acceptodds
Under review as a conference paper at ICLR 2027

SPECTOR: Spectral Tree Operator Realization for Time-Series Priors

Abstract

Time series foundation models increasingly rely on synthetic data to supplement limited real-world corpora and to improve the scale and diversity of pretraining. Yet existing generators face a trade-off: compositional kernels offer rich diversity but require cubic-time covariance factorization, while fast parametric simulators scale linearly at the cost of narrower structure and weaker transfer. We introduce SPECTOR (SPECtral Tree Operator Realization), a realization compiler for compositional time-series priors that resolves this trade-off. SPECTOR treats a sum-product kernel expression as a stochastic-process program and recursively compiles its subtrees into efficient realization spaces, without materializing covariance matrices while enabling structural diversity and linear-time sampling. We formalize the semantics of this compiler through a covariance-preserving result for general realization trees, with spectral and feature-space semiring constructions providing stronger law-preserving guarantees for important kernel classes. With 18 kernel primitives, SPECTOR spans stationary, periodic, trending, transient, and stochastic-drift behaviors. Empirically, SPECTOR scales to long contexts and is hundreds of times faster than covariance-based kernel generators, enabling on-the-fly pretraining. Across six architectures and three benchmarks, SPECTOR achieves 74–91% pairwise win rates against competing synthetic priors and narrows the gap to published models pretrained on hybrid real-synthetic corpora.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, synthetic data generation

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_osxGOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes SPECTOR, a compiler for compositional time-series priors. Kernel subtrees are dispatched to spectral sampling, finite-rank features, Gaussian modulation, state-space simulation, or independent path composition, avoiding dense covariance factorization where the representation permits. The general construction preserves covariance; stronger Gaussian-law guarantees apply to suitable subclasses or asymptotically in the spectral feature count. The instantiated generator is evaluated through synthetic pretraining across six architectures and three forecasting benchmark suites.

Critical strengths
  1. The compiler offers an extensible way to generate diverse priors efficiently. The theory distinguishes covariance preservation from Gaussian-law preservation and states the conditions for linear scaling.

  2. SPECTOR leads all 18 architecture–benchmark aggregates. The gains remain with shared augmentation and when only the sampler changes under a fixed 10M-example budget (Tables 24–26). These controls make the downstream benefit convincing.

Critical weaknesses
  1. SPECTOR’s instantiated prior is manually specified through 18 primitives, sampling weights, composition rules, and augmentation settings such as SCM mixing, quantization, and spike injection. The bank and augmentation ablations are useful, but sensitivity to the specific weights and augmentation probabilities remains unclear. Testing alternative configurations would help establish how robust the results are to these design choices.
  2. All forecasting evaluations use univariate inference: multivariate targets are split into separate series, and exogenous covariates are ignored (p.6, L292–294). Although the SCM augmentation generates correlated series, its value for forecasting with cross-variable dependencies remains untested. A native multivariate or covariate-aware evaluation would strengthen the evidence for this component.
Questions for authors
  1. How sensitive are the forecasting gains to the primitive sampling weights and augmentation probabilities? Do reasonable alternative configurations retain comparable performance?
  2. Does the SCM augmentation benefit native multivariate or covariate-aware forecasting? A comparison with and without SCM in a setting that explicitly uses cross-variable information would help establish its value beyond univariate prediction.
2 comments
AuthorsOpenReview ↗

Response 1/2

We are grateful for your positive assessment of our work, and for recognizing the extensibility and efficiency of our framework together with its strong pretraining performance. Following your suggestions, we added a sensitivity analysis of SPECTOR to its base-kernel sampling weights and augmentation probabilities (Appendix J.11), and experiments with native multivariate and covariate-aware pretraining and inference (Appendix J.10). Our detailed point by point response appears below.

W1 + Q1. How sensitive are the forecasting gains to the primitive sampling weights and augmentation probabilities? Do reasonable alternative configurations retain comparable performance?

