Pith. sign in

REVIEW 6 major objections 4 minor 23 references

Hopformer claims that forecasting with high-dimensional covariates is best split into a sparsity-aggregated trend stage plus a LoRA-fine-tuned residual Transformer, with an oracle inequality and dependence-aware generalization bounds, and r

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 05:12 UTC pith:2HRYQJOA

load-bearing objection The headline SOTA claim is contradicted by the paper's own appendix; the two-stage wrapper idea is solid and worth peer review, but this version overclaims. the 6 major comments →

arxiv 2607.22299 v1 pith:2HRYQJOA submitted 2026-07-24 stat.ML cs.LG

Hopformer: Homogeneity-Pursuit Transformer for Time Series Forecasting

Wan Zhang , Qinjie Lin , Chan Lee , Weijian Li , Han Liu , Kai Zhang This is my paper
classification stat.ML cs.LG MSC 62M1062J0768T07
keywords time series forecastinghigh-dimensional covariatessparsity pattern aggregationoracle inequalitylow-rank adaptationgeneralization boundβ-mixingfoundation models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Hopformer sets out to solve a specific failure of modern forecasting Transformers: they ignore or mishandle high-dimensional external covariates. The paper proposes to split forecasting into two stages—first, a sparsity-aware aggregation of regression experts extracts a common low-variance trend driven by covariates; second, a pretrained Transformer fine-tuned with low-rank adaptation models the leftover residual. The authors prove an oracle inequality showing the aggregation stage is near-optimal relative to the best sparse combination of experts, and a generalization bound for the fine-tuning stage under stationary, β-mixing residuals. If the claims hold, Hopformer is a modular, theory-backed way to bolt covariate modeling onto any foundation model, with an average 6.56% MASE improvement reported across six benchmarks.

Core claim

On its own terms, the paper claims that the right decomposition for covariate-driven forecasting is a homogeneity-pursuit trend plus a residual: a pool of cross-sectional regressors (linear, tree-based, neural) is combined through sparsity pattern aggregation, which weights models by empirical performance with a sparsity prior, producing a shared low-variance trend that absorbs covariate effects. The residual is then handed to a pretrained Transformer fine-tuned with low-rank adaptation. The theoretical core is two theorems: SPA satisfies an oracle inequality of the form E||ĝ_SPA − η||² ≤ min_p (E||ĝ_ω̂p − η||² + (4σ²/n) log π_p^{-1}), and the LoRA stage has a generalization bound scaling as

What carries the argument

The load-bearing object is Sparsity Pattern Aggregation (SPA), a convex exponential-weighting scheme over binary patterns of expert subsets. SPA defines weights ω^SPA proportional to the sum over patterns p of least-squares estimates ω̂_p times prior π_p times exp(−empirical risk/4σ² − |p|/2), which simultaneously selects and averages experts. Its role is to turn a heterogeneous collection of series into a common residual space by extracting a low-variance covariate-driven trend, with a proved oracle inequality. The second pillar is low-rank adaptation (LoRA), which updates attention weights through low-rank matrices B A; the paper uses the fact that LoRA shrinks mutual information I(R^T;W)

Load-bearing premise

The generalization guarantee for the residual Transformer rests on the residual series being stationary and β-mixing with fast enough dependence decay that blocks of length a can be treated as nearly independent; if the residuals have long memory or slowly decaying mixing, the bound's constant m can absorb so much bias that the stated inequality says nothing.

What would settle it

Take a real or synthetic series with long-range dependence (e.g., fractional noise with Hurst exponent above 0.8), run Hopformer's Stage 1, estimate the β-mixing coefficients of the residuals, and check whether β_a decays fast enough that the block length a and block count m satisfy 2am ≤ T with the bound's right-hand side smaller than the observed generalization gap. If the gap exceeds the bound for all admissible a and m, the theorem's premise is violated. Simpler: on a covariate-free dataset, SPA should confer no benefit; if it still improves forecasts, the gain is not coming from covariate

