (MIMIC), but +0.3125 (USHCN).Dependence on exact timing varies greatly by benchmark.
USHCN ablations, Table 6: removing sparse or phase-sensitive candidates increases MSE from 0.1565 to 0.1848 or 0.2181.Identifies the specific historical mechanisms behind gains.
200 bootstrap ERM refits, Table 6: candidate selections vary across calibration resamples.Reveals calibration sensitivity rather than assuming perfect selection.
Table 5 and Appendix: bootstrap confidence intervals, paired ANCHOR-versus-persistence tests, and volatility/density strata.Tests robustness and heterogeneity beyond a leaderboard.

Theoretical intuition: when low-complexity summaries already capture most target-relevant historical information, learned representations have limited extra signal to exploit. When latent states, cross-variable coupling, or longer dependencies matter, that sufficiency breaks down. Perturbations and ablations support this interpretation, but we do not present it as a formal theorem.

Q4. Would the conclusion survive sufficiently complex patterns?

Not universally, and the revision demonstrates why. In Figure 3, AutoAnchor achieves the lowest MSE in 13 of 19, not all, lookback–horizon settings. On USHCN, its MSE rises from 0.1565 at lookback/horizon (150,3) to 1.1591 at (100,3); extending the horizon at lookback 150 also restores a neural MSE advantage. Under 10% history retention, even after matched recalibration, AutoAnchor reaches 0.4265 ± 0.0831 MSE versus 0.2351–0.2673 for neural baselines. These are explicit counterexamples defining the boundary of our claim.

Appendix Table 12 shows MSE gains over persistence even in high-volatility strata (29.0% P12; 22.7% MIMIC; 65.0% USHCN; 35.4% HumanActivity). Yet volatility is not a complete test of nonlinear or coupled dynamics, so we avoid that stronger claim.

Q5. Figure quality and implications for future architectures.

We polished Figure 1 into a TikZ-style technical schematic; new Figure 2 visualizes ANCHOR's rule library, calibration, and frozen test-time forecasting following same theme as Figure 1 for optimal visual consistency. The numerical evidence comes from the documented experiments, not the illustrations; anonymous source code is provided. Section 5 identifies priorities for neural design: latent states, cross-variable dependencies, regime changes, and limited-history/long-horizon forecasting.

We respectfully request a positive reconsideration: the revised contribution is a mechanistic benchmark audit, with a reusable control framework, quantitative irregularity analysis, interventions, uncertainty, and documented failures. It asks when complexity earns its predictive value, rather than presuming it from leaderboards.

I thank the authors for their additional explanations and modifications. After reviewing the updated PDF, I maintain my position that the figures and tables are not yet polished enough for a top-tier conference, as they retain a distinctly AI-generated style; the authors should at least refine and adjust them manually. Furthermore, the additional evidence indicates that the study's main conclusion is not universal, which may limit its broader impact and influence.

Reviewer_9sWROpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper conducts a benchmark audit of IMTS forecasting, questioning whether complex neural representations are strictly necessary to achieve strong benchmark performance. The authors introduce the ANCHOR family, a hierarchy of transparent, gradient-free diagnostic controls built from fixed history-derived forecasting rules (e.g., persistence, exponential smoothing, mode repetition, and phase lookup), culminating in AUTOANCHOR, which selects rules via non-test ERM and structural diagnostics. Evaluating across four standard benchmarks against 12 neural baselines, the authors show that ANCHOR variants match or outperform neural models at standard windows.

