acceptodds
Under review as a conference paper at ICLR 2027

MarkovTS: Pretrained State–Transition Representations for Zero-Shot Time Series Generation

Abstract

Across heterogeneous real-world domains, time-series generation supports simulation, model development, and data augmentation by generating sequences that preserve target-domain statistics and temporal dependencies when real observations are scarce or costly. However, existing methods predominantly address in-domain generation, limiting their reuse in unseen domains. We study reference-conditioned zero-shot time series generation: target domains are unseen during pretraining, generation uses a single target-domain reference sequence, and no model parameters are updated. MarkovTS characterizes sequence evolution through local morphologies and their ordered transitions at multiple temporal scales. Its pretrained encoder quantizes normalized time-series patches into multiscale morphological states with a state codebook, while a transition codebook encodes contextual transitions between adjacent intervals. These transferable state–transition representations condition a pretrained flow-matching model. At generation, the frozen model combines a reference-derived condition with independent Gaussian noise to produce stochastic samples that retain the reference's temporal organization. Across 12 unseen target datasets, MarkovTS outperforms strong baselines in most distributional comparisons and generates diverse, reference-consistent sequences. As downstream augmentation, its samples yield overall gains across forecasting, classification, and anomaly detection benchmarks and diverse model families.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series, Time series representation, Time series generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_AxbnOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies reference-conditioned zero-shot time series generation, where a pretrained model generates samples from a single reference sequence in an unseen target domain. MarkovTS represents temporal structure through multiscale morphological states and contextual transitions, which condition a pretrained flow-matching generator. The paper reports evaluations on 12 unseen target datasets, together with data augmentation experiments for forecasting, classification, and anomaly detection, to assess generation quality, transferability, and downstream utility.

Critical strengths
  1. Broad reported empirical validation. The paper reports evaluation on 12 unseen target datasets and downstream augmentation across forecasting, classification, and anomaly detection. This breadth is valuable because it examines both generation quality and practical utility, providing support for the relevance of the approach across different applications.

  2. Clear motivation and coherent problem formulation. The paper connects target-domain data scarcity to a well-defined generation setting with a single reference and no parameter updates. The proposed multiscale state–transition representation follows naturally from this motivation, making the intended contribution and its practical significance clear.

Critical weaknesses

The single-reference conditioning mechanism limits the scope of generation. The method relies on one reference to supply target-domain information and explicitly aims to preserve its temporal organization. For domains containing multiple regimes or substantially different evolution patterns, this reference may provide incomplete coverage. Reference consistency and stochastic variation alone therefore do not establish that the generated samples capture the broader variability needed for simulation or effective augmentation.

Questions for authors

How do the authors characterize the variability that can be generated from a single reference? What evidence shows that the samples provide meaningful structural diversity while preserving reference consistency? Can author also provide few-shot performance, as reference is kind of one-shot

5 comments

To W1Q1: Structural Diversity, Single-Reference Conditioning, and Few-Shot Generation

Thank you for suggesting that we further characterize the range of variation under reference conditioning and the relationship between structural diversity and reference consistency. In its default setting, MarkovTS uses patch-level temporally aligned state–transition conditions to preserve the reference sequence’s temporal organization, with Gaussian sources providing variation in local morphology and amplitude. We adopt this design primarily for data augmentation in multivariate forecasting, classification, and anomaly detection, where the relative timing of patterns across variables matters.

First, the default generation setting demonstrates downstream utility. For multivariate forecasting, we compare MarkovTS with five generators, Gaussian perturbation, and Temporal Patch Shuffle (TPS). Across 30 backbone–dataset combinations, MarkovTS achieves the lowest average MAE and MSE, reducing the Real-only errors by 3.30% and 3.39%, respectively. Relative to the alternative augmentation methods, its reductions range from 2.01% to 4.73% for MAE and 2.46% to 6.26% for MSE. MarkovTS uses three-run means, and the alternatives use one run per setting.

Training dataMAE ↓MSE ↓
Real only0.35680.3511
Real + MarkovTS0.34500.3392
Real + TimeDP0.35540.3538
Real + ImagenFew0.35330.3502
Real + Diffusion-TS0.36210.3618
Real + TimeVQVAE0.35620.3548
Real + TimeVQVAE-C0.35600.3542
Real + Gaussian 0.20.35210.3503
Real + TPS0.35230.3477

The generated data also improve aggregate performance in multivariate classification and anomaly detection. These results support the practical augmentation value of the current reference-aligned generation setting.