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A covariate-driven trend stage can be added to any pretrained time-series Transformer without retraining the backbone; the paper reports consistent zero-shot MASE reductions across four different foundation models.
  • Because the SPA oracle inequality penalizes model complexity by log prior mass, larger and more diverse expert pools should be usable without overfitting—the paper's ablation shows SPA's advantage widening from 4 to 20 experts.
  • LoRA fine-tuning of the residual stage recovers essentially all of the accuracy of full fine-tuning (within 1% MAPE) while updating only a handful of low-rank parameters.
  • The generalization bound implies any parameter-efficient fine-tuning method that reduces mutual information between data and weights should inherit a similar guarantee, not just LoRA.
  • Experiments varying context length and forecast horizon indicate the gains concentrate where data is scarce: the largest reported improvement over the backbone occurs at the shortest context length on the electricity-price dataset.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension of the framework is to use SPA not only as a front-end for forecasting but as a preprocessing diagnostic: if the first stage is doing its job, residual series should show measurably lower autocorrelation and volatility than the original series, and that drop should predict downstream gains.
  • The paper's theory ties generalization to rank and quantization; an untested corollary is that aggressively quantizing or reducing the rank of LoRA updates should improve out-of-sample performance up to a point, a prediction that could be checked by sweeping these hyperparameters.
  • Because SPA is a convex aggregation of arbitrary experts, the same homogeneity-pursuit stage could be applied to classification or survival tasks with shared covariates, where a 'common signal plus residual' decomposition is equally natural.
  • If the stationarity/mixing assumption fails in practice, the second-stage bound degrades gracefully only if the local-stationarity extension holds; measuring empirical mixing coefficients of residuals on real benchmarks would tell whether the theoretical regime is actually inhabited.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

6 major / 4 minor

Summary. The paper proposes Hopformer, a two-stage framework for forecasting multiple time series with high-dimensional covariates. Stage 1 uses Sparsity Pattern Aggregation (SPA) over a pool of cross-sectional regression experts to extract a common low-variance trend; Stage 2 fits a pre-trained Transformer (Chronos-bolt-small) to the residuals using LoRA. The authors claim two theoretical contributions—an oracle inequality for SPA and information-theoretic generalization bounds for LoRA under dependent data—and an empirical headline that Hopformer improves MASE by an average of 6.56% and sets a new state of the art. The experiments cover three synthetic datasets and three real-world datasets, with ablations comparing aggregation strategies and LoRA versus full fine-tuning.

Significance. The two-stage decomposition is conceptually attractive: it targets a real gap in the literature, namely the use of high-dimensional future covariates with pretrained foundation models, and the modular design plus released code are useful. If the theoretical guarantees were rigorous and the empirical comparisons complete, the paper could make a worthwhile contribution. However, the central empirical claim is contradicted by the paper's own Appendix F, where TimeMixer, ChronosX, and TimeXer achieve substantially lower MASE than Hopformer on several datasets, meaning the '6.56% improvement' and 'new state of the art' statements are artifacts of baseline selection. The theoretical results also have load-bearing gaps: Theorem 1 is proved under an i.i.d. multi-sample model that does not match the single-trajectory rolling-window protocol used in the experiments, and Theorem 2 hides the dependence on mixing coefficients behind an existential constant. Given these issues, the paper's main claims are not currently supported.