Critical strengths
  • The paper addresses a critical methodological question in IMTS forecasting by moving beyond single-window aggregate leaderboards to isolate what specific low-complexity historical signals drive performance on each benchmark.
  • Unlike prior baselines that rely on grid alignment, linear layer training, or restricted deep networks, the proposed ANCHOR ladder operates natively on unaligned, asynchronous timestamps without gradient learning or test-time routing leakage, achieving highly competitive errors in milliseconds on a CPU.
  • The paper provides a well-executed empirical results check for the IMTS forecasting literature, showing that simple historical heuristics frequently match complex neural models.
Critical weaknesses
  • What explains the discrepancy in Table 3 where neural baseline errors drop substantially relative to Tables 2 and 4, reversing the P12 MSE outcome in favor of GraFITi?
  • Why does non-test ERM fail to select the sparse/phase candidates on USHCN, requiring heuristic prior overrides (πv(H)\pi_v(\mathcal{H})) to achieve strong test performance? The paper lacks a theoretical or statistical explanation for why ERM underperforms on calibration data.
  • The major concern is related to baselines. The paper omits several recent 2025–2026 cross-variable and hierarchical architectures cited in Section 2 (e.g., Hi-Patch, HyperIMTS, TimeCHEAT) and more complex, strongly coupled IMTS datasets.
  • Does the neural advantage under aggressive history thinning (Figure 3) persist if AUTOANCHOR is recalibrated directly on thinned historical data rather than evaluated with frozen full-history specifications?
  • How current LLMs behave on this task? Is there any study authors can conduct?
Questions for authors

See weaknesses.

1 comment
AuthorsOpenReview ↗

[Request for Acceptance Appraisal] Resolved Evaluation Discrepancies, Expanded Neural Baselines, and New Evidence on ERM Selection and History Thinning

We thank the reviewer for concerns that prompted substantive revisions: a rebuilt Table 3, six new neural baselines, thinning-matched recalibration, and USHCN mechanism/stability analyses.

Q1. Why did Table 3 contradict Tables 2 and 4?

We agree that the original comparison was inconsistent. The former Table 3 used a paired common-valid-coordinate/entity-level evaluation, whereas Tables 2 and 4 reported globally masked benchmark metrics. These evaluation bases are not interchangeable. We have replaced Table 3 with a direct summary of Table 2's global results, retaining neural mean ± SD and labeling deltas descriptive, with no paired neural significance claims.

Under this consistent protocol, P12 MSE is 0.302 (AutoAnchor), 0.308 ± 0.002 (GraFITi), and 0.309 ± 0.001 (APN) (Tables 2–3); Table 4 reports GraFITi's unperturbed MSE as 0.3075. Thus, the earlier GraFITi reversal does not occur under the uniformly reported global metric. We corrected the comparison rather than rationalizing incompatible numbers.

Q2. Why do structural diagnostics outperform non-test ERM on USHCN?

ERM minimizes non-test calibration risk, not held-out risk. Sparse, irregular histories can make the ranking of plausible rules sensitive to the calibration sample; minimizing observed calibration error does not guarantee the most stable held-out choice. This is a selection/generalization issue, not evidence that ERM is incapable of choosing phase-aware rules.

Our new Table 6 reports 200 bootstrap ERM refits: for four of five USHCN variables, no candidate dominates more than 36.5% of resamples. Modal candidates include phase-nearest (36.5%), last–phase mixtures (32.0%, 26.0%), and phase–trend (36.0%). ERM therefore considers phase structure, but selections are calibration-sensitive.

ERM-only MSE/MAE is 0.1662/0.2634 versus AutoAnchor's 0.1565/0.2187. Removing sparse or phase-sensitive candidates increases AutoAnchor MSE to 0.1848 or 0.2181, respectively. The diagnostic overrides are fixed from non-test information, not test labels. We explicitly label their thresholds exploratory rather than theoretically optimal (Secs. 3, 5).

Q3. Missing recent baselines and strongly coupled datasets.