For few-shot evaluation, we further consider support sets of 3, 10, and 100 target-training windows on Stock and Web, both unseen during MarkovTS pretraining. Zero-shot denotes transfer to unseen target domains without parameter updates, and the number of references specifies the information budget at inference. MarkovTS keeps its parameters fixed and conditions each output on one reference from the support set. The following table reports three-run mean MMD / flat-KL, both lower-is-better.

Support windowsStock, MarkovTSStock, TimeDPWeb, MarkovTSWeb, TimeDP
30.085 / 1.0280.101 / 1.4100.077 / 0.1510.087 / 0.344
100.059 / 0.4260.111 / 0.7410.061 / 0.1340.077 / 0.182
1000.009 / 0.0250.077 / 0.1140.007 / 0.0160.026 / 0.071

MarkovTS improves both metrics over TimeDP’s frozen domain prompting at every support size. The appendix also provides comparisons with other generators and limited-support experiments on the 12 generation benchmarks.

We agree with the reviewer’s emphasis on broader structural diversity. We understand structural diversity to include changes in phase, pattern placement, occurrence intervals, and pattern combinations while retaining recognizable reference patterns. OATS[1] offers a valuable exploration of this direction. It uses TimeDP as its generative backbone, trains on multi-domain subsets of LOTSA-100M, and combines example guidance with subset-class conditioning to generate sequences with varied patterns. Its online augmentation improves forecasting performance in time series foundation models, demonstrating the value of reference-guided generation for broader training-data diversification.

The default outputs already exhibit measurable shape variation, with average Shape RMS and Shape APD of 0.314 and 0.302, respectively. These values are similar in magnitude to TimeVQVAE’s 0.452/0.371 and TimeVQVAE-C’s 0.457/0.380. The additional patch-aligned visualizations also show waveform variation under a fixed reference condition. We further examine how to control generated structure and expand the range of variation supported by reference conditioning.

To this end, we introduce State–Transition Recombination (STR), a variant of codebook conditioning in MarkovTS. Given a single reference of length LrefL_{\mathrm{ref}}, STR samples NN windows of length LL to form a reference-window pool, allowing overlap between windows. STR then recombines state–transition code blocks at patch boundaries shared across scales, using the same source window across scales within each block. The frozen flow-matching generator produces a complete sequence from the recombined condition and a Gaussian source. The multi-reference setting further allows selection of reference sources and code positions at each scale, combining temporal patterns from different references.

The new visualizations show changes in pattern placement, local morphology, amplitude, and occurrence intervals while retaining recognizable reference patterns. In the multi-reference comparison, we fix the references, source assignments, and Gaussian inputs, and vary the code positions to examine changes in generated structure. The Supplementary Material provides an extended collection of single-reference STR 16 and STR 32 examples, displaying the source references, multiple input crops, and generated samples side by side across diverse datasets.

In the single-reference experiments, we set Lref=1024L_{\mathrm{ref}}=1024 and L=512L=512, recombine 256-point code blocks, and use window-pool sizes N=16N=16 and N=32N=32. These settings correspond to MarkovTS (STR 16) and MarkovTS (STR 32) below. Across 12 datasets, we evaluate Shape RMS, Shape APD, and ACF/PSD PRDC Coverage. Shape RMS measures the shape difference between an output and its conditioning reference, Shape APD measures differences among outputs sharing a reference, and ACF/PSD PRDC Coverage measures coverage of real autocorrelation and spectral structures. The summaries weight the datasets equally, and Coverage uses 100 candidates per reference.

Model and settingShape RMSShape APDACF/PSD PRDC Coverage ↑
MarkovTS (default)0.3140.30241.25%
MarkovTS (STR 16)1.3391.32432.92%
MarkovTS (STR 32)1.3261.34032.62%
ImagenFew0.6050.69341.23%
TimeDP1.0081.00855.20%
Diffusion-TS0.8250.6349.66%
TimeVQVAE0.4520.37144.30%
TimeVQVAE-C0.4570.38045.08%
Reference replication0.0000.00024.00%
Gaussian 0.20.1960.27627.39%

STR increases average Shape APD from 0.302 in the default setting to 1.324 and 1.340, with structural coverage exceeding reference replication and Gaussian perturbation. Together with the visualizations, these results demonstrate how recombining reference conditions expands shape variation and changes pattern placement and combinations. Single-reference STR combines patterns within that reference, and multi-reference STR further combines patterns supplied by different references.

Evaluation using the main benchmark’s MMD and flat-KL shows that STR expands structural variation and maintains competitive distributional matching. STR 16 and STR 32 achieve average MMD of 0.0674 and 0.0673, compared with the default MarkovTS score of 0.0605. Their average flat-KL scores of 0.1908 and 0.1891 are comparable to the default score of 0.1905. Both STR settings achieve lower averages on both metrics than the nine comparison generators in Table 1.