Following your suggestion, we ran experiments that vary the primitive sampling weights and augmentation probabilities to test how robust SPECTOR is to these hyperparameters. The bank has 18 base kernels, and to analyze the sensitivity in a more efficient way, we rescale all weights jointly with a temperature-style exponent, pk(α)∝wkαp_k^{(\alpha)} \propto w_k^{\alpha}: α<1\alpha<1 flattens the distribution toward uniform and α>1\alpha>1 concentrates it on the highest-weighted primitives. We trained PatchTST on four variants, using the training procedure in Appendix H.1. B1 and B2 flatten or sharpen the sampling weights, with α=0.5\alpha=0.5 and α=1.5\alpha=1.5 respectively. B3 lowers the SCM probability from 1.0 to 0.5, and B4 lowers the value-transform probability from 0.5 to 0.25. The table below (also added to Appendix J.11) reports aggregate skill on each benchmark (PatchTST, skill %, higher is better), with CauKer, the strongest baseline generator, for reference. All four variants remain above CauKer. Compared with SPECTOR default, the largest change in aggregate skill is 1.71 percentage points on M-Series, 1.21 on GIFT-Eval and 0.55 on FEV-Bench. No variant dominates: each improves on at least one benchmark and lowers skill on another. The default is therefore not a sharp optimum, and SPECTOR's advantage over other synthetic priors does not depend on the exact weights or augmentation probabilities.

ConfigurationM-SeriesGIFT-EvalFEV-Bench
SPECTOR (default, α=1.0\alpha=1.0, SCM p=1.0p=1.0, value transforms p=0.5p=0.5)20.9643.7239.99
CauKer (Table 1)13.2339.7636.59
CauKer with SPECTOR augmentations (Table 26)19.0741.8337.75
SPECTOR variants
B1: primitive weights, α=0.5\alpha=0.522.6742.5139.44
B2: primitive weights, α=1.5\alpha=1.519.5544.1840.08
B3: SCM p=0.5p=0.521.3942.9139.70
B4: value transforms p=0.25p=0.2521.5943.7739.73
AuthorsOpenReview ↗

Response 2/2

W2 + Q2. Does the SCM augmentation benefit native multivariate or covariate-aware forecasting? A comparison with and without SCM in a setting that explicitly uses cross-variable information would help establish its value beyond univariate prediction.

Thank you for the insightful question. Among the six architectures we study, Chronos-2 is the only one that supports both multivariate targets and covariates. To test whether the SCM structure in SPECTOR helps when the model actually uses cross-variable information, we pretrained two variants of Chronos-2 on the same SCM-generated SPECTOR series. They differ only in whether the SCM group structure is exposed to the model (we believe this is a cleaner setup, since removal of SCM from the generation pipeline would also change marginal distributions):

  • SPECTOR, 100% univariate: All SPECTOR series are passed as univariate series (no shared group IDs) during Chronos-2 training
  • SPECTOR, 50% univariate, 25% multivariate, 25% covariate: A training batch is composed of 50% univariate series, 25% multivariate groups and 25% covariate groups passed to Chronos-2 during training. For the multivariate and covaraite group, the series belonging to the same SCM share the same group IDs, randomly split into multivariate targets or covariates.

All models are evaluated on FEV-Bench with multivariate and covariate-aware inference, so the model receives all target channels and the provided covariates. The table below (added as Table 29, Appendix J.10) shows that mixed covariate/multivariate aware pretraining improves aggregate skill in all three task groups. The largest gain was on covariate-aware tasks (+5.33 points), with a smaller gain on multivariate tasks (+0.56 points). These preliminary results demonstrate a direct path from SPECTOR’s existing generator to multivariate pretraining. Designing richer dependencies between targets and covariates is a separate research question that warrants a dedicated study. We have updated the Limitations section accordingly.

TrainingInferenceUnivariate (32)Multivariate (26)Covariate (42)
SPECTOR, 100% univariatemultivariate + covariate33.6655.4638.40
SPECTOR, 50% univariate + 25% multivariate + 25% covariate (via SCM)multivariate + covariate35.53 (+1.87)56.02 (+0.56)43.73 (+5.33)

FEV-Bench SQL skill score (%, higher is better). Values in parentheses are gains/losses in percentage points over SPECTOR 100% univariate.

Summary

Thank you again for your detailed and positive review; we believe it has substantially improved our paper. The revised manuscript adds a sensitivity analysis to the base-kernel weights and augmentation probabilities (Appendix J.11) and native multivariate and covariate-aware pretraining and inference (Appendix J.10). We hope these additions address your concerns, and we would be glad to answer any further questions.

Reviewer_Zi64OpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses an efficiency bottleneck in synthetic data generation: time series priors built by adding and multiplying a small set of base kernels are highly expressive, but exact sampling requires factorizing the covariance matrix of every series, at a cost that grows cubically with sequence length and therefore becomes impractical for long-context pretraining. The authors treat such a kernel expression as an executable program and sample frequencies recursively along its syntax tree: at an addition node one branch is selected at random with probability proportional to the variance on each side, at a multiplication node the frequencies drawn from both sides are added, and the sampled frequencies are finally stacked into a sum of cosines. Because no covariance matrix is constructed or factorized, the sampling cost grows linearly with sequence length while the correlation structure of the generated series still matches the kernel defined by the original expression. Across six forecasting architectures and three zero-shot benchmarks, models pretrained on this generator consistently outperform other synthetic priors and remain practical at long context lengths.