major comments (6)
  1. [Abstract and Section 5, Table 2 vs Appendix F, Table 10] The headline claim that Hopformer 'sets a new state of the art, improving MASE by an average of 6.56%' is computed from Table 2, whose baseline set excludes TimeMixer, TimeXer, and ChronosX. Appendix F (Table 10) includes these models and reports clearly lower MASE for them on multiple datasets. For example, on Sale1, Hopformer LoRA is 0.819, while TimeMixer full LoRA is 0.301 and ChronosX full LoRA is 0.496; on Electricity, Hopformer LoRA is 0.730, while TimeMixer LoRA is 0.152. Unless the appendix uses a materially different protocol (which is not stated), the paper's own results falsify the 'new state of the art' claim. The 6.56% figure is at best an improvement over the baselines selected for the main table, not over the state of the art.
  2. [Section 4.1, Eq. (1) and Eq. (3)] Theorem 1 is stated for data generated as X_t^{(i)} = η(Z_t^{(i)}) + ξ_i with ξ_i i.i.d. Gaussian and, in the sample version of Eq. (1), an i.i.d. sample of size n for each time point (X_{t,k}^{(i)}, Z_{t,k}^{(i)}). This does not describe the experimental protocol, which uses a single observed trajectory per series with 20 rolling windows. The oracle inequality therefore does not apply to the estimator actually evaluated. The proof in Appendix A.1 also imports Lemma 4 without showing how the unknown regression function η and the prior normalization are handled, and the statement that the SPA estimator satisfies the bound in expectation is asserted rather than derived from the displayed lemma.
  3. [Section 4.2, Lemma 8 and Theorem 2, Eq. (4)] The mixing argument is not quantitative. Lemma 8 says 'we choose a large enough so that the odd blocks are almost independent', but no relation between the block length a, the mixing coefficient β_a, and the final constant is given. The theorem merely asserts 'there exists a constant a > 0 and an integer m' with 2am ≤ T. If the mixing decay is slow, a must be chosen very large, making m small and potentially rendering the bound vacuous. A generalization bound that depends on the mixing rate (e.g., through β_a or a matching condition on a versus T) is needed for the statement to have content. As written, Eq. (4) hides the principal difficulty of dependent data.
  4. [Corollary 3 and Appendix A.2] Corollary 3 introduces a quantization assumption—that LoRA parameters are quantized to q bits and uniformly distributed—that is not present in Theorem 2 or in the experiments, where LoRA weights are continuous. The entropy bound H(ΔW) ≤ q r Σ(d_in + d_out) is valid only under such a quantization model, which is not shown to hold for the actual training procedure. The mutual-information manipulations, including the notation I(W; R^T | P_{W|R^T}), are also nonstandard and insufficiently justified. This weakens the claimed theoretical support for the LoRA stage.
  5. [Section B, synthetic data, and Appendix C.1] The synthetic datasets are generated from additive covariate models (e.g., Sale1 sales equals baseline plus sinusoid plus promotion/temperature/price effects; Electricity load equals base plus seasonal plus temperature/calendar effects) that are precisely the functional forms the expert regressors are designed to learn. The large gains on these datasets therefore partly verify the data-generation setup rather than establish a general forecasting advantage. In addition, the M5 results in Table 2 are based on an unspecified subset; Appendix C.1 states the full 30K-series M5 dataset exceeded the computational budget. The paper should clearly label the M5 evaluation as a subset and should test on naturally occurring covariate data at scale.
  6. [Section 5, Tables 2, 3, and 7-8] The empirical section lacks consistency and statistical support. Table 2 reports context length 512, while Table 3 uses context length 256; the zero-shot Chronos MASE on EPF differs between these tables (0.662 vs 0.785), so the tables are not directly comparable. No error bars, standard deviations, or significance tests are provided for the 20 rolling windows. Given the small number of datasets (three of which are synthetic) and the strong claims made, this is insufficient support for the headline average improvement.
minor comments (4)
  1. [Table 2] Several cells are malformed, e.g., 'MAPE 0.486 0.485 635 0.912' in the Sale2 row, and the table caption misspells 'Chronos'. These should be corrected.
  2. [Section 5, Ablation Study 2] The sentence 'This suggests that Hopformer offers a more effective and interpretable way to incorporate exogenous information than direct multivariate modeling' is duplicated verbatim.
  3. [Appendix A.2, Lemma 4] The proof of Lemma 4 contains notation that is not fully defined (e.g., the relationship between the empirical risk estimates and the loss ℓ), and the step 'log w_m' has a sign that is not derived. The appendix would benefit from a complete proof rather than a sketch.
  4. [Section 3.1, prior definition] The prior π_p is defined with a normalization constant H and a separate 1/2 term for |p|=M; it is not clear that these pieces are consistent or that π_p sums to one. Please clarify the normalization.