Model and settingAverage MMD ↓Average flat-KL ↓
MarkovTS (default)0.06050.1905
MarkovTS (STR 16)0.06740.1908
MarkovTS (STR 32)0.06730.1891
Winformer0.06910.1940
ImagenFew0.08500.4201
TimeDP0.07440.2123
Diffusion-TS0.09510.4253
TimeGAN0.41352.4536
GT-GAN0.36181.3796
TimeVAE0.36071.3715
TimeVQVAE0.18580.8419
TimeVQVAE-C0.09960.9073

Dataset-level changes further illustrate the effects of condition recombination. On Temperature, both STR settings increase MMD from 0.134 to 0.161, with flat-KL increasing from 0.098 to 0.138 and 0.139. On Air, MMD decreases from 0.015 to 0.010 and flat-KL from 0.026 to 0.020, improving both metrics. The effects are therefore dataset-dependent. For example, NN5 MMD increases from 0.145 to 0.189 and 0.186, while flat-KL decreases from 0.045 to 0.040 and 0.039. Electricity flat-KL increases from 0.001 to 0.079 and 0.080. These results characterize how the structural variation introduced by STR affects matching to specific target distributions.

The current experiments support MarkovTS’s augmentation advantage over alternative generators and augmentation methods in in-domain multivariate forecasting, and demonstrate its augmentation value in classification and anomaly detection. Systematic comparisons among generators on classification, anomaly detection, and broader foundation forecasting model training remain for future work. The revision introduces State–Transition Recombination at the end of the method section and reports distributional matching results, structural diversity metrics, and generation visualizations in the experiments.

The appendix also provides multi-reference and cross-domain generation examples, such as combining Exchange and Traffic signals through state–transition conditions to generate sequences that retain source patterns and exhibit new pattern combinations. Future work will investigate diverse time series generation and fine-grained editing controlled through state–transition conditions.

[1] OATS: Online Data Augmentation for Time Series Foundation Models. NeurIPS 2026.

Manuscript revisions: Section 4.5 (State–Transition Recombination); Sections 5.2 (Generation Quality), 5.4 (Downstream Utility), and 5.5 (Structural Diversity, Table 8); Appendices B.7 (STR Details), D.2–D.3 (Evaluation Protocol and Metrics), E.1.2 (Shape Metrics and Structural Coverage, Tables 24–25), E.2.4 (STR Distributional Results, Table 28), F.1–F.2 (Limited-Support Comparisons), G.5 (STR Visualizations, Figures 14–16), and H.3 (Alternative Augmentation, Table 36). Supplementary Material: markovts_str_samples.pdf (extended single-reference STR 16/32 generation examples).

Reviewer_7EQWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MarkovTS, a pretrained framework for reference-conditioned zero-shot time series generation. The method learns multiscale morphological states and contextual transitions through two separate codebooks, which are subsequently used to condition a pretrained flow-matching generator. Ultimately, given single reference from unseen target domain, MarkovTS could generate new sequences without parameters updating. Evaluations on various datasets and tasks demonstrate that MarkovTS achieves superior performance compared with existing methods.

Critical strengths
  1. Generating time series from unseen domain using only single reference and without parameter updates is a challenging setting. It is practically meaningful in data-scarce scenarios.
  2. The experimental evaluation is extensive. Besides evaluating generation performance on multiple datasets with different metrics, the paper also includes representation-level intervention experiments and investigates downstream applications in forecasting, classification, and anomaly detection.
Critical weaknesses
  1. The proposed framework mainly combines existing techniques, including vector quantization, multiscale representations and flow matching. The claimed advantages of state-transition decomposition over conventional representations are not sufficiently established. Furthermore, the connection between the proposed codebook and actual Markov process remains unclear, while the theoretical analysis provides limited support for cross-domain transferability.
  2. Despite the rich set of experiments, it lacks depth in validating. The downstream experiments mainly compare performance w/ and w/o MarkovTS, without sufficiently comparing alternative generation or augmentation methods. Moreover, MarkovTS processes variables independently without explicitly validating cross-variable dependencies metric to support its rationality.
Questions for authors
  1. Why baselines are conducted across different target-domain training and adaptation protocols? Could the authors provide controlled comparisons under identical single-reference and without fine-tuning conditions?
  2. Why length-512 generated trajectories are evaluated primarily through length-168 windows? How does the method perform on full-sequence metrics?
  3. What is the precise connection between the proposed transition codebook and Markov process?
  4. Could cross-variable dependencies been preserved while variable is processed independently in MarkovTS? Please report joint multivariate distribution fidelity of generated results.
  5. Why downstream experiments do not compare against alternative generators or conventional augmentation strategies?
5 comments