Critical strengths
  1. The paper reduces the cost of sampling compositional kernel priors from cubic growth in sequence length to linear growth, and does so while preserving the compositional structure. Existing linear-cost approaches such as random Fourier features require the frequency distribution to be written in an explicit form; for a kernel obtained by adding and multiplying base kernels, however, the frequency distribution is a convolution of several distributions and usually has no explicit form. The authors instead draw frequencies recursively along the expression tree, selecting one branch at an addition node with probability proportional to the variance on each side and adding the frequencies drawn from the two sides at a multiplication node, so that sampling is completed without writing down the composite distribution or constructing a covariance matrix. This changes not only the speed but also the range of sequence lengths that can be generated, so that compositional kernel priors previously confined to short sequences can now be used for long-context pretraining. The scheme is also simple enough to implement to have the potential for wide adoption.
  2. The experiments offer good breadth of coverage and rigour of comparison. The authors evaluate six different forecasting architectures on three public benchmarks and obtain the best result in each of the eighteen architecture and benchmark combinations, with pairwise win rates between seventy and ninety percent. More importantly, they compare models pretrained on synthetic data alone against published checkpoints pretrained on mixed real and synthetic corpora, and in the ablations they apply the same set of augmentations to the baselines as a control.
Critical weaknesses
  1. Presentation and terminology

    • The paper places in-situ pretraining at the center of the abstract and introduction and lets it carry the main motivational argument, yet every experiment in the body is based on synthetic datasets with a fixed sample budget. There is no description of how generation is coupled to training, and no comparison between online generation and a fixed offline set. The term is borrowed from high-performance computing, where it means processing simulation data as they are produced. Here it seems refer to generating samples on demand during training, but how this differs from the on-the-fly generation and procedural data generation already in the literature is never made clear, leaving the reader to infer it.
    • Terminology departs from standard usage in the field. Throughout the paper the leaf kernels of the compositional grammar are called kernel primitives or primitive kernels, whereas the literature uses base kernels . More seriously, the term shifts in extension within the paper: the abstract restricts it to positive-semidefinite kernels, while later sections rename it stochastic primitives and include processes that share the covariance of white noise but have heavy-tailed marginals, so that one term comes to refer to two different classes of objects.
    • The density of programming metaphors raises the reading barrier. Words such as compiler, program, syntax, semantics, dispatch and fallback run through the paper; they are consistent with how the method is implemented, but they place an additional burden on readers with a statistical background and obscure the actual content behind formal language. The clearest illustration is that the two rules that matter most for sampling, namely selecting one branch at an addition node in proportion to variance and adding the frequencies at a multiplication node, appear only after a long development of semirings and homomorphisms, so that a reader like me is often still unsure how data are generated halfway through Section 2.
  2. Technical shortcomings

    • The coverage and novelty figures in Figure 2 are computed relative to other synthetic generators. Real data appear only as evaluation sets and never as a distributional reference, so the claim of a broader pretraining distribution rests on indirect evidence. In addition, both quantities being high is equally consistent with a prior that is simply wider and more dispersed, and the paper does not distinguish these two explanations using any quantitative distribution distance or structural statistic.
    • The stated contribution does not match the actual technical increment. Drawing several primitives, combining them at random by addition and multiplication, and then sampling from the combination already exists in KernelSynth; what this paper replaces is the final sampling step. Yet in-situ pretraining, the expanded kernel bank and the augmentations are all presented as contributions, while the genuinely new element, recursive frequency drawing along the tree, is compressed into a single proposition in Section 2.
    • The scope is restricted to univariate series and the risk of extrapolation is not discussed. All multivariate tasks are converted to univariate forecasting and covariates are discarded, whereas practical use of time series foundation models frequently involves multivariate series and covariates. The paper neither discusses how this restriction affects its conclusions nor indicates a path to the multivariate setting, so the claim of broad coverage of temporal structure holds only within the univariate case.
    • Reproducibility is insufficiently supported.
Questions for authors
  1. I would be grateful for clarification on how the training data are actually produced in the experiments. Are sequences generated on the fly at each training step, or generated once offline and then loaded from disk? If the former, it would help to see how generation throughput is matched to training consumption, for example how many series are required per training step and how many CPU workers are used; if the latter, it would be worth stating explicitly that in-situ pretraining is offered as a design motivation rather than as an experimental setting.
  2. The coverage analysis would be considerably more convincing if it were measured against real data rather than only against other synthetic generators, for instance by reporting the fraction of real training series that fall inside the SPECTOR density contour, or a quantitative distribution distance such as MMD, energy distance, or the AUC of a two-sample classifier, with the baseline generators reported alongside. Even a single benchmark would be informative here.
7 comments
AuthorsOpenReview ↗

