acceptodds
Under review as a conference paper at ICLR 2027

Compositional Discrete Forecasting with Orthogonal Frequency-Band Codebooks

Abstract

Long-term time-series forecasting must extrapolate recurring local structures whose frequency components reappear in different combinations. Continuous representations compose these factors flexibly but do not expose a finite vocabulary of reusable patterns. Existing discrete methods typically bind complete patches to single codes, tokenize transform coefficients, or retain continuous residual paths, leaving compositional codebook-based forecasting unresolved. We propose PatchVQFormer, which factorizes each patch into orthogonal frequency bands and represents it as a composition of frozen, band-specific waveform prototypes. Training-set energy and predictability determine the band partition, while band-dependent codebook capacities are selected on validation data. A spatial-then-temporal backbone uses masked future queries to predict all future band distributions in parallel. Training combines hard token classification with differentiable sequence reconstruction, whereas inference uses parameter-free soft expectations over frozen waveform prototypes. Across eight multivariate benchmarks and four horizons, PatchVQFormer achieves an average MSE/MAE of and ranks first in MSE and MAE comparisons. On the four ETT benchmarks at , it improves average MSE over the strongest listed competitor by . In controlled ETT ablations, frequency-band factorization reduces average MSE over the no-Fourier variant by . These results establish compositional frequency-band vocabularies as an accurate and structured approach for long-horizon forecasting without a continuous residual decoder.

Primary area:
learning on time series and dynamical systems
Keywords:
Long-term time series forecasting, discrete representation learning, vector quantization, frequency-band decomposition, multivariate time series forecasting

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_Cun9OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes PatchVQFormer. Each instance-normalized patch is split by fixed FFT masks into M orthogonal frequency bands; band boundaries come from a heuristic energy/autocorrelation score on training patches. For each band a k-means codebook of time-domain waveform prototypes is fitted offline and frozen. Historical patches enter the model only through prototype-indexed embeddings (summed across bands); a Transformer attends across variables and then across patches, with future query tokens built from positional and calendar information. Independent heads output a categorical distribution per band, and the forecast is the softmax-weighted expectation of the frozen prototypes, summed over bands and overlap-added. Training combines cross-entropy on hard target codes with a sequence reconstruction loss. Results are reported on eight benchmarks and four horizons against seven continuous baselines (values taken from the literature at differing input lengths), with ETT ablations of the Fourier split, the codebook and patching, a band-count study on ETTm2 and a capacity study on one ETTh2 setting. Short analytical statements (additive distortion across orthogonal bands, a nearest-prototype margin condition, monotone distortion in capacity) accompany the method.

Critical strengths
  1. An interesting representational idea. Factorizing a patch into orthogonal bands so that a small set of per-band prototypes yields a combinatorially large vocabulary, with exactly additive distortion across bands, is a neat way to give discrete time-series tokens compositional structure. If validated properly, it would be a useful design point between whole-patch codes (TOTEM-style) and coefficient tokens (WaveToken-style).

  2. Unusual frankness in the appendix. Appendix A.2 and B list the limits of the theory, the absence of multi-seed, cost and robustness measurements, and even code paths that conflict with the stated protocol. This transparency is what makes the paper assessable, and it would make a corrected resubmission easy to trust.

Critical weaknesses
  1. The main text and the appendix describe materially different methods and protocols, and the appendix discloses test-informed selection. I cannot tell which system produced Table 1.

    • Capacity selection. Abstract, Sections 3 and 4.1: band capacities are "selected on validation data". Appendix B.3: the main runs "do not sweep … on validation data"; capacities are script defaults or an "auto" rule, and they differ per dataset–horizon cell (Table 4).
    • History encoding. Section 4.1: one hard nearest-prototype index per band. Appendix B.5: a distance-weighted Top-K mixture of embeddings (K = 4, or 8 / 2 for "selected ETTh1 cells"). The stability result of Section 4.4 concerns the hard assignment that is not what is implemented, as the appendix itself notes.
    • Attention mask. Section 4.2: future queries attend to history and to all future queries. Appendix B.5: a future query attends to history and to itself only.
    • Objective. Section 4.3: L = L_CE + λ·L_seq with λ = 1. Appendix Eq. 9: cross-entropy + 2·(1 − cosine) + 8·MSE + 2·first-difference MSE, with per-cell overrides of the weights.
    • Temperature. Section 4.3: fixed, shared, "used unchanged in training and inference". Appendix: temperature 1 in training and a per-cell "test-time" temperature between 0.9 and 1.3 (Table 4), with no statement of how it was chosen.
    • Embeddings. Main text: trainable. Appendix B.4: frozen by default, trainable in three ETTm2 cells.
    • Model selection. Appendix B.2 states that the repository "contains a baseline-selection script that chooses the best input length by test MSE, and some configuration comments record test-informed tuning". The model's own input length varies over {96, 336, 512, 720} per cell (the L = 336 ablation gives ETTh1 average MSE 0.427, whereas the Table 1 ETTh1 rows average 0.404), patch length, band count, capacities, temperature, loss weights, batch size and width are all set per cell, there is one seed, and the test loader drops the last incomplete batch. The non-monotone errors in Table 1 (ETTh1 and ETTh2 are better at H = 336 than at H = 192) are consistent with cell-specific tuning. Under these conditions the reported numbers cannot be treated as held-out estimates.
  2. Even taken at face value, the results do not support the headline claims, and the central "fully discrete, no continuous decoder" claim is not tested.