Circularity Check

0 steps flagged

No circularity in the derivation chain; the theoretical results are imported from external published theorems, and the empirical SOTA claim is contradicted by the paper's own Appendix F but that is a baseline-selection/correctness issue, not a circular reduction.

full rationale

Walking the paper's derivation chain, I find no step where a claimed prediction or first-principles result reduces to its inputs by construction. Stage 1's SPA estimator is the Rigollet-Tsybakov exponential-screening procedure, and Theorem 1 is the standard oracle inequality for that estimator, cited to Rigollet and Tsybakov [2011] and proved via the Leung-Barron Lemma 4. This is external, published support whose assumptions do not include Hopformer's target results; per the rules, an external theorem is independent evidence and does not create circularity. Stage 2's Theorem 2 is an information-theoretic generalization bound adapted from Xu and Raginsky [2017] via a mixing-block argument; the proof is incomplete (the dependence on mixing coefficients is hand-waved as 'we choose a large enough so that the odd blocks are almost independent'), but an incomplete or vacuous proof is a correctness risk, not a circular reduction, because the bound is not obtained by substituting the conclusion into the assumptions. Corollary 3's entropy bound I <= q r sum(d_in+d_out) is a valid upper bound, not a definitional identity. The empirical headline ('improving MASE by an average of 6.56%') is a summary of Table 2, which omits TimeMixer, TimeXer, and ChronosX; Appendix F Table 10 shows those models achieve lower MASE on Sale1, Electricity, and EPF. This undermines the 'new state of the art' claim as a matter of baseline selection and internal consistency, but the claimed improvement is an empirical comparison, not a quantity defined as the input of the calculation. No fitted parameter is renamed as a prediction, no load-bearing self-citation chain exists, and no ansatz is smuggled in via citation. Accordingly, the circularity score is 0.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The theoretical claims import the SPA oracle inequality as a black box and assume a data model (i.i.d. replicates at fixed time) that the experiments do not satisfy. The generalization bound's constants depend on existential choices and a quantization assumption absent from the implementation. The empirical headline depends on a hand-picked expert pool, an unestimated σ², and a nonstandard M5 subset.

free parameters (6)
  • SPA noise variance σ² = not reported
    Equation (1) defines SPA weights via exp(−1/(4σ²)·Σ residuals); Theorem 1 assumes σ² known. The experiments never state how σ² is estimated.
  • LoRA hyperparameters (rank, alpha, dropout, steps) = r=8, α=16, dropout 5%, 100 gradient steps
    Section 5: chosen by hand; no sensitivity analysis. Rank directly controls the entropy bound in Corollary 3.
  • Expert pool size M = 8
    Section 5: authors state additional regressors (up to 16) showed weak performance under default settings and 8 was selected as a balanced pool.
  • Block length a and number of blocks m in Theorem 2 = unspecified; 'there exists a constant a>0 and integer m'
    Theorem 2 and Lemma 8: the bound's constant depends on an existential choice; no guidance for actual mixing rates.
  • Quantization level q (Corollary 3) = not in implementation
    Corollary 3 bounds mutual information via q-bit quantization; experiments use standard float LoRA with no quantization.
  • M5 evaluation subset = 730 x 200 (timesteps x series)
    Table 1 lists M5 as 730×200, while Section C.1 admits the full M5 has ~30K series and several million points and was not run.
axioms (6)
  • standard math Oracle inequality of Rigollet-Tsybakov (2011)
    Theorem 1 is presented as a consequence of the SPA prior; the proof refers to Lemma 4 (Leung-Barron) and one line 'The proof follows from letting...'. The near-optimal bound is assumed from the SPA theory.
  • domain assumption Each time point (i,t) has n i.i.d. samples from a conditional distribution with Gaussian noise of known variance σ²
    Section 3.1 introduces {(X_t,k, Z_t,k)}^n_{k=1} as i.i.d. realizations; this does not match the actual single-trajectory time-series protocol used in experiments.
  • domain assumption Residuals are strictly stationary and β-mixing with mixing coefficients that decay fast enough for a block argument with no penalty term
    Theorem 2 / Lemma 8: proof says 'we choose a large enough so that odd blocks are almost independent' and concludes σ/√m-subgaussianity without explicit mixing terms.
  • domain assumption Loss ℓ(w, R) is σ-subgaussian
    Theorem 2 setup; no verification for MASE or MAPE losses used in experiments.
  • ad hoc to paper LoRA parameters are quantized to q bits and uniformly distributed
    Corollary 3 and remark: entropy bound H(W) ≤ q r Σ(d_in+d_out) assumes discrete uniform support; implementation uses continuous float LoRA (rank 8).
  • standard math Xu-Raginsky inequality for subgaussian f
    Lemma 5, used to convert mutual information into a generalization gap.