Response 1/5

We are grateful for your thoughtful review and for recognizing the wide adoption potential of our method and experimental rigour. In our response and revised manuscript, we have added clarifications of terminology (Appendix A and Section 2), in-situ training setup and generation throughput measurements (Appendix B), contribution clarifications (Introduction and Method), coverage analysis in relation to real-world data (Appendix M), reproducibility statement, and evaluation of SPECTOR adaptation to multivariate and covariate-aware forecasting (Appendix J.10). Our detailed point by point response appears below.

W1 The paper places in-situ pretraining at the center of the abstract and introduction and lets it carry the main motivational argument, yet every experiment in the body is based on synthetic datasets with a fixed sample budget. There is no description of how generation is coupled to training, and no comparison between online generation and a fixed offline set.... how this differs from the on-the-fly generation and procedural data generation already in the literature is never made clear, leaving the reader to infer it.

Thank you for pointing this out. We respond to each concern in this review as below.

yet every experiment in the body is based on synthetic datasets with a fixed sample budget

To clarify, in the main results and in all ablations except the sampling-mechanism ablation, SPECTOR generates data on the fly at each training step rather than loading a fixed corpus. The sampling-mechanism ablation (Fig. 4a) uses a fixed 10M-series budget for both samplers, so that SPECTOR and Cholesky sampling are compared on identical data budgets, which is impractical to do on the fly for Cholesky sampling. We have revised the Baselines paragraph in Section 3.1 to describe how the data for SPECTOR and for each baseline generator are produced.

no comparison between online generation and a fixed offline set

We included a comparison in pretraining performance between SPECTOR on-the-fly data generation (OTF) and SPECTOR loading from a fixed corpus of 10M series (10M) in Table 28 (Appendix J.9), showing that Skill Scores differ by at most 0.86 percentage points across the six GiftEval and FEV-Bench subsets. We also added online generation pipeline Figure 6 coupling generation to training, in Appendix B.2 that presents SPECTOR walk-through examples.

how this differs from the on-the-fly generation and procedural data generation already in the literature is never made clear

We revised the Introduction adding: Here, in-situ pretraining refers to on-the-fly synthetic data generation, generating samples on demand during training.

W2 Terminology departs from standard usage in the field. Throughout the paper the leaf kernels of the compositional grammar are called kernel primitives or primitive kernels, whereas the literature uses base kernels . More seriously, the term shifts in extension within the paper: the abstract restricts it to positive-semidefinite kernels, while later sections rename it stochastic primitives and include processes that share the covariance of white noise but have heavy-tailed marginals, so that one term comes to refer to two different classes of objects.

Following you suggestion, in the revised manuscript, we replace primitives, kernel primitives or primitive kernels with base kernels and reserve stochastic processes for the elements supplied via the realization space to avoid confusion.

W3 The density of programming metaphors raises the reading barrier. Words such as compiler, program, syntax, semantics, dispatch and fallback run through the paper; they are consistent with how the method is implemented, but they place an additional burden on readers with a statistical background and obscure the actual content behind formal language....

Thank you for this concrete feedback. The compiler view is what makes SPECTOR flexible and extensible, as well as efficient. New priors can be built by adding a base kernel and its sampler, without new covariance machinery. The generation process is linear in sequence length, because compilation process allows us to parse the sampled kernel expression without ever materializing any covariance matrices. To address your concern, we have introduced terminology section at the beginning of Section 2 as well as Appendix A, where we define the key terms used in the programming and statistical fields. Additionally, to make the generation procedure easy to follow without the formal machinery, we added an illustration in Appendix B.2 (Toy Examples). It walks through Algorithms 1–4 on four small examples, from a single base kernel to the full pipeline. We now refer to the algorithmic implementation and toy examples early on in Figure 1 caption: SPECTOR's algorithmic implementation is provided in Appendix B and complete sampling pipeline walk through via toy examples appears in Appendix B.2.