    • Averaging the 32 rows of Table 1, PatchTST (2023) obtains 0.289 / 0.340 MSE/MAE, better than PatchVQFormer's 0.291 / 0.342 on both metrics; WPMixer has a better MAE (0.336). The abstract quotes PatchVQFormer's average without any baseline average. The method is the worst listed on Exchange at H = 96 and H = 720 and trails clearly on Weather and Electricity at long horizons; the "first-place count" is carried by ETT and Solar.
    • Baseline input lengths differ ("a broad literature comparison", in the authors' words), so the ETT long-horizon improvements of 1.0–1.8% are not controlled.
    • The paper is positioned against discrete forecasters (TOTEM, WaveToken, ReCast, Sparse-VQ, VQ-AR, Chronos) but compares with none of them. The question it poses, whether a codebook-only output path can match discrete alternatives, is therefore unanswered.
    • The output is a softmax-weighted linear combination of fixed vectors trained with a heavily weighted MSE term, and the input is a soft Top-K mixture. Functionally this is continuous regression onto a fixed dictionary, not discrete prediction. The appendix says the evaluation code records hard-argmax decoding and oracle-code reconstruction, but neither is reported; these are the numbers that would show whether the discrete vocabulary carries the forecast.
    • Removing the codebook costs 3.4% MSE on ETT and improves ETTh1 MAE; removing patching improves ETTm1; with one seed these effects are not distinguishable from noise. Several cells use P = H (e.g. 720), where the "patch" is the whole horizon and the compositional-reuse argument does not apply.
    • The analysis in Section 4.4 is the Pythagorean identity, a triangle-inequality margin and nestedness of feasible sets; it does not bear on forecasting accuracy.
Questions for authors

Clarifications that could change my assessment:

  1. Which description is correct for the Table 1 runs: Section 4 or Appendix B (hard vs. Top-K history, future-to-future attention, loss, temperature, frozen vs. trainable embeddings, capacity selection)? Please make the paper describe one system.
  2. How was the test-time temperature in Table 4 chosen for each cell? How was the input length chosen for each cell? Was any quantity in Table 1 selected using test metrics?
  3. Please report a fully held-out protocol: one input length for all methods, hyperparameters selected on validation only, ≥3 seeds, no dropped test batches, with baselines re-run under it.
  4. Please report hard-argmax decoding and oracle-code reconstruction error next to the soft-expectation result for every dataset.
  5. Please compare with TOTEM, WaveToken and ReCast under the same protocol.
  6. In "w/o Codebook", what replaces the codebook input and output paths?

Suggestion for future work (does not affect my rating): modeling the joint distribution over band codes and sampling from it would make the discrete formulation pay off for probabilistic forecasting.

Reviewer_mERhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on the problem of long-term time-series forecasting (LSTF). The authors propose PatchVQFormer, a compositional discrete forecaster that represents each patch as a sum of prototypes selected from orthogonal frequency-band codebooks. The authors conduct experiments on 8 benchmarks and 32 settings and demonstrate the effectiveness of the proposed solution.

The motivation and problem definition are clear. It provides a good survey of the related work. Also, the proposed framework seems overall reasonable. However, I am concerned that the proposed solution's accuracy improvement is modest and its effectiveness may be limited. Also, the font size of Figure 2 is too small, and it is hard to read.

Critical strengths

(S) The motivation and problem definition are clear

(S) The authors conduct experiments on 8 benchmarks and 32 settings and demonstrate the effectiveness of the proposed solution

Critical weaknesses

(W) The proposed solution's accuracy improvement seems modest, and its effectiveness may be limited

Questions for authors

N/A

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

The paper proposes PatchVQFormer, which represents time-series patches using frozen codebooks over orthogonal frequency bands. A Transformer predicts future code distributions, which are combined to reconstruct forecasts. Experiments on eight datasets show competitive performance, although inconsistencies between the main method and Appendix B make the exact experimental implementation unclear.

Critical strengths
  1. Band-specific waveform codebooks allow patterns to be reused across frequency combinations. The frozen prototypes and the absence of a continuous residual decoder also make the role of the discrete vocabulary easy to inspect.

  2. The evaluation covers eight datasets and four horizons, with good results on long-horizon ETT forecasting and Solar.

  3. The orthogonal decomposition, additive distortion identity, and hard-assignment stability result are valid under their stated assumptions.

Critical weaknesses
  1. The main text and Appendix B describe different historical token retrieval, future-attention masks, training losses, loss weights, and decoding temperatures. These are substantial modeling differences rather than minor implementation details. Without knowing which configuration produced Tables 1–3, it is difficult to assess or reproduce the paper's main claims.

  2. Appendix B.2 acknowledges a script that selects input length by test MSE and mentions test-informed tuning, but the paper does not identify which reported results used these procedures. Table 1 also compares models with different input lengths. Until the affected results are clearly identified and rerun with validation-only selection, the headline comparisons are not sufficiently reliable.

Questions for authors
  1. Under the same forecasting backbone and total prototype budget, how does the proposed band-specific tokenizer compare with whole-patch quantization and a simple equal-width frequency partition? This comparison would help isolate the benefit of the proposed compositional design and statistics-guided band discovery.

  2. Section 4.1 argues that separate codebooks support many cross-band compositions. What fraction of future hard-code tuples in the test set are unseen during training, and how accurate is PatchVQFormer on these unseen combinations?

  3. Appendix B.5 states that hard, Top-K, soft, and oracle decoding results are recorded. Can the authors report these results quantitatively? In particular, the gap between hard and soft decoding would show whether performance comes mainly from discrete prototype selection or from continuous interpolation within each codebook.

  4. The paper motivates PatchVQFormer by contrasting it with discrete forecasting methods such as TOTEM, WaveToken, and ReCast, but none is included in Table 1. Can the authors provide a controlled comparison with at least one codebook-based forecaster under the same datasets and input lengths? Without this comparison, the improvement over existing discrete forecasting methods remains unclear.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.