To W1Q3: The Contribution of State–Transition Decomposition and Its Connection to Markov Modeling

Thank you for asking us to clarify the contribution, the motivation for state–transition decomposition, and its connection to Markov modeling. MarkovTS contributes a transferable multiscale state–transition representation pretrained for single-reference zero-shot time series generation. Its motivation is that locally normalized morphological states do not uniquely determine how adjacent intervals change. Identical state pairs can occur with different relative levels, amplitudes, and preceding contexts. Accordingly, the state codebook encodes normalized local morphology, while the transition codebook encodes residuals between continuous patch-pair relations and their state-pair baselines, incorporating preceding context through causal attention (Eq. 6). Both codebooks are pretrained and reused across target domains, and each reference expresses its distinctive temporal structure through their ordered multiscale assignments.

The connection to Markov modeling lies in the state–transition description of sequence evolution. Classical Markov models describe evolution through states and their transition probabilities. MarkovTS adopts this state–transition perspective to construct generation conditions. Each transition code represents a contextual residual change between adjacent intervals, and transition queries can access the full preceding context. The resulting state–transition code sequence conditions the flow-matching generator, with Gaussian sources providing stochastic variation among samples. We use Markov-inspired to describe this conceptual connection in the representation design.

To evaluate the advantage over conventional representations, Patch-Flow uses a standard patch Transformer and a single codebook while retaining the remaining model and training settings. Under the paired five-seed comparison, MarkovTS obtains lower mean MMD and flat-KL on 11 of 12 datasets (Table 3 and Appendix E.2). Figure 5 further directly examines the frozen encoder. Within identical ordered state pairs, different transition codes distinguish changes in level, amplitude, slope, and patch boundaries. Within-code dispersion decreases by 18.1%–29.7% relative to count-preserving label permutations across the five scales, with all five 95% bootstrap confidence intervals excluding zero. These results support the generation benefits of the representation and the complementary information supplied by transition codes.

Regarding cross-domain transferability, all evaluation target domains are unseen during MarkovTS pretraining, and the encoder, both codebooks, and flow-matching generator remain fixed during target-domain generation. Beyond the 12 generation benchmarks, additional Stock and Web experiments demonstrate competitive distributional matching against generators using domain prompting or target-domain adaptation. The same pretrained MarkovTS also generates additional training data in a zero-shot manner, yielding overall improvements across multiple model families in forecasting, classification, and anomaly detection. Together, these results support the transferability of the learned representations and the practical utility of zero-shot generation. Section 4.2 clarifies the Markov connection, while the revised experiments present the representation comparisons, complementarity analysis, and cross-domain evidence.

Manuscript revisions: Sections 4.2 (Multiscale State–Transition Representation Pretraining), 5.2–5.4 (Generation Quality, State–Transition Analysis, and Downstream Utility); Appendices B–C (Representation Objectives and Analysis), E.2 (Code Interventions, Table 27), G.1 (State–Transition Complementarity), and F.1–F.2 (Cross-Domain and Limited-Support Comparisons).

To Q1: Fairness of Generator Training, Target-Domain Adaptation, and Comparison Protocols

Thank you for highlighting the need to clarify the comparison protocols. The main benchmark follows the TimeDP and Winformer evaluation settings, including the target test splits, evaluation length, and metrics. Each baseline follows its established target-domain training or adaptation procedure. Among the evaluated implementations, TimeDP supports frozen domain prompting, and the other baselines use target-domain training or fine-tuning. We additionally examine limited-data settings. All target datasets are unseen during MarkovTS pretraining, and its parameters remain frozen.

First, on the 12 generation benchmarks, baselines use 10% or 20% of each designated target training split for prompting or fine-tuning. The table reports equal-weight dataset averages. MarkovTS uses its fixed five-seed benchmark result at both support levels, and the baseline settings use one seed.

Method10% MMD ↓10% flat-KL ↓20% MMD ↓20% flat-KL ↓
MarkovTS, fixed benchmark0.06050.19050.06050.1905
TimeDP0.13850.29360.14040.2976
ImagenFew0.11510.34330.11990.4389
Diffusion-TS0.11860.36450.14650.3416
TimeVQVAE0.09100.26930.09920.3291
TimeVQVAE-C0.09430.28380.09540.3112

The fixed MarkovTS benchmark obtains lower average MMD and flat-KL than the baselines at both support levels.

Second, on Stock and Web, all methods receive the same support sets of K=3, 10, or 100 target training windows. MarkovTS conditions each output on one reference with frozen parameters, and TimeDP constructs a frozen domain prompt from the support set. The remaining generators are fine-tuned on that set. Each setting produces 2,000 length-168 evaluation windows against the same real test windows. Cells report mean MMD / flat-KL over seeds 42, 2026, and 2027, with support sets fixed across seeds. Both metrics are lower-is-better.

