acceptodds
Withdrawn from ICLR 2027

CODA: Time-Series Forecasting through Auditable Concept-State Transitions

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate long-horizon forecasting requires explanations tied to the prediction mechanism. We propose CODA, a Consistency-Driven Concept-State Architecture that learns five softly anchored temporal concepts and predicts through a direct, horizon-specific affine transition. The history encoder remains expressive, while the final head exposes the exact contribution of each concept and the last observation. Analytic targets stabilize concept meaning, and local compatibility penalties regularize the explanatory state. Across six public benchmarks, CODA achieves the lowest error in 10 of 12 primary-horizon MSE/MAE comparisons among the evaluated methods. In a separate multi-horizon evaluation against N-HiTS and ModernTCN, it attains the lowest MSE in 19 of 24 dataset–horizon settings and ties in one more. Concept-target ranges from 0.62 to 0.94, and removing semantic anchoring increases average MSE from 0.268 to 0.741. Controlled concept and residual ablations, an explicit Electricity coefficient analysis, and input-noise tests connect predictive performance with inspectable model behavior. CODA thus offers accurate forecasting through a compact semantic interface whose downstream effects are analytically verifiable.

Primary area:
interpretability and explainable AI
Keywords:
Time-Series Forecasting, Interpretable Machine Learning, Concept-State Transitions, Semantic Alignment, Consistency Regularization

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_FX27OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CODA proposes a concept-based interface for time-series forecasting. Five softly anchored concepts and an explicit observed anchor feed a horizon-specific affine decoder. It produces H forecasts from one past-only history and concept vector. The effects of concept edits are exactly inspectable within the fitted model. The paper evaluates predictive performance and the interface through analytic-statistic controls, concept removal and extension, residual alternatives, and alignment and coefficient analyses. It limits this interpretation to model behavior rather than causal interventions on the underlying physical system.

Critical strengths
  1. The concept interface is coherent. The affine decoder makes the fitted model's response to concept edits mathematically explicit. The manuscript acknowledges the limits of soft alignment and the absence of harmonic phase. It also treats compatibility terms as priors rather than physical identities. These qualifications give the interpretation a clear scope.

  2. The analytic-statistic control, concept-removal and extension experiments, residual alternatives, and alignment/coefficient analyses examine relevant aspects of the interface. These experiments provide positive evidence for the intended method, even though its correspondence with the empirical artifact remains unresolved.

Critical weaknesses

The main unresolved issue is the correspondence between the reported forecasting task and the empirical artifact. The manuscript specifies fixed-origin direct H-step forecasting: all H outputs use one initial history and one concept vector. The supplied reassessment's preserved source excerpts instead describe a scalar prediction head. For consecutive one-step predictions, this pipeline updates the history with newly observed values. Those observations were unavailable at the initial origin. Rolling one-step forecasting is legitimate at each new origin, but its task and information set differ from the stated fixed-origin comparison. The encoder, previous-window reference, and channel-sharing definitions also differ. These discrepancies make it necessary to identify the version used for the reported results.

An older supplementary version could explain the discrepancy. The README also identifies its results as debug-scale experiments rather than the full benchmark. The current material, however, gives neither a version chronology nor a connection between an existing direct-H implementation and the main tables or ablations. The excerpts do not show which pipeline generated those results. A general old-version explanation or newly corrected code would not establish that connection. Identifiable records are needed before this concern can be resolved.

Questions for authors
  1. Please identify the already-existing fixed-origin direct-H version used for the reported results and provide its version timeline. How does it relate to the supplied scalar, moving-observed-history implementation?

  2. Please provide saved H-step predictions, the configuration, and metric records for one or two identifiable main-table or ablation entries. These records should link the model version and initial forecast origins to the reported values. Existing records may resolve the concern without rerunning the full benchmark. If those entries used rolling-observed history, the affected comparisons would need fixed-origin evaluation with matched baselines.

Reviewer_YN5kOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

The authors introduce a time-series forcasting model called CODA. This consists in a first transformer based step that maps time series to five learned coordinates. These are somewhat encouraged through some regularization and correspond to some temporal statistics. From these a horizon-dependent affine head predicts outcomes. The results are encouraging on a series of tests.

Critical strengths
  • The presented interface is inspectable because the learned transformations carry approximately a meaning provided by the regularization. This could be useful for interpretability or debugging.
  • It is interesting in its own, when a small-sized model, especially if it is restricted to only 5 features, performs well.
Critical weaknesses
  1. I think the novelty is relatively narrow. They cite, e.g., van Sprang et al., 2024, which seems to be already the same overall concept of concept supervision. The novelty is mostly concentrated on the specific choice of five concepts.
  2. The main advantage of this is supposed to be interpretability. Another advantage is that the concepts aren't hard-coded, but learned, so that a learner can deviate from them. But don't these clash? If I want to interpret a signal in a certain way, I should know what a feature means, but I essentially only know what a feature means most of the time.
  3. The theoretical analysis is elementary and essentially entirely independent from the method. Proposition 1 uses only the triangle inequality and Cauchy-Schwarz, and Proposition 2 uses only properties of linear maps. None of this has anything todo with the five concepts and their specific form.
Questions for authors

Q1. What are the conceptual novelties beyond van Sprang et al., 2024 that distinguish this work from a simple application of that work? Q2. Is my interpretation in weakness 2 correct, or can we use the features for interpretation regardless? Q3. What is the role of Proposition 1 and 2? What theoretical insight do they provide specifically for the proposed method?

