acceptodds
Under review as a conference paper at ICLR 2027

Mostly the Norm: Decomposing Frequency-Domain Forecast Losses into Basis, Norm, and Residual Weighting

Abstract

Frequency-domain forecast losses such as FreDF change three things at once: the basis in which the residual is measured, the norm applied to it and the weight of each coordinate. We separate the three. (i) Under a squared norm any unitary basis is inert (Parseval), and training confirms it. (ii) Under , on iTransformer and at an equal mix of the two terms, the same term with no transform recovers most of FreDF's gain and is behind FreDF on none of nineteen held-out dataset–horizon cells. At FreDF's released weights it is ahead on ten of twenty cells and behind on four, mostly at the short horizons of the hourly datasets. (iii) To measure explicit weighting we use \method as an instrument, a squared loss on the residual whitened by detached, shrunk, Kronecker-factored second-moment estimates. Linear theory gives the invariance condition under which a fixed quadratic metric can move the fitted predictor. Empirically, the estimated weighting is ahead of the time-domain term on no iTransformer cell. Its largest gain, on PatchTST at ETTh1 and horizon over three seeds, does not survive a Holm correction, depends on the metric's eigenvectors rather than its spectrum, and disappears once the test period moves three months later. A time-domain term at matched strength is therefore the baseline against which frequency-domain and whitened losses should be judged.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, frequency-domain losses, generalized least squares, residual whitening, multivariate forecasting

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

This paper studies frequency-domain objectives for multivariate time-series forecasting, using FreDF as the main example. The authors decompose such objectives into three factors: the basis in which the residual is represented, the norm applied to it, and the weighting of residual coordinates.

The paper argues that a unitary basis change is inert under a squared norm, and then uses a time-domain L1 loss as a control to study how much of FreDF's gain is due to the norm rather than the Fourier basis. It further introduces WHITEDF, a residual-whitening objective based on estimated temporal and cross-variable second moments, and connects the analysis to classical GLS results.

The empirical study mainly uses iTransformer, with additional experiments on PatchTST and DLinear, and reports a large number of diagnostic and follow-up experiments on ETT, Weather, ECL, and Traffic.

Critical strengths
  1. The paper asks a useful diagnostic question.

Separating the effects of basis, norm, and residual weighting is a reasonable way to analyze frequency-domain forecasting objectives. In particular, the time-domain L1 control is simple and useful, and on the main iTransformer/FreDF experiments it reproduces a substantial part of FreDF's improvement.

  1. The authors are transparent about many negative and unstable results.

The paper explicitly reports failed experiments, multiple-testing corrections, test-period sensitivity, implementation issues, and cases where the proposed whitening objective hurts performance. This transparency is valuable, even though in my view these results ultimately weaken the broader claims of the paper.

Critical weaknesses
  1. The baseline and dataset coverage is not sufficient for the broad claim made by the paper.

The paper is framed as an analysis of frequency-domain forecasting losses in general, but the central norm-versus-basis conclusion is primarily established for FreDF on iTransformer. The most systematic experiments are concentrated on the ETT subsets and Weather, while several closely related methods discussed by the paper itself, such as Time-o1, OLMA, SFALoss, and QDF, are either not evaluated or only appear in limited auxiliary experiments.

This is particularly problematic because the ETT subsets do not provide strong independent domain diversity. The paper itself notes that the held-out WHITEDF gains on iTransformer are concentrated on datasets sharing the development cell's station and test months. ECL and Traffic broaden the whitening analysis, but they are not used to provide an equally systematic test of the main norm-versus-basis claim.

As a result, I think the evidence supports a much narrower statement: on the authors' main iTransformer/FreDF setup, a time-domain L1 term explains much of the gain. I do not think it supports the broader framing that frequency-domain forecasting losses are "mostly the norm."

  1. The evaluation is heavily exploratory/post-hoc, and the defensive presentation does not resolve this issue.

The WHITEDF configuration is selected using the test error of the ETTh1, T=96 development cell. Several later experiments are explicitly motivated after observing previous results: additional candidate criteria are introduced after earlier comparisons, some follow-up cells are selected after observing harm, and the Holm correction families are defined after seeing the results. The authors disclose these choices, but disclosure does not turn them into independent confirmatory evidence.

This issue is amplified by the structure of the manuscript. Many sections read as responses to anticipated reviewer objections: checkpoint-rule controls, scale matching, rotations, shifted test periods, alternative criteria, and extensive caveats about what each result does or does not establish. These controls can be useful diagnostically, but they do not substitute for a pre-specified evaluation on genuinely unseen methods, datasets, and time periods.

The instability of the whitening results illustrates the problem. The largest PatchTST gain does not survive the Holm correction and disappears when the test period is shifted forward. The paper ultimately acknowledges that it cannot predict when whitening will help. In my view, this leaves the weighting part of the paper as an exploratory investigation rather than a validated general result.

Questions for authors
  1. What exactly is the intended scope of the "mostly the norm" conclusion?