Stock

MethodK=3K=10K=100
MarkovTS0.085 / 1.0280.059 / 0.4260.009 / 0.025
TimeDP, prompting0.101 / 1.4100.111 / 0.7410.077 / 0.114
ImagenFew, fine-tuned0.252 / 1.2960.077 / 0.3010.086 / 0.193
Diffusion-TS, fine-tuned0.652 / 2.0550.109 / 0.9250.079 / 0.185
TimeVQVAE, fine-tuned0.062 / 1.3690.066 / 0.6280.026 / 0.078
TimeVQVAE-C, fine-tuned0.062 / 1.5320.065 / 0.6600.010 / 0.056

Web

MethodK=3K=10K=100
MarkovTS0.077 / 0.1510.061 / 0.1340.007 / 0.016
TimeDP, prompting0.087 / 0.3440.077 / 0.1820.026 / 0.071
ImagenFew, fine-tuned0.082 / 0.2840.032 / 0.1360.039 / 0.151
Diffusion-TS, fine-tuned0.075 / 0.3960.097 / 0.1860.038 / 0.051
TimeVQVAE, fine-tuned0.052 / 0.3130.023 / 0.1650.004 / 0.047
TimeVQVAE-C, fine-tuned0.051 / 0.2860.016 / 0.1480.002 / 0.031

MarkovTS achieves lower MMD and flat-KL than frozen TimeDP at every support size on both datasets. It also remains competitive with fine-tuned generators, including the lowest flat-KL across all Web settings. TimeVQVAE-C achieves lower MMD on Web. These results support effective generation with frozen MarkovTS parameters under limited target data. We clarify the training and adaptation protocols in the experimental setup and provide dataset-wise comparisons in the appendix.

Manuscript revisions: Section 5.1 (Experimental Setup); Appendices D.2 (Evaluation Protocol) and F.1–F.2 (Stock/Web and Limited-Support Comparisons, Tables 32–33).

To Q2: Evaluation of Complete Generated Sequences beyond Short Benchmark Windows

Thank you for raising this question. The main comparison uses length-168 windows to follow the TimeDP and Winformer benchmark settings and enable direct comparison with existing results. MarkovTS targets zero-shot generation in data-scarce settings and operates at internal generation lengths of 512–4096. The minimum length of 512 follows from the multiscale encoder, whose coarsest scale requires two consecutive 256-point patches to construct state–transition representations. We therefore generate length-512 sequences and evaluate their length-168 windows under the common benchmark protocol.

To address full-sequence performance directly, we additionally evaluate complete generated sequences at lengths 512, 1024, and 2048, without reducing them to length-168 windows. The following results report Shape RMS, Shape APD, ACF/PSD Coverage, MMD, and flat-KL for the default MarkovTS and its STR variants.

LengthMethodShape RMSShape APDACF/PSD Coverage ↑MMD ↓flat-KL ↓
512MarkovTS0.3370.32736.72%0.2040.092
512STR 161.4381.34743.86%0.1170.094
512STR 321.4421.37242.94%0.1170.095
1024MarkovTS0.3290.31833.28%0.0280.099
1024STR 161.4211.34937.25%0.0100.101
1024STR 321.4231.37636.18%0.0100.102
2048MarkovTS0.3260.31631.10%0.0520.091
2048STR 161.4201.35134.99%0.0190.091
2048STR 321.4201.37634.43%0.0190.091

In this exploratory evaluation, STR exhibits greater shape variation and higher ACF/PSD Coverage, together with lower MMD and similar flat-KL relative to the default setting at each length. We present these measurements as supplementary full-length diagnostics rather than a controlled cross-method benchmark. Reference selection is stochastic, the evaluated segmentations differ across lengths, and the baseline implementations do not provide directly matched long-sequence outputs under this protocol. Thus, the length-168 benchmark remains the basis for baseline comparisons, and the full-length Coverage values should not be compared directly with those in Table 8. The Supplementary Material provides full-length waveform visualizations at 512, 1024, and 2048 points, comparing the reference, MarkovTS, Gaussian perturbation, STR 16, and STR 32.

Manuscript revisions: Section 5.1 (Experimental Setup); Appendices B.1 (Aligned Patchification), D.2 (Generation Evaluation Protocol), and E.3 (Generation Across Sequence Lengths, Table 29, evaluated with length-168 windows). Supplementary Material: markovts_full_length_512-1024-2048.pdf (full-length generation visualizations at 512, 1024, and 2048 points); full-length diagnostic metrics are reported above.