pith-pipeline@v1.3.0-alltime-deepseek · 26287 in / 22141 out tokens · 172954 ms · 2026-08-01T05:12:16.637433+00:00 · methodology

0 comments
read the original abstract

Forecasting multiple time-series with high-dimensional covariates presents a core challenge: unifying common temporal patterns while retaining meaningful series-specific information. We introduce Hopformer (Homogeneity-Pursuit Transformer), a two-stage framework that addresses this challenge. In the first stage, we perform a Sparsity Pattern Aggregation (SPA) scheme extracting a common low-variance trend that incorporates the covariates. This acts as a homogenization layer. In the second stage, a LoRA-fine-tuned Transformer models the remaining complex dependencies in the residual. Our method is theoretically grounded. We prove that SPA achieves a near-optimal bias-variance trade-off via an oracle inequality. We also provide generalization bounds for the second stage under dependent time series data. Hopformer sets a new state of the art, improving MASE by an average of 6.56% across synthetic and real-world forecasting benchmarks.

Figures

Figures reproduced from arXiv: 2607.22299 by Chan Lee, Han Liu, Kai Zhang, Qinjie Lin, Wan Zhang, Weijian Li.

Figure 1
Figure 1. Figure 1: Overview of the Hopformer workflow. In Stage I, SPA extracts a low-variance common trend from covariates. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Zero-shot forecast comparison between Hopformer and Chronos on four representative stores from the [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (Top) Effect of context length on model performance across datasets (prediction length = 24). (Bottom) Effect [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 7
Figure 7. Figure 7: This suggests that the covariate-driven expert pool provides valuable signal when historical information is scarce. (ii) Fine-tuned performance: After full-parameter or LoRA fine-tuning, Hopformer still yields lower error than Chronos. Removing covariate effects in the first stage appears to simplify the residual dynamics, making the subsequent 8 [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of synthetic dataset (SALE1) generation. Left: parameter distribution across stores. Right: sample time series showing temporal dynamics. effect Ecov i,t incorporates multiple nonlinear and interactive effects: E cov i,t = βpromo · promoi,t + βholiday · holidayt + βweekend · weekendt + βcompetitor · comp_promoi,t + βinventory · inventoryi,t + mi,t + βtemp · (1 − |Ti,t − T ∗ i |) + βprice · pri… view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of synthetic dataset (SALE2) generation. Left: parameter distribution across stores. Right: sample time series showing temporal dynamics. Here, Br is the base load, αr is the regional trend (load growth), and Dr(t), Wr(t), Yr(t) are the daily, weekly, and yearly seasonal patterns. The covariate component Ecov r,t accounts for complex external influences: E cov r,t = βtemp · fT (Tr,t) + βhumid … view at source ↗
Figure 6
Figure 6. Figure 6: Load decomposition over one representative week for two regions across summer and winter. Each panel [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: This suggests that the covariate-driven expert pool provides valuable signal when historical information is scarce. (ii) Fine-tuned performance: After full-parameter or LoRA fine-tuning, Hopformer still yields lower error than Chronos. Removing covariate effects in the first stage appears to simplify the residual dynamics, making the subsequent transformer easier to adapt. (iii) Cross-sectional aggregation… view at source ↗
Figure 7
Figure 7. Figure 7: Model robustness across varying context lengths and forecast horizons. Only the informative portion of each [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Decomposition of Hopformer’s prediction pipeline on the SALE1 dataset. Each column corresponds to a store. Row 1: ground truth and final forecast; Row 2: cross-sectional stage (covariate aggregation); Row 3: residual transformer output; Row 4: covariate trajectories. Removing covariate effects (Row 1 → 3) yields a smoother, quasi-periodic residual series, illustrating the division of labour between the two… view at source ↗
Figure 9
Figure 9. Figure 9: Simulated grocery-sales example (grey curves are individual regressors encoding trend, seasonality, and promotion effects). Left: Linear mixture of 3 regressors over 100 time steps with added noise—SPA matches ordinary least-squares (LR) aggregation. Middle: Non-linear combination of 20 regressors over 50 time steps—SPA outperforms LR by capturing interaction effects. Right: Same non-linear setting over 20… view at source ↗
Figure 10
Figure 10. Figure 10: Forecasting results on the Sales1 dataset. top: Moirai-MoE; middle: Moirai; bottom: Lag-Llama. Prediction Phase: For forecasting, the algorithm applies the learned decomposition in reverse: it first extracts residuals from the context data, predicts future residuals using the foundation model, then reconstructs the final forecasts by adding back the predicted covariate effects. This design enables the fra… view at source ↗
Figure 11
Figure 11. Figure 11: Forecasting results on the EPF dataset. top: Moirai-MoE; middle: Moirai; bottom: Lag-Llama. 1 from autogluon.timeseries import TimeSeriesDataFrame 2 from hopformer import Predictor # Hopformer wrapper 3 4 # ---------- load & split data ---------- 5 df = TimeSeriesDataFrame.from_path("store_sales_data.csv") 6 train, test = df.train_test_split(prediction_length=24) 7 8 # ---------- fit Hopformer ---------- … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 3 canonical work pages