AuthorsOpenReview ↗

Response 2/5

W4 + Q2 The coverage and novelty figures in Figure 2 are computed relative to other synthetic generators. Real data appear only as evaluation sets and never as a distributional reference, so the claim of a broader pretraining distribution rests on indirect evidence. In addition, both quantities being high is equally consistent with a prior that is simply wider and more dispersed, and the paper does not distinguish these two explanations using any quantitative distribution distance or structural statistic. The coverage analysis would be considerably more convincing if it were measured against real data rather than only against other synthetic generators, for instance by reporting the fraction of real training series that fall inside the SPECTOR density contour, or a quantitative distribution distance such as MMD, energy distance, or the AUC of a two-sample classifier, with the baseline generators reported alongside. Even a single benchmark would be informative here.

Following your suggestion, we evaluated each generator against real context windows, using both coverage and distributional distance metrics. We use 5,000 windows per source from M4 Monthly, GiftEval short, and FEV-Bench univariate.

Coverage: We measure contour recall as the fraction of real windows inside each generator's 95% kernel-density contour. This is the same metric as in Figure 2, now applied to real windows instead of baseline samples. We also report kNN coverage [1]: the fraction of real windows whose k-nearest-neighbor ball (k = 5) contains at least one synthetic window. The two metrics complement each other. A sufficiently dispersed prior can enclose real windows within a broad contour, so contour recall alone can reward a prior that is merely wider. kNN coverage credits a real window only when a synthetic sample lies within its local neighborhood, so it is robust to such dispersion.

For M4 Monthly, SPECTOR achieves the highest recall and coverage under both metrics, including the dispersion-robust kNN coverage (0.442 vs. 0.357 for CauKer, the best baseline). For GiftEval and FEV-Bench, SPECTOR and CauKer both have good recall and coverage, within 0.03 of each other.

Table 1. Recall/Coverage of real windows (higher is better).

GeneratorM4 MonthlyGiftEval (Short)FEV-Bench (Univ.)
Contour Recall ↑kNN Cov. ↑Contour Recall ↑kNN Cov. ↑Contour Recall ↑kNN Cov. ↑
SPECTOR0.9510.4420.6120.3270.5550.271
KernelSynth0.3610.2450.3460.1440.4400.088
CauKer0.8740.3570.6280.3100.5810.305
SarSim00.5750.2530.5020.2330.5080.209
ETS0.7990.2500.4190.2250.2920.153
ForecastPFN0.6610.2890.4580.2890.4050.307
Real data, held-out half (reference)0.9420.9690.9190.9350.8810.851

[1] Naeem, M. F., Oh, S. J., Uh, Y., Choi, Y., & Yoo, J. (2020). Reliable Fidelity and Diversity Metrics for Generative Models. ICML.

Distributional distance As requested, we also measure the distributional distance between each generator and the real data (Table 2). Distance metrics, however, penalize any synthetic pattern outside the benchmark's support. For a zero-shot pretraining prior, such patterns are desirable: the model must generalize to structures beyond any single benchmark. Accordingly, ETS (M4 Monthly) and ForecastPFN (GiftEval Short, FEV-Bench Univariate) are closest to real data yet are among the weakest generators in zero-shot forecasting (Table 1 of the paper).

Table 2. Distributional distance to real windows (lower is better).

GeneratorM4 MonthlyGiftEval (Short)FEV-Bench (Univ.)
MMD² ↓AUC ↓MMD² ↓AUC ↓MMD² ↓AUC ↓
SPECTOR0.2590.8040.1920.7810.2090.804
KernelSynth0.4460.9160.3300.8960.3430.904
CauKer0.2680.8310.1750.7830.1910.806
SarSim00.3000.8580.1870.8020.1880.783
ETS0.0550.7140.1720.7910.2170.793
ForecastPFN0.0940.7940.1080.7170.1350.796
Real data, held-out half (reference)0.0000.5060.0030.5120.0030.514

MMD²: Gaussian kernel. AUC: two-layer LSTM classifier (0.5 = indistinguishable).

AuthorsOpenReview ↗

Response 3/5

W5 The stated contribution does not match the actual technical increment. Drawing several primitives, combining them at random by addition and multiplication, and then sampling from the combination already exists in KernelSynth; what this paper replaces is the final sampling step. Yet in-situ pretraining, the expanded kernel bank and the augmentations are all presented as contributions, while the genuinely new element, recursive frequency drawing along the tree, is compressed into a single proposition in Section 2.