We expanded Table 2 from 12 to 18 neural models, adding Hi-Patch, HyperIMTS, TimeCHEAT, GRUwE, TFMixer, and ASTGI, using matched splits, masks, scaling, and evaluation. The strongest newly added neural result in each dataset remains behind AutoAnchor:

DatasetAutoAnchor MSE/MAEBest new neural MSE/MAE
P120.302 / 0.359HyperIMTS: 0.305 / 0.361
MIMIC0.379 / 0.366HyperIMTS: 0.394 / 0.373
USHCN0.157 / 0.219TimeCHEAT: 0.164 / 0.272
HumanActivity0.042 / 0.115ASTGI: 0.043 / 0.119

This directly addresses the omission of recent hierarchical/cross-variable architectures. New Table 1 also characterizes genuine irregularity: MIMIC has 96 variables and gap CV 0.93, while USHCN has gap CV 0.91 and only 1.1 variables observed per timestamp. We agree, however, that a purpose-built strongly coupled benchmark would further test the limits of variable-wise controls; we have not added such a dataset, and do not claim universal coverage.

Q4. Does the neural advantage persist after thinning-matched recalibration?

On USHCN, yes; on the other three datasets, no. We added recalibration directly on histories thinned to 10% (Sec. 4.3, Fig. 4):

DatasetRecalibrated Auto MSENeural MSE
P120.4564 ± 0.00750.5279
MIMIC0.6232 ± 0.01110.8375
HumanActivity0.0981 ± 0.00290.2740
USHCN0.4265 ± 0.08310.2351–0.2673

Neural checkpoints still originate from full histories, so this tests robustness to missing history, not retraining all models under thinning. The USHCN reversal is deliberately retained as evidence of where neural models add value.

Q5. How do current LLMs perform?

We agree LLM forecasting merits investigation but do not claim unperformed experiments. A meaningful comparison would serialize irregular timestamp–variable–value records and future queries, enforce the same splits/masks and MAE/MSE, and control context and compute/token budgets. Serialization and evaluation-harness choices make this a distinct study; Section 5 explicitly states this scope limitation.

In summary, the revision resolves the reported metric inconsistency, includes the requested modern architectures, tests ERM selection stability, adds thinning-specific calibration, and documents where neural models retain an advantage. We respectfully ask the reviewer to reassess the strengthened, clearly bounded benchmark-audit contribution.

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

The paper considers forecasting methods for Irregular multivariate time-series (IMTS). The development of the paper is around their ANCHOR methodology which they demonstrate through extensive benchmarking against competing methods for IMTS. They demonstrate ANCHOR to be a strong performer in regards to IMTS.

Critical strengths

The critical strength is the development of the ANCHOR family of methods along with the extensive validation of the method both against its variants and against other competitors. This is a very nice contribution for the evaluation of time-series forecasting. It is a much necessary benchmarking system for improvement of forecasting for IMTS systems.

Critical weaknesses

No time-series data are actually shown, examples of some of the time-series data would be helpful to see in order to provide context. Also, there is no comparison or discussion of how the ANCHOR suite compares if time-series data is regularly sampled.

Questions for authors

How does the ANCHOR suite compare to regularly sampled data? What are the right metrics for time-series forecasting when not considering IMTS data? And does ANCHOR have broader applicability?

1 comment
AuthorsOpenReview ↗

[Request for Acceptance Appraisal] Clarification on Dataset Characteristics, Regular-Sampling Generalizability, and the Broader Applicability of ANCHOR

We thank the reviewer for these constructive questions. We agree that data examples and clearer scope help. The revision therefore distinguishes the methodological applicability of Anchor from the already established empirical settings, seeking your stronger score appraisal.

Q1. Representative time-series examples and dataset characteristics

The revision now separates model design from the observed data. Figure 2 summarizes the Anchor family: its shared candidate library, calibration hierarchy, and frozen test-time forecasting operator. Figure 5 then shows representative observed input histories from the datasets.

