acceptodds
Under review as a conference paper at ICLR 2027

Spectral Dynamics or Structured Linear Maps? Tiny Source-Only Time-Series Transfer

Abstract

Can a model trained on one time-series dataset forecast an unseen dataset without target fitting? We study this question under a strict source-trained, target-free protocol using DynaFK, a compact forecaster that combines a seasonal anchor with a phase-dependent residual. At any fixed horizon, its encoder, diagonal recurrence, and decoder compose exactly into a phase-specific linear map. This motivates a matched direct control: with 397 real scalars it matches DynaFK at , while larger linear and spectral models can improve accuracy at higher storage cost. The recurrence’s clearest benefits are horizon-independent parameter storage and execution across different channel counts. Source choice and horizon still change the ranking, so the results describe an accuracy–storage–portability trade-off rather than a universal advantage for spectral dynamics.

Primary area:
learning on time series and dynamical systems
Keywords:
time-series forecasting, source-only transfer, spectral dynamics, structured linear models, distribution shift

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_533yOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the problem of time-series forecasting trained on a source domain without requiring target-domain data and proposes the DynaFK model. A key contribution of the study is the revelation that, for a fixed forecasting horizon, DynaFK degenerates into a linear mapping with a specific phase; however, I do not understand why this is claimed as a contribution of the paper.

Critical strengths

None

Critical weaknesses
  1. I believe the paper's scientific contribution is insufficient. Its core theoretical result is quite intuitive: once the forecast horizon and model parameters are fixed, the encoder, linear recurrent module, and decoder can be collapsed into a single linear mapping. Although the derivation is correct, it essentially amounts to nothing more than substituting and rearranging finite summations. More concerningly, many of the subsequent experimental results undermine the rationale behind the proposed "spectral mechanism" itself. Sinkhorn alignment and L2 smoothing yielded no substantial improvements; the "global-spectrum" and "phase-linear" variants performed on par with DynaFK; and for longer forecast horizons, DynaFK significantly underperformed compared to the Linear or Koopa models. Consequently, beyond the realization that "a simpler linear model can explain most of DynaFK's performance over short horizons," it remains unclear to me what the paper's core scientific contribution actually is.

  2. The paper fails to convincingly demonstrate the value of DynaFK as a novel method. DynaFK integrates a host of reasonable technical components—including seasonal naive forecasting, instance normalization, phase rearrangement, complex-valued encoders/decoders, diagonal spectral dynamics, smoothing, cross-sampling-rate interpolation, and history resampling—yet the paper does not clearly articulate why this specific combination is necessary. In fact, the authors' own experimental results show that a simple "phase-linear model" with only 397 parameters performs essentially on par with DynaFK, while standard baselines such as Linear and FITS often achieve higher forecasting accuracy. For instance, on the ETTh2 dataset, DynaFK achieves an MSE of 0.3124, whereas Linear/FITS achieves 0.2947; on the ECL dataset, DynaFK yields 0.2426, compared to 0.2260 for Linear. The model's sole advantage lies in the fact that its parameter count is independent of the prediction horizon length; however, even this aspect was not validated by using a single "frozen checkpoint" across different horizons. Figure 3(c) presents results from models trained separately for each horizon, and the paper itself characterizes this merely as "parameter-shape evidence." Consequently, it is difficult to understand why DynaFK—rather than the simpler baseline models—should be regarded as the paper's primary contribution.

Questions for authors

What novel method contributing scientifically to the field of time-series forecasting does the paper propose? Are the experimental datasets and baseline comparisons adequate? Do not submit AI-generated coursework as an academic paper to ICLR or any other publication platform; doing so wastes the reviewers' time and effort.

Reviewer_NR9gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a tiny spectral forecaster for source-only transfer, then shows that at a fixed horizon it's just a linear map, and that a small size parameter linear model can also do it.

Critical strengths

The paper is well structured and written clearly such as the protocol description, with experimental verification and discussion.

Critical weaknesses

The main issue is that the proposed method does not offer any real improvement. Several baselines are better than it on most domains.

Questions for authors
  1. When would you actually want to have DynaFK instead of a plain linear model?

  2. Could you also run the 397-parameter linear control at H=192 to 720?

  3. If linear models are channel-independent, then what makes DynaFK more portable across channel counts?

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

The paper studies whether latent spectral dynamics explain the transfer performance of tiny time-series forecasters in a strict source-only setting, where a model is trained on one dataset and applied to an unseen target without any fitting. Using a compact spectral forecaster (DynaFK) as a test case, it shows that at a fixed horizon the encoder, diagonal recurrence, and decoder collapse exactly into a phase-specific linear map, and a directly learned linear map with fewer parameters matches its accuracy. Experiments on ETT, ECL, Traffic, Weather, and Exchange Rate frame the results as an accuracy–storage–portability trade-off.

Overall: The paper is carefully executed with a rigorous evaluation protocol, but its general claim about spectral dynamics rests on a single model, and the claimed benefits of the recurrence are not directly tested.

Critical strengths

(a) The source-only protocol is rigorous. It cleanly separates source training, checkpoint selection, target inference, and evaluator-only scaling, which addresses common sources of target leakage in transfer forecasting. The taxonomy of transfer regimes is also useful for the community. (b) The exact fixed-horizon reduction gives a principled matched control for attributing performance, and it is verified numerically. The paper reports its negative findings honestly (alignment adds nothing, phase structure does not drive accuracy) and backs the comparisons with moving-block bootstrap intervals and detailed reproducibility artifacts.

Critical weaknesses

(i) The title asks about spectral dynamics in general, but the analysis covers only one linear diagonal-recurrence model, for which the collapse to a linear map holds almost by construction. Koopman-style models with nonlinear or input-dependent dynamics are not analyzed the same way, so the conclusion may not generalize. (ii) The claimed benefits of the recurrence are not directly shown. Each horizon is trained separately rather than rolling out one frozen checkpoint, and gains over seasonal naive shrink to around 1% at longer horizons. Channel portability is shared by the simple Linear baselines, and the smaller storage does not translate into faster or more memory-efficient inference. (iii) The experiments are limited to strongly seasonal sensor datasets at mostly one horizon, and much of the evidence beyond the main setting comes from archived runs. On less seasonal data such as Weather and Exchange Rate, the learned models fall below seasonal naive or persistence, and there is no comparison to zero-shot foundation models as a reference.

Questions for authors

(1) Does the collapse result extend to spectral or Koopman models with nonlinear or input-dependent dynamics, and if not, how should the general question in the title be interpreted? (2) Can a single frozen checkpoint trained at a short horizon be rolled out to much longer horizons, and how does it compare with horizon-specific linear models? (3) How would the approach handle non-seasonal or regime-switching data such as financial time series, where the seasonal anchor and the choice of period may not be meaningful?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.