Thank you for this comment. We recognize that the original contribution statement did not make the scope of the technical advance sufficiently clear. We have revised the Introduction to clarify that SPECTOR changes the full computation from a sampled kernel expression to a generated path, eliminating covariance-matrix construction, composition, and factorization by carrying out composition directly in realization space. Describing this as replacing the "final sampling step" understates the scope of the change. Recursive frequency drawing handles stationary subtrees. Supporting mixed stationary/nonstationary trees and non-Gaussian leaves requires additional realization and composition rules, each supported by a corresponding correctness result in Section 2.

The original manuscript credits KernelSynth for the compositional grammar and attributes the augmentations to prior work. However, listing the augmentations alongside the compiler’s instantiation in the contribution statement could obscure their role, which we have clarified in the revision. The expanded bank demonstrates the compiler’s extensibility across 18 base kernels and process realizations. We also implement on-the-fly (in-situ) pretraining as an application enabled by the compiler’s efficiency, as detailed in our response to Q1.

Finally, our three contributions are the realization framework, its instantiation with a bank of 18 base kernels and process realizations, and the empirical results. We have revised the Introduction to make the scope and roles of our contributions unambiguous. The following points explain these contributions and their significance in detail.

  • A realization compiler for compositional kernel priors. Our method spans several complementary results in Section 2, each defining a component of the generation pipeline and establishing its correctness. SPECTOR changes the full computation from a sampled expression tree to a generated path. KernelSynth evaluates the expression via covariance-matrix construction and composition, then applies Cholesky factorization and sampling. SPECTOR carries out composition directly through spectral, feature-space, and time-domain realizations, eliminating covariance-matrix construction and factorization. Recursive frequency sampling handles stationary subtrees (Prop. 2.2, Thm. 2.3), with exact path covariance and finite-dimensional convergence guarantees (Prop. E.2). Extending the pipeline to mixed stationary/nonstationary trees and non-Gaussian leaves requires Gaussian basis expansions, Gaussian modulation, and direct path sampling and composition (Props. 2.1, 2.4–2.5).

  • Scalability and extensibility. The 18-entry bank demonstrates the range of processes that the realization rules support within the same compositional framework. Direct recursions generate Wiener and integrated-Wiener paths in linear time, while direct Student-t sampling captures heavy-tailed marginals beyond those obtainable through Gaussian covariance sampling. Generation remains linear in sequence length for fixed tree size and feature budgets, enabling on-demand generation of long sequences. Table 28 in Appendix J.9 and Figure 4b demonstrate the forecasting benefits of the expanded bank and long-sequence training on real-world benchmarks. Table 32 and the qualitative examples in Appendix L further show improved forecasting under synthetic regime changes and structural breaks.

  • Empirical gains. SPECTOR achieves the best aggregate skill among synthetic priors in all 18 architecture–benchmark comparisons. Table 28 in Appendix J.9 isolates the value of the sampling procedure. With KernelSynth’s kernel bank and the 10-million-example budget held fixed, replacing Cholesky sampling with SPECTOR’s sampler improves aggregate skill on all three benchmarks.

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

This paper proposes SPECTOR, an efficient realization framework for generating synthetic time series from compositional stochastic-process priors. Instead of constructing and factorizing large covariance matrices, SPECTOR compiles kernel expression trees into efficient sampling procedures using spectral, feature-space, and direct time-domain realizations. The framework supports a diverse set of stochastic processes with linear-in-length generation complexity under a fixed composition size and feature budget. Extensive experiments across six forecasting architectures and three benchmarks demonstrate strong zero-shot forecasting performance, as well as substantial improvements in generation efficiency over covariance-based sampling.

Critical strengths
  • An interesting and well-motivated direction. The paper addresses an important problem in time-series foundation model pretraining: how to generate sufficiently diverse synthetic data without making data generation a computational bottleneck. I particularly appreciate the idea of treating compositional kernel expressions as executable stochastic-process programs, which provides a flexible way to construct large-scale synthetic pretraining distributions.

  • Solid and comprehensive experiments. The experimental evaluation is quite convincing, covering six different forecasting architectures, three benchmark suites, and a wide range of synthetic data generators. Beyond the strong overall results, the authors provide useful ablations on sampling methods, kernel diversity, augmentation, and sequence length. I also appreciate the additional experiments on out-of-distribution periodicity and structural changes, which help explain what kinds of forecasting capabilities benefit from the proposed prior.