  1. [1]

    T. Aksu, G. Woo, J. Liu, X. Liu, C. Liu, S. Savarese, C. Xiong, and D. Sahoo. Gift-eval: A benchmark for general time series forecasting model evaluation.arXiv preprint arXiv:2410.10393,

    Pith/arXiv arXiv

  2. [3]

    URLhttps://arxiv.org/abs/2310.04948. T. Chen and C. Guestrin. Xgboost: A scalable tree boosting system. InProceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794,

    Pith/arXiv arXiv

  3. [7]

    URLhttps://arxiv.org/abs/2401.03955. S. Gao, T. Koker, O. Queen, T. Hartvigsen, T. Tsiligkaridis, and M. Zitnik. Units: A unified multi-task time series model,

    Pith/arXiv arXiv

  4. [8]

    URLhttps://arxiv.org/abs/2403.00131. R. Godahewa, C. Bergmeir, G. I. Webb, R. J. Hyndman, and P. Montero-Manso. Monash time series forecasting archive. arXiv preprint arXiv:2105.06643,

    Pith/arXiv arXiv

  5. [11]

    URL https://arxiv.org/abs/2310.01728. G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y . Liu. Lightgbm: A highly efficient gradient boosting decision tree.Advances in neural information processing systems, 30,

    Pith/arXiv arXiv

  6. [16]

    URLhttps://arxiv.org/abs/2508.07490. Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers.arXiv preprint arXiv:2211.14730,

    Pith/arXiv arXiv

  7. [17]

    URL https: //arxiv.org/abs/2310.08278. P. Rigollet and A. Tsybakov. Exponential screening and optimal rates of sparse estimation.The Annals of Statistics, 39 (2):731–771,

    Pith/arXiv arXiv

  8. [18]

    Y . Wang, H. Wu, J. Dong, G. Qin, H. Zhang, Y . Liu, Y . Qiu, J. Wang, and M. Long. Timexer: Empowering transformers for time series forecasting with exogenous variables.arXiv preprint arXiv:2402.19072,

    Pith/arXiv arXiv

  9. [19]

    URLhttps://arxiv.org/abs/2210.02186. J. Xiao, Y . Chen, G. Cong, W. Nejdl, and S. Gottschalk. Flextsf: A universal forecasting model for time series with variable regularities,

    Pith/arXiv arXiv

  10. [20]

    URLhttps://arxiv.org/abs/2410.23160. A. Xu and M. Raginsky. Information-theoretic analysis of generalization capability of learning algorithms. In I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume

    Pith/arXiv arXiv

  11. [21]

    24 Running Title for Header Y

    URL https://arxiv.org/abs/2210.08964. 24 Running Title for Header Y . Yao, D. Li, H. Jie, H. Jie, T. Li, J. Chen, J. Wang, F. Li, and Y . Gao. Simplets: An efficient and universal model selection framework for time series forecasting.Proc. VLDB Endow., 16(12):3741–3753, Aug