Reviewer_U2zPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies zero-shot time-series generation from a single reference: a model pretrained on a large corpus generates new sequences for an unseen dataset from one reference window, without any fine-tuning. On 12 held-out datasets, the model gets the best MMD and flat-KL in most comparisons against baselines trained on the target data. Reference-swap tests show that outputs follow the given condition, and the generated data is used as augmentation for forecasting, classification and anomaly detection.

Critical strengths
  1. The fixed-noise reference-swap test is a nice and clean way to check that the generator really uses the condition and does not just output generic samples. The state–transition analysis with permutation controls is also careful.
  2. The setting is well-defined: : one frozen pretrained model, one reference, no target-specific training. Evaluating on 12 datasets plus three downstream tasks gives a broad picture.
Critical weaknesses
  1. The main metrics cannot separate generation from copying. In the appendix, simply copying the reference gets MMD close to or better than MarkovTS on several datasets (e.g., Traffic, Taxi, Air), and it is also competitive with many baselines in Table 1, yet this control is not in the main table. Since samples stay close to their reference (high retrieval accuracy, 95% bands mostly covering the reference) and there is no check of distance to the nearest reference, it is unclear how new the generated sequences really are.
Questions for authors
  1. How does MarkovTS compare with TimeDP and other pretrained generators in their own unseen-domain or few-shot settings?
  2. Can you run the forecasting experiments with several seeds?
3 comments

To W1: Reference Consistency, Distinguishing Generation from Replication, and Structural Diversity

Thank you for raising this concern about reference consistency and downstream augmentation, and the diversity of the generated samples. Reference replication can achieve low MMD because references and test samples originate from the same target dataset. We therefore supplemented the evaluation with reference distances, diversity measures, and structural coverage.

First, our emphasis on reference consistency serves multivariate downstream augmentation. Our forecasting, classification, and anomaly detection experiments involve multivariate sequences, where temporal correspondence across variables carries task-relevant dependencies. We therefore preserve strict patch-level temporal alignment in the reference-derived conditions used during generation. This design encourages generated samples to remain close to the reference’s temporal organization.

Second, we quantified reference consistency and sample variation across 12 datasets, comparing MarkovTS with five generative baselines, reference replication, and Gaussian perturbation, with 100 candidates per fixed reference for each method. Shape RMS measures the distance to the conditioning reference, while shape APD measures differences among outputs sharing that reference. Both measures independently z-standardize each sequence. Closeness to the reference varies across models and datasets. For example, on Temperature and NN5, MarkovTS obtains shape RMS of 0.453 and 0.560, compared with 0.342 and 0.370 for TimeVQVAE. Meanwhile, its shape APD is 0.442 and 0.516, compared with 0.276 and 0.275 for Gaussian perturbation, demonstrating stochastic variation under a fixed reference.

Furthermore, we evaluated coverage of real temporal structure using PRDC Coverage in an ACF/PSD feature space. This measures the proportion of real held-out windows whose feature neighborhoods are covered by generated samples. As the number of candidates per reference increases from 1 to 10 and 100, MarkovTS’s coverage increases from 3.84% to 19.46% and 41.25%. With 100 candidates, it exceeds reference replication and Gaussian perturbation by 17.25 and 13.86 percentage points, respectively, with coverage comparable in magnitude to ImagenFew and the TimeVQVAE variants. These results show that repeated generation expands coverage of real temporal structure while maintaining reference consistency.

The table summarizes equal-weight averages across 12 datasets from the two analyses. Shape RMS averages dataset-wise medians.

MethodShape RMSShape APDACF/PSD PRDC Coverage
MarkovTS0.3140.30241.25%
ImagenFew0.6050.69341.23%
TimeDP1.0081.00855.20%
Diffusion-TS0.8250.6349.66%
TimeVQVAE0.4520.37144.30%
TimeVQVAE-C0.4570.38045.08%
Reference replication0.0000.00024.00%
Gaussian 0.20.1960.27627.39%

We further conduct a nearest-reference and nearest-training-window analysis under a common five-candidate center-window protocol. Across 2,783 common outputs per method, MarkovTS obtains a nearest-reference shape RMS of 0.318, a nearest-native-training-window shape RMS of 0.315, and 0.00% exact training-window matches. In contrast, reference replication has zero distance to both sets and 100.00% exact matches (Appendix E.1.2, Table 26). This provides a direct check against literal replication alongside the diversity and coverage measures. The Supplementary Material also presents extended fixed-reference visual comparisons among MarkovTS, the generative baselines, reference replication, and Gaussian perturbation.