If the claim is specifically about FreDF on iTransformer, I think the title and framing should be narrowed substantially. If the intended claim is about frequency-domain forecasting losses more generally, why are closely related objectives such as Time-o1, OLMA, SFALoss, and QDF not evaluated systematically under the same norm-versus-basis controls across multiple backbones and datasets?

  1. Can the authors provide a genuinely held-out evaluation in which all loss definitions, hyperparameters, and hypotheses are fixed before looking at the evaluation results?

In particular, the current WHITEDF setting is chosen using the test error of the development cell, and several later controls are explicitly introduced after observing earlier outcomes. A validation-only selection protocol followed by evaluation on unseen datasets or later non-overlapping time periods would materially affect my assessment.

  1. Can the authors clarify which experiments and hypotheses were specified before observing test results and which were added post hoc?

The current manuscript mixes confirmatory and exploratory analyses, and I found it difficult to determine which conclusions should be interpreted as independently validated.

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

The paper decomposes frequency-domain forecasting losses into basis choice, norm, and residual weighting. It argues that a unitary Fourier basis is inert under a squared norm, shows that a time-domain L1 loss recovers most of the reported FreDF benefit, and studies a covariance-whitening objective as a more principled form of residual weighting. The experiments use matched seeds, multiple datasets and horizons, and several controls to separate these effects.

Critical strengths
  1. The basis–norm–weighting decomposition is simple but useful. In particular, the Parseval argument and the matched time-domain controls make it much easier to identify which part of a frequency-domain loss can actually affect training.
  2. The experimental practice is unusually careful. The paper uses paired seeds, common checkpoint rules, multiplicity correction, alternative test periods, and negative results, and it does not overstate the inconsistent gains from whitening.
Critical weaknesses
  1. The broad conclusion that the basis contributes little is tested most thoroughly with iTransformer, while the direct basis swap is concentrated on one development setting. The comparison with FreDF also uses the released operating point rather than a compute- and hyperparameter-matched search for most cells. Repeating the full decomposition on another backbone and dataset family would make the conclusion less architecture-specific.
  2. The evidence for whitening is statistically fragile. The selected setting is chosen on test error, most final runs use only three to five seeds, and rolling-origin evaluation is limited to ETTh1. The few gains that remain are concentrated in the same station/month family, so it is not yet clear whether they reflect a general covariance effect or a local regime.
Questions for authors
  1. Can the authors repeat the basis-versus-norm comparison, including validation-selected weighting strength, on PatchTST or DLinear and on a non-ETT dataset?
  2. Can WHITEDF hyperparameters be selected without using test error and then evaluated over rolling origins on at least one additional dataset?
  3. How sensitive are the FreDF conclusions to a matched tuning budget rather than the released reference configuration?
Reviewer_yBVgOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper analyzes three components changed by frequency domain losses (particularly FreDF): basis, norms and residual weighting. The authors claim that of the three components affected by the FreDF loss, the only one that is relevant is the ℓ1\ell_1 norm. The authors conclude that frequency domain losses must beat a time-domain ℓ1\ell_1 to be meaningful.

Critical strengths

N/A

Critical weaknesses
  • This is a diagnostic paper with a weak and unsurprising results: by the paper's own writing. "Figure 3: under ℓ1\ell_1 the basis adds little". The results are also limited to the iTransformer architecture, which is not state-of-the-art, limiting scope and relevance. The claimed "gain given by ℓ1\ell_1 is 0.9%", which is ~2x the noise level (line 236).
  • The theory is essentially restatements of elementary facts. e.g. Corollary one is simply Parseval's theorem. It is trivial to say that the residual's squared energy is the same in the time and frequency domain. It is not surprising that FreDF contributes little as a loss.
  • Very unclear and writing: heavy use of unknown jargon, unclear and undefined. For instance: "matched strength", "arms", "development cell", "surviving Holm", etc.
  • Questionable methodological choices call into question the whole work:
    • As written, it appears the WhiteDF setting was chosen on the development cell's test error, not its validation error: "Of the twenty ETT and Weather cells, ETTh1 at T = 96 is the development cell, on whose test error the WhiteDF setting was chosen, and the other nineteen are held out." L360:"The setting was chosen on this cell's test error, so the evidence for it comes from nineteen cells that did not influence it, to which Figure 4 applies it unchanged."
    • The use of Holm corrections is not justified.
Questions for authors
  1. "Whether whitening helps therefore deepnds on more than dimension and the linear sign." Could the authors expand on what do they mean by this? On what else does it depend?
  2. Can the authors explain the -15.3% negative gai on the Traffic dataset?
  3. "A time-domain ℓ1\ell_1 term at matched strength is therefore the baseline against which frequency- domain losses should be compare." What do the authors mean by "matched strength"?
  4. "The norm carries most of the gain, and the transform keeps a measurable part at the shortest hourly horizons." (Section 3.2) Could thet authors explain why would the transform help only at short horizons of the hourly datasets?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.