    Pith/arXiv arXiv

  12. [22]

    doi: 10.14778/3611540.3611561

    ISSN 2150-8097. doi: 10.14778/3611540.3611561. URLhttps://doi.org/10.14778/3611540.3611561. B. Yu. Rates of Convergence for Empirical Processes of Stationary Mixing Sequences.The Annals of Probability, 22 (1):94 – 116,

    arXiv

  13. [1991]

    doi: 10.1162/neco.1991.3.1.79

    ISSN 0899-7667. doi: 10.1162/neco.1991.3.1.79. URL https://doi.org/10.1162/neco. 1991.3.1.79. M. Jin, S. Wang, L. Ma, Z. Chu, J. Y . Zhang, X. Shi, P.-Y . Chen, Y . Liang, Y .-F. Li, S. Pan, and Q. Wen. Time-llm: Time series forecasting by reprogramming large language models,

    doi:10.1162/neco.1991.3.1.79 1991

  14. [1994]

    doi: 10.1214/aop/1176988849. Y . Zhang and J. Yan. Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting. InThe eleventh international conference on learning representations,

    arXiv

  15. [2006]

    doi: 10.1109/TIT.2006.878172. S. Li, X. Jin, Y . Xuan, X. Zhou, W. Chen, Y .-X. Wang, and X. Yan. Enhancing the locality and breaking the memory bottleneck of transformer on time series forecasting.Advances in neural information processing systems, 32,

    arXiv 2006

  16. [2008]

    doi: 10.1007/s10994-008-5051-0

    ISSN 1573-0565. doi: 10.1007/s10994-008-5051-0. URL http: //dx.doi.org/10.1007/s10994-008-5051-0. A. Das, W. Kong, R. Sen, and Y . Zhou. A decoder-only foundation model for time-series forecasting. InForty-first International Conference on Machine Learning,

    work page doi:10.1007/s10994-008-5051-0

  17. [2014]

    URLhttps://doi.org/10.1214/14-EJS886

    doi: 10.1214/14-EJS886. URLhttps://doi.org/10.1214/14-EJS886. A. Dalalyan and A. B. Tsybakov. Aggregation by exponential weighting, sharp pac-bayesian bounds and sparsity. Machine Learning, 72(1–2):39–61, Apr

    work page doi:10.1214/14-ejs886

  18. [2019]

    X. Liu, J. Liu, G. Woo, T. Aksu, Y . Liang, R. Zimmermann, C. Liu, S. Savarese, C. Xiong, and D. Sahoo. Moirai-moe: Empowering time series foundation models with sparse mixture of experts.arXiv preprint arXiv:2410.10469, 2024a. Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long. itransformer: Inverted transformers are effective for time series f...

    Pith/arXiv arXiv

  19. [2021]

    URL https://arxiv.org/abs/2104.08691. G. Leung and A. Barron. Information theory and mixing least-squares regressions.IEEE Transactions on Information Theory, 52(8):3396–3410,

    Pith/arXiv arXiv

  20. [2022]

    URL https://arxiv.org/abs/ 2204.10398. V . Ekambaram, A. Jati, P. Dayama, S. Mukherjee, N. H. Nguyen, W. M. Gifford, C. Reddy, and J. Kalagnanam. Tiny time mixers (ttms): Fast pre-trained models for enhanced zero/few-shot forecasting of multivariate time series,

    Pith/arXiv arXiv

  21. [2023]

    doi: 10.1016/j.ijforecast.2022.01.008

    ISSN 0169-2070. doi: 10.1016/j.ijforecast.2022.01.008. URLhttp://dx.doi.org/10.1016/j.ijforecast.2022.01.008. D. Gupta, A. Bhatti, and S. Parmar. Beyond loRA: Exploring efficient fine-tuning techniques for time series foundational models. InNeurIPS Workshop on Time Series in the Age of Large Models, 2024a. URL https://openreview. net/forum?id=YZJ8Re0gQv. ...

    Pith/arXiv arXiv 2070

  22. [2024]

    URL https://arxiv.org/abs/2403.07815. D. Cao, F. Jia, S. O. Arik, T. Pfister, Y . Zheng, W. Ye, and Y . Liu. Tempo: Prompt-based generative pre-trained transformer for time series forecasting,

    Pith/arXiv arXiv

  23. [2025]

    23 Running Title for Header B

    doi: 10.1007/s10618-025-01120-8. 23 Running Title for Header B. Lester, R. Al-Rfou, and N. Constant. The power of scale for parameter-efficient prompt tuning,

    work page doi:10.1007/s10618-025-01120-8