Manuscript revisions: Sections 5.2 (Generation Quality) and 5.5 (Structural Diversity); Appendices D.2–D.3 and E.1.2 (Tables 23–26). Supplementary Material: fixed_patch_reference_K100.pdf (extended fixed-reference generation and replication comparisons).

To Q1: Unseen-Domain Generation and Comparisons under Limited Target Data

Thank you for suggesting this comparison. We evaluate unseen-domain generation on Stock and Web, alongside few-shot target-domain adaptation on the 12 generation benchmarks. All target datasets are excluded from MarkovTS pretraining, and its parameters remain frozen throughout these comparisons.

First, on Stock and Web, we use K=3, 10, or 100 support windows from the target training split. MarkovTS and TimeDP perform frozen reference-conditioned generation and domain prompting, respectively, while ImagenFew, Diffusion-TS, TimeVQVAE, and TimeVQVAE-C are fine-tuned on the corresponding support sets. Each setting generates 2,000 univariate sequences of length 168 and uses the same test windows. The tables report mean MMD / flat-KL over seeds 42, 2026, and 2027, with support sets fixed across seeds. Reference replication reports deterministic point estimates. Both metrics are lower-is-better, and bold indicates the lowest displayed mean.

Stock

MethodK=3K=10K=100
MarkovTS0.085 / 1.0280.059 / 0.4260.009 / 0.025
TimeDP, prompting0.101 / 1.4100.111 / 0.7410.077 / 0.114
ImagenFew, fine-tuned0.252 / 1.2960.077 / 0.3010.086 / 0.193
Diffusion-TS, fine-tuned0.652 / 2.0550.109 / 0.9250.079 / 0.185
TimeVQVAE, fine-tuned0.062 / 1.3690.066 / 0.6280.026 / 0.078
TimeVQVAE-C, fine-tuned0.062 / 1.5320.065 / 0.6600.010 / 0.056
Reference replication0.086 / 1.4390.070 / 0.6160.011 / 0.053

Web

MethodK=3K=10K=100
MarkovTS0.077 / 0.1510.061 / 0.1340.007 / 0.016
TimeDP, prompting0.087 / 0.3440.077 / 0.1820.026 / 0.071
ImagenFew, fine-tuned0.082 / 0.2840.032 / 0.1360.039 / 0.151
Diffusion-TS, fine-tuned0.075 / 0.3960.097 / 0.1860.038 / 0.051
TimeVQVAE, fine-tuned0.052 / 0.3130.023 / 0.1650.004 / 0.047
TimeVQVAE-C, fine-tuned0.051 / 0.2860.016 / 0.1480.002 / 0.031
Reference replication0.134 / 0.4660.076 / 0.1890.010 / 0.038

MarkovTS achieves lower MMD and flat-KL than frozen TimeDP and reference replication at every support size on both datasets. At K=100, its MMD/flat-KL scores are 0.009/0.025 on Stock and 0.007/0.016 on Web. Other baselines also perform strongly in some settings, with TimeVQVAE-C obtaining lower MMD on Web.

Second, on the 12 generation benchmarks, the baselines use 10% and 20% of the target training split for prompting or few-shot fine-tuning. The table reports equal-weight averages across datasets. The MarkovTS row uses the five-seed results from Table 1 as a fixed reference for both support levels, with average MMD/flat-KL of 0.0605/0.1905. The baseline results use one seed per setting.

Average over the 12 generation benchmarks

Method10% MMD ↓10% flat-KL ↓20% MMD ↓20% flat-KL ↓
MarkovTS (Table 1 reference)0.06050.19050.06050.1905
TimeDP0.13850.29360.14040.2976
ImagenFew0.11510.34330.11990.4389
Diffusion-TS0.11860.36450.14650.3416
TimeVQVAE0.09100.26930.09920.3291
TimeVQVAE-C0.09430.28380.09540.3112

The fixed MarkovTS reference has lower average MMD and flat-KL than the baselines at either support level. Together, these comparisons demonstrate competitive distributional matching with frozen MarkovTS parameters on unseen target domains.

Manuscript revisions: Section 5.1 (Experimental Setup); Appendix D.2 (Generation Evaluation Protocol and Baselines) and F.1–F.2 (Stock/Web and Limited-Support Comparisons, Tables 32–33).

To Q2: Forecasting Stability across Random Seeds and Augmentation Comparisons

Thank you for this suggestion. We added forecasting experiments with seeds 2026 and 2027 to the original seed 42, covering all five forecasting backbones, six datasets, and four prediction horizons. After averaging over horizons and equally aggregating across datasets and backbones, the results are as follows.