We also add Table 1, which quantifies structural differences among the benchmark configurations:

DatasetGap CVAvg. variables/timestamp
P120.147.1
MIMIC-III0.935.6
HumanActivity0.383.0
USHCN0.911.1

Gap CV is the median within-sample coefficient of variation of successive observed-timestamp gaps; zero denotes regular spacing. Thus, P12 has less variable observed-timestamp gaps in this configuration, whereas MIMIC-III and USHCN have substantially more variable gaps. USHCN records only 1.1 variables per observed timestamp on average, indicating especially asynchronous observations. These results show that the audit covers several IMTS regimes rather than one common missing-data pattern. They do not turn P12 into a fully regularly sampled benchmark.

Q2. Comparison with regularly sampled forecasting

The underlying Anchor rules--persistence, exponential smoothing, recent-level summaries, trimmed means, local trends, and phase-nearest lookup--can in principle be evaluated on regularly sampled histories. Likewise, ERMAnchor's non-test empirical selection is a general way to construct transparent forecasting controls.

Our present implementation, diagnostics, and experiments are designed for irregularly sampled multivariate time series. We do not claim that Anchor outperforms specialized regular-time forecasting methods, and we now state this explicitly in Section 5, “Scope and Future Directions.” A fair regular-time comparison would require matched regular benchmarks, preprocessing, horizons, competing methods, and tuning budgets.

The distinction nevertheless connects our contribution to the regular-time literature: simple predictors and standardized evaluations have challenged the assumption that increasingly complex architectures are automatically necessary for strong forecasting performance (Zeng et al., 2023; Qiu et al., 2024; Chen et al., 2025). We extend that evaluation question to IMTS, where unequal gaps and asynchronous channels add challenges absent from conventional regularly sampled benchmarks. P12 provides a useful IMTS case with less variable timestamp gaps, but it is not a substitute for a fully regular-time study.

Q3. Metrics beyond IMTS forecasting

MAE and MSE apply regardless of whether timestamps are regular. Our protocol evaluates globally masked MAE and MSE in the training-scaled space (Equations 7--8). With fully observed regular-time targets, the mask is simply mb,k,v=1m_{b,k,v}=1, which recovers the standard average MAE and MSE.

More broadly, our point is not that these two metrics alone settle model quality. Evaluation should also test whether added architectural complexity improves over strong transparent historical controls under the same conditions. The revised paper supplies this complementary evidence through lookback--horizon sweeps (Figure 3), entity-bootstrap confidence intervals (Table 5), temporal perturbations and thinning analyses (Figure 4 and Table 4), and paired AutoAnchor-versus-NaiveAnchor tests (Appendix Table 7).

Q4. Broader applicability

We view Anchor's broader value primarily as a diagnostic evaluation principle, not as a universal or SOTA forecasting architecture. Across clinical, activity, and climate benchmarks, AutoAnchor reduces MSE relative to last-value persistence by 24.3% on P12, 17.0% on MIMIC-III, 30.6% on HumanActivity, and 57.0% on USHCN (Table 5). The mechanisms differ: P12 and MIMIC-III primarily benefit from empirical selection over the shared rule library; HumanActivity is largely explained by horizon-aware smoothing; and USHCN benefits from sparse and phase-sensitive candidates. Candidate-removal, perturbation, and bootstrap analyses identify these differences rather than treating all gains as evidence for a single heuristic.

The results also define limits. Neural models regain measurable advantages when history is shortened, observations are severely thinned, or horizons become more demanding, especially on USHCN. Anchor therefore should not be read as evidence that learned representations are unnecessary. Rather, it motivates a stronger standard for future IMTS studies: before attributing a forecasting gain to complex learned representations, establish what transparent history-derived controls achieve under matched protocols.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Official Announcement by The Authors

Dear Chairs and Reviewers,

Thank you for your reviewing service. The authors have provided summaries, rebuttals, and clarifications on November 11. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your current recommendations.

Thank you.

Dear Reviewers,

I'm writing to gently remind you that the rebuttal discussion phase deadline is approaching. Please take a moment to read the authors' rebuttal, and feel free to respond if any concerns remain that you'd like them to address. Thank you very much for your time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.