Critical weaknesses
  • The novelty of the core sampling mechanism needs to be better positioned. While the overall realization compiler is interesting, several of its key mathematical ingredients are closely related to existing work on Gaussian processes and random Fourier features. In particular, Kim and Teh (2016), Scalable Structure Discovery in Regression using Gaussian Processes, already discussed how spectral samples for sums and products of compositional kernels can be obtained through mixtures and sums of samples from the individual kernels. I would encourage the authors to discuss these connections more carefully and clarify what is fundamentally new in the spectral tree compilation, beyond organizing existing sampling identities and realization techniques into a unified framework.

  • The interesting finite-feature effect deserves a deeper investigation. One of the most surprising findings is that finite-feature realization consistently outperforms exact GP sampling via Cholesky, even with the same kernel bank and data budget. The authors suggest that this may be explained by spectral regularization, but the current analysis is mainly observational and does not fully establish the mechanism. Since finite-feature sampling preserves the target covariance in expectation, it would be particularly interesting to understand which higher-order distributional or per-sample spectral properties are responsible for the improved downstream generalization.

Questions for authors

What exactly makes finite-feature realization more effective than exact GP sampling for foundation model pretraining? Could the authors provide further experiments to distinguish the proposed spectral regularization effect from other distributional differences introduced by finite-feature sampling

4 comments
AuthorsOpenReview ↗

Response 1/2

We are grateful for your encouraging and thoughtful review, especially the acknowledgement of the flexibility of our framework and the comprehensive experimental evaluation. In response to your feedback, we added discussion on positioning of our method (Section 2) and added more theoretical and empirical analysis of the difference between Cholesky and SPECTOR sampling (Appendix K). Our detailed point by point response appears below.

W1. The novelty of the core sampling mechanism needs to be better positioned.

We thank you for pointing us to Kim & Teh (2016). We have revised Section 2 around Eq. (1) to credit their pairwise spectral sampling rules. Kim & Teh used them to construct a stochastic RFF upper bound on the negative log-determinant term of the GP log marginal likelihood. Our work builds a recursive sampler for compositional time-series priors on top of pairwise spectral sampling rules, and shows that the resulting paths are effective for foundation-model pretraining. Specifically, we contribute the following:

  • Spectral tree compilation. For stationary subtrees, recursive frequency draws exactly follow the composite spectral law of any finite sum-product tree, proved by structural induction (Prop. 2.2, Thm. 2.3). Its cost is O(nM+MT)\mathcal{O}(nM + MT) for nn leaves, MM frequencies and TT time points.

  • Guarantees on the generated paths (Prop. E.2). The covariance is exact at every finite M, finite-dimensional distributions converge to the GP law, and the convergence follows a Berry-Esseen O(M−1/2)\mathcal{O}(M^{-1/2}) rate. Kim & Teh's use relies only on an unbiased kernel estimate.

  • Mixed stationary/non-stationary trees. Finite-rank subtrees are compiled exactly in feature space (Prop. 2.4). Finite-rank × Gaussian products use Gaussian modulation (Prop. 2.5), which preserves the product kernel without tensor-product growth. It also preserves the GP law: exactly for Gaussian factors, and as M→∞M \rightarrow \infty for spectral factors. Direct realizations (Wiener, integrated Wiener, non-Gaussian leaves such as Student-t) compose with the other subtrees while preserving the target covariance (Prop. 2.1). Our full compiler therefore applies to arbitrary finite trees over the bank, preserving the covariance in general and the law on the subclasses above.

  • Evidence for synthetic pretraining. The full compiler makes on-the-fly sampling practical at long sequence lengths, and the resulting prior achieves the best aggregate skill among the synthetic generators in all 18 architecture-benchmark comparisons, narrowing the gap to published models pretrained on hybrid real-synthetic corpora.

AuthorsOpenReview ↗

Response 2/2

W2 + Q1. On finite-feature realization vs. exact GP sampling.

We appreciate this question as we have spent some time developing our understanding of it since submission. The difference between Cholesky and spectral sampling is about how higher frequency modes are introduced into the sample paths. Informally, the argument is as follows. A stationary kernel evaluated on a grid is symmetric Toeplitz matrix which, for large TT, would be approximately a Circulant matrix and hence diagonalizable by the Fourier basis. Since the spectral density is proportional to the eigenspectra, then when SPECTOR samples a mode, lower frequency modes have a higher likelihood of being present whereas Cholesky would contain all the modes. Please see section K in the appendix for a formal treatment. To validate empirically this theoretical argument, we sample 100K paths from the original KernelSynth bank (which consists of only stationary kernels except dot product) and compute the Fourier decomposition of each path. We consider that a frequency is present in a path if it is numerically not close to zero (relative to the observed largest frequency). The new appendix Figure 10 in the appendix shows how low frequencies are almost always present in all the sampled paths and that as we increase M a higher proportion of paths start to contain higher frequency elements. Given our theoretical argument this behavior is intuitive, as we increase M we sample more frequencies from the spectral density which then increases the amount of higher frequency elements that are sampled.

