acceptodds
Under review as a conference paper at ICLR 2027

SAWS: Forecasting Inaccurate Time Series with Sharpness-Aware Weak Supervision

Abstract

Inaccurate observations degrade time-series forecasting twice: inaccurate historical observations obscure the dynamics to predict, and inaccurate future observations used as labels degrade supervision. While applying means of observation accuracy improvement can recover useful information, treating resulting improved observations as ground-truth observations can amplify any errors that may have been introduced. We introduce SAWS, a forecasting framework that combines vector-quantized historical retrieval with sharpness-aware weak supervision. A past-only codebook compresses similar historical contexts and their observed dynamics into coordinate-supported prototypes. Retrieved prototypes correct inputs and supply weak targets without accessing the current forecast horizon. A detached, sample-level sharpness score measures the sensitivity of an exponential-moving-average forecaster on observed targets; together with prototype support and dispersion, it controls the influence of weak supervision. Multiscale temporal features preserve the distinction between short-term dynamics and longer-term context without requiring dense attribute-graph propagation. We characterize the bias–variance tradeoff of prototype completion and establish when sharpness weighting reduces weak-label error, explicitly contending with selection bias and confidence miscalibration. A leakage-controlled experimental evaluation offers evidence that SAWS is able to offer better accuracy and reliability on inaccurate data than are state-of-the-art baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Incomplete Data, Sharpness

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_KQ1WOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This submission addresses an important problem—forecasting when both historical inputs and future training labels may be incomplete or corrupted—and proposes to combine prototype-based completion with sharpness-aware weak supervision. The high-level motivation is reasonable, and the manuscript shows awareness of temporal leakage and the potential failure modes of pseudo-label weighting. However, I cannot recommend acceptance because the submission contains a fundamental and highly concerning contradiction regarding its empirical evaluation.

Critical strengths

This submission addresses an important problem

Critical weaknesses

On page 7, the authors explicitly state that “numerical experiments have not yet been executed” and that “all result cells are intentionally empty.” However, Table 1 on the following page contains complete numerical results and is followed by claims of state-of-the-art performance. The conclusion again states that the paper reports no empirical outcomes.

This fundamental contradiction makes the validity and provenance of the reported results impossible to assess. The manuscript appears unfinished and insufficiently checked. I therefore recommend rejection.

Questions for authors

On page 7, you state that “numerical experiments have not yet been executed” and that “all result cells are intentionally empty.” However, Table 1 reports complete numerical results, and the text claims state-of-the-art performance. Were these experiments actually conducted?

Reviewer_zE7gOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SAWS for time series forecasting with incomplete observations. A vector-quantized historical memory supplies input completions and weak future targets using past-only retrieval. An exponential-moving-averag teacher provides a head-local sharpness score, which combines with prototype support and dispersion to control training weights. The framework also includes an observed-label anchor and multiscale temporal modeling. The theoretical analysis characterizes prototype estimation error and the tradeoff between weak-label quality and selection bias.

Critical strengths

S1. The framework separates retrieval keys from future payloads, excludes overlapping reference windows before prototype aggregation, while freezes the evaluation archivem, and address an important source of leakage in retrieval-based forecasting.

S2. The analysis distinguishes prototype bias, finite-support variance, and selection effects. It also acknowledges that increased support does not eliminate systematic bias and that sharpness weighting requires an additional calibration assumption.

Critical weaknesses

W1. The sharpness-based weighting lacks sufficient justification. Sharpness is measured on observed targets but used to weight weak targets at missing coordinates, without evidence that it predicts their errors. Proposition 4.2 assumes this relationship, while Appendix B.2 shows residual sensitivity in a scalar linear-head setting, which cannot distinguish unreliable supervision from difficult valid examples.

W2. No ablation demonstrates an advantage over prototype reliability alone or residual-based weighting.

Questions for authors

Q1: Please explain their provenance and reconcile them with the statements in Section 5 and the conclusion that no numerical or empirical experiments have been executed.

Q2: What evidence shows that observed-coordinate sharpness predicts weak-target error at missing coordinates after accounting for prototype reliability? Does it improve forecasting over matched variants using prototype reliability alone and residual-based weighting?

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

SAWS addresses forecasting when historical inputs and training targets can be missing or noisy. A VQ codebook stores prototypes of past contexts and their continuations. Retrieval completes missing inputs and provides weak training targets. Prototype support and dispersion estimate reliability, while a sharpness score from an EMA teacher controls training weights. The paper also analyzes completion error and weighting bias.

Critical strengths
  1. The paper addresses a useful problem: missing future observations reduce the supervision available during training, in addition to missing history making prediction harder.
  2. The method separates observed values from completed values and restricts retrieval to past data. The analysis also acknowledges that confidence weighting can change which samples dominate training.
Critical weaknesses
  1. The reported status of the experiments is contradictory. Page 7 says numerical experiments have not been run and all result cells are empty, but Table 1 on page 8 contains numerical results and page 9 claims leading performance. The conclusion again says no empirical outcomes are reported. Please clarify which experiments were completed and how the table values were obtained.
  2. The experiments do not establish whether the main components are needed. There are no reported ablations removing VQ completion or sharpness weighting, or separating input completion from training-target completion. The only performance table uses 5% missingness. Could you provide these ablations and results under stronger missingness and noise?
  3. The reliability estimate may give too much confidence to repeated, similar windows. Equation (4) uses reference counts to reduce uncertainty, but heavily overlapping windows do not provide independent evidence. Excluding references near the query does not remove overlap among the references themselves. Could you test non-overlapping references or adjust the counts for dependence, and show how this affects completion reliability and forecasting?
Questions for authors

Please see weaknesses.

Comments on the paper

1 comment

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.