Reviewer_du3KOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors study how to make time-series forecasts interpretable through a small set of concept states. They propose CODA, which aligns learned representations with five analytically defined concepts and uses a horizon-specific affine head to produce forecasts. The reported forecasting results are promising. However, in its current form, the empirical performance is the most compelling aspect of the paper; I am not yet convinced that the learned states have stable meanings or that the proposed transition analysis delivers the claimed level of auditability.

Critical strengths

1)The model has a clear and relatively simple forecasting interface. 2)Using analytic targets to guide concept learning is an interesting design choice. 3)The reported results are competitive across several forecasting datasets.

Critical weaknesses

1)The mathematical notation needs more careful revision. Vectors, matrices, and scalars should be distinguished and formatted consistently throughout the paper. I also encourage the authors to check the equations for correctness and ensure that all symbols are defined when they first appear. 2)The mathematical presentation needs more care. In Eq. (4), the relationship between the moving-average change and the latest growth is only a soft prior, while the division in R2,t may be unstable for anchors near zero. The authors should explain these choices more clearly and quantify their effects in practice. In addition, the practical relevance of Eq. (9) would be easier to assess if the rank and singular values of the fitted coefficient matrices were reported. 3)The references should be checked and formatted consistently. Some entries have incomplete or uneven bibliographic details, and the edition notation for the Box et al. book should be corrected. 4)The five concepts are chosen in advance, but the paper does not establish that the learned coordinates retain these meanings consistently across datasets, horizons, or individual examples. More direct tests of concept stability would strengthen the interpretability claim. 5)Several compatibility constraints are described as soft local priors rather than exact relationships. The paper should clarify what can actually be audited when these constraints are violated, and report the size and frequency of such violations on test data. 6)The effect of changing one concept while holding the others fixed follows algebraically from the affine forecasting head. It does not by itself show that this intervention corresponds to a meaningful or feasible change in the input time series. The wording around concept effects should reflect this distinction. 7)The propositions characterize error bounds and distinguishability under a fitted model, but provide limited evidence that the learned concepts are semantically faithful in practice. Empirical tests connecting the propositions to actual concept errors would make the theoretical section more compelling. 8)The forecasting gains are encouraging, but the central claim is about auditable concept-state transitions. The evaluation should therefore include stronger case studies or quantitative measures of whether these transitions help users understand, diagnose, or correct forecasts.

Questions for authors

My questions for the authors are detailed in the weaknesses above.

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

This paper focuses on designing an inspectable forecasting architecture to reveal the hidden properties of the given history sequence determine the future, which is of great importance to applications like finance and public health. Concept bottleneck models are representative technical pathway, which places named variables on the predictive path. Inheriting this approach, this paper proposes a Consistency-Driven Concept-State Architecture (CODA) that learns 5 concepts via expressive encoders like transformer and further predicts through a direct, horizon-specific affine head. The concept-removal and extension experiments are provided to support the necessity of each concept.

Critical strengths
  1. The system design is model-agnostic and lightweighted, which is possible to integrate into many existing time series forecasting framework.
  2. The 5 concepts contains both basis temporal statistics (level, growth) and frequency domain information (first-harmonic amplitude).
Critical weaknesses
  1. The writing in this paper requires significant improvement. Generally, a well-written paper should allow a reader to grasp the research gap of existing work simply by reading the abstract. However, I found it's difficult to extract this essential information from the abstract. The only takeaway for me is that this paper aims to leverage simple, concept-based designs to make predictions auditable. However, the abstract doesn't illustrate the challenge resolved. More concerning is the inclusion of numerous terms that are left explanation within the given context, such as analytic targets, local compatibility penalties, and semantic anchor. More than half of the abstract contents are just enumerating the experiments conducted rather than explaining the research gap of existing work and the design principle.
  2. The paper doesn't explain why we need to softly supervise the 5 learned concepts rather than fully supervising. Moreover, this paper doesn't provide either theoretical or empirical evidence to support these five concepts are informationally complete for the forecasting task, especially given that the final prediction is obtained via a simple affine mapping.
Questions for authors
  1. For the concept-removal and extension experiment, only adding the acceleration can't sufficiently demonstrate the current 5 concepts are adequate.
  2. Does the contributions from the last observation and 5 concepts to the final prediction are complementary? Why the final affine head includes the last observation additionally?
  3. The introduction section is hard to understand, especially from line 92 to line 95.

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

Related papers

Discussion (0)

Sign in to comment.