However, it remains unclear what constitutes the optimal distribution of frequency modes for forecasting pretraining, or whether other factors contribute to the observed improvements. We agree that understanding the underlying mechanisms is an important research direction. However, since our key contribution and main focus of the paper is the unified realization framework, we leave a systematic causal investigation of these sampling effects to future work.

Summary

Thank you again for your detailed and encouraging review; we believe it has substantially improved our paper. In the revised manuscript, we have provided new empirical and theoretical analyses comparing Cholesky and spectral sampling (Appendix K). We have made significant improvements to the writing to better position our contributions and clarify our method. We hope these additions address your concerns, and if so, we would be grateful if you would consider updating your assessment. We would be glad to answer any further questions.

Thanks for detailed rebuttal from the authors. I wil keep my positive scores

Dear Reviewer krvs,

Thank you for your positive assessment of our work!

Best regards,
The Authors

Comments on the paper

2 comments
Area_Chair_GX1BOpenReview ↗

Reminder: Please Participate Actively in the Discussion

Dear Reviewers,

Please actively participate in the discussion by responding to the authors’ replies and engaging with your fellow reviewers. This discussion is essential for clarifying concerns, resolving differences, and reaching a well-supported consensus on the final recommendation.

Thank you for your time and valuable contributions.

Best regards, AC

Dear Reviewers, we appreciate your thoughtful and constructive feedback and your recognition of the flexibility, extensibility, efficiency, and broad applicability of our approach, as well as its strong pretraining performance and comprehensive experimental evaluation. In response to your feedback, we have substantially strengthened the revised manuscript through:

  1. Positioning and contributions (Introduction, Section 2) Clarified that SPECTOR's key contribution is realization compiler that replaces the entire covariance-based generation pipeline with direct path-space composition. Its extensibility enables efficient generation from diverse stationary, nonstationary, and non-Gaussian processes, leading to significant improvement in forecasting performance across architectures and benchmarks.

  2. Efficiency and on-the-fly training (Appendix B, Section 3) Clarified that SPECTOR generates data on the fly with training times comparable to fixed corpora, whereas Cholesky sampling increases PatchTST-FM training time by 11.2× and 79.2× at sequence lengths of 1,024 and 4,096, respectively.

  3. Robustness to hyperparameters (Appendix J.11) Demonstrated that SPECTOR is robust to variations in base kernel sampling weights and augmentation probabilities, consistently outperforming the strongest baseline across all tested configurations.

  4. Coverage of real-world data (Appendix M) Demonstrated that SPECTOR has strong coverage of real-world time series, achieving the best coverage on M4 Monthly and competitive coverage on GiftEval and FEV-Bench.

  5. Multivariate and covariate-aware pretraining (Section 3, Appendix J.10) Demonstrated that SPECTOR can be natively adapted to multivariate and covariate-aware pretraining. Exposing the multivariate structure created by SCM augmentation to Chronos-2 during pretraining improves FEV-Bench skill by 5.33 pp on covariate-aware tasks and 0.56 pp on multivariate tasks. The univariate scope of the main study is now stated explicitly in the Limitations.

  6. Reproducibility (Reproducibility Statement) Added a reproducibility statement and further implementation details. Code to be released with the camera-ready version.

  7. Readability and terminology (Section 2, Appendices A-B) Section 2 now opens with the sampling rules in plain terms, before the theory. A terminology bridge and glossary map the compiler vocabulary to standard statistical notions, and Appendix B.2 adds a step-by-step walk-through on toy examples.

  8. Finite-feature vs. Cholesky sampling (Appendix K) Added theoretical and empirical analysis showing that, while both samplers have the same covariance, they differ in the frequency content of individual paths. A Cholesky path contains every frequency with spectral mass. A SPECTOR path contains at most MM frequencies, each present with probability given by its share of the variance. Low-frequency modes appear in almost every path, and high-frequency modes in a growing fraction of paths as MM increases.

We believe your feedback and these revisions strengthen both the technical foundations and the practical significance of SPECTOR, and we would be grateful if you would take them into account in your final assessment. We are happy to answer any further questions in the individual threads.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.