Training seedMAE ↓MSE ↓
420.34500.3387
20260.34550.3400
20270.34460.3388

The aggregate forecasting performance is similar across the three seeds. The revised manuscript reports three-seed averages in the main text and dataset- and backbone-wise results in the appendix.

We also compare MarkovTS-based augmentation with real-only training, five alternative generators (ImagenFew, TimeDP, Diffusion-TS, TimeVQVAE, TimeVQVAE-C), and two augmentation schemes (Gaussian 0.2 and TPS [1]). Averaged across six datasets, five forecasting backbones, and four prediction horizons, MarkovTS achieves three-seed mean MAE and MSE of 0.3450 and 0.3392, reducing the corresponding real-only errors by 3.30% and 3.39%. Against the other seven generation and augmentation approaches, the reductions in matched mean MAE range from 2.01% to 4.73%, with mean MSE reductions of 2.46–6.26%. Complete dataset-, backbone-, and horizon-wise results appear in the appendix.

[1] Bakhshaliyev et al. (2026). Temporal Patch Shuffle (TPS): Leveraging Patch-Level Shuffling to Boost Generalization and Robustness in Time Series Forecasting.

Manuscript revisions: Section 5.4 (Downstream Utility, Tables 4–5); Appendix D.4 (Forecasting Protocol) and H.1–H.3 (Horizon-Wise Results, Repeated Runs, and Alternative Augmentation Sources, Tables 34–36).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their careful reading, constructive comments, and suggestions. We have revised the manuscript to clarify the state–transition representation and strengthen the evaluation of reference consistency, structural diversity, cross-domain transfer, and downstream augmentation. The main revisions are summarized below.

  • Clarifying the representation and its Markov-inspired motivation. We expanded the explanation of why locally normalized states alone cannot capture contextual changes between adjacent intervals (Section 4.2). We also strengthened the representation analysis through the paired Patch-Flow comparison (Table 3) and a direct examination of state–transition complementarity (Figure 5), showing how transition codes distinguish changes beyond state-pair morphology.

  • Distinguishing reference consistency from replication. We added reference-distance and within-reference diversity measurements, ACF/PSD structural coverage, and explicit replication and Gaussian-perturbation controls (Sections 5.2 and 5.5; Appendix E.1.2). With 100 candidates per reference, MarkovTS reaches 41.25% coverage, compared with 24.00% for reference replication and 27.39% for Gaussian perturbation. A separate nearest-training-window analysis finds no exact training-window matches under the evaluated five-candidate protocol.

  • Extending structural diversity through state–transition recombination. We introduced State–Transition Recombination (STR) as a conditioning variant using the frozen generator (Section 4.5), and evaluated its effects on pattern placement, shape variation, coverage, and distributional fidelity (Section 5.5, Table 8). STR increases mean within-reference Shape APD from 0.302 to 1.324/1.340 for its two settings. The evaluation also characterizes the resulting trade-offs in structural coverage and distributional matching.

  • Broadening transfer and multivariate evaluations. We added limited-support comparisons on Stock and Web, as well as the 12 generation benchmarks, while keeping MarkovTS parameters frozen (Appendices F.1–F.2). We also evaluated contemporaneous and lagged cross-variable correlations on 11 multivariate datasets (Appendix E.4), providing direct evidence for the dependence preservation discussed in Section 4.4.

  • Strengthening downstream and reproducibility evidence. We extended forecasting evaluation to three training seeds and compared MarkovTS augmentation with five alternative generators and two conventional augmentation strategies across six datasets, five backbones, and four horizons (Section 5.4; Appendices H.1–H.3). MarkovTS attains aggregate MAE/MSE of 0.3450/0.3392, reducing real-only errors by 3.30%/3.39%. Appendix H.4 additionally reports data-preparation costs, including 148.34 seconds to generate the six forecasting datasets on one RTX 4090.

The appendices provide the detailed protocols, dataset-wise results, and additional analyses. Three supplementary visualization files further illustrate fixed-reference generation, extended STR samples, and complete sequences at lengths 512, 1024, and 2048. The individual responses identify the relevant sections, figures, tables, and supplementary files. New and revised passages are highlighted in the manuscript for ease of review.

We thank the reviewers again for their time and feedback.

Area_Chair_8yGdOpenReview ↗

Discussion period ending soon

Dear Reviewers,

Thank you for your reviews of this submission. The authors have posted responses addressing the concerns raised.

With limited time remaining in the discussion period, please take a moment to review the authors' responses and post a brief follow-up indicating whether your concerns have been addressed or what remains outstanding. If your assessment changes as a result, please update your score accordingly.

Active engagement at this stage is essential for a fair and well-substantiated final decision. Thank you for your time and effort in reviewing this submission.

Best regards, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.