Fusing Large Language Models with Temporal Transformers for Time Series Forecasting

Chen Su; Jun Zhang; Qinyu Liu; Yan Song; Yuanhe Tian

REVIEW 3 major objections 5 minor 3 cited by

Fusing Large Language Models with Temporal Transformers for Time Series Forecasting

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A single learned gate that mixes an LLM's semantic representation into a temporal Transformer's hidden state is claimed to yield more accurate forecasts than either model alone.

desk verdict The architecture is a plausible incremental idea and the internal ablations support it, but the paper's headline SOTA claim is contradicted by its own reported numbers: over the six datasets actually tabulated, Time-LLM has lower average MSE and MAE than the proposed model. read the letter →

arxiv 2507.10098 v1 pith:CKVH3LDO submitted 2025-07-14 cs.CL

Chen Su , Yuanhe Tian , Qinyu Liu , Jun Zhang , Yan Song This is my paper

classification cs.CL

keywords timeseriesforecastinglargelanguagemodelsTransformergatedfeaturefusionsemanticrepresentationtemporaldynamicsprompt-basedencoding

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that the right way to involve a large language model in time-series forecasting is not to let the LLM make predictions, but to let it contribute a complementary semantic reading of the series while a conventional Transformer keeps track of temporal dynamics. Its architecture feeds both the original patches and the Transformer's own hidden representation into the LLM, prompts the LLM to reason about the series, and then merges the LLM's output into an intermediate Transformer layer with a learned scalar gate. The claim is that this fused representation outperforms either branch alone, outperforms simple addition of the two representations, and improves on existing Transformer-, MLP-, and LLM-based forecasting baselines across the datasets tested. If true, it gives a practical recipe for transplanting LLM pattern knowledge into numerical forecasting without forcing the LLM to decode continuous values.

What carries the argument

The carrying mechanism is the gated fusion layer defined by Eq. 5: a single scalar $g\in(0,1)$ is computed from the concatenated LLM and Transformer representations through a linear layer and sigmoid, and the fused vector is the convex combination $g Z_{\mathrm{LLM}} + (1-g)Z_{l-1}$. The surrounding machinery that makes the LLM representation worth fusing is the prompt-structured LLM encoding: the LLM receives a task prompt, the Transformer's temporal features, the original patch data, and a data prompt, and only the outputs at the patch positions are kept. The argument assigns to this fusion layer both the preservation of temporal dynamics and the injection of semantic understanding, and the paper treats the fusion layer, rather than the LLM or Transformer alone, as the source of the forecasting gain.

What would settle it

Retrain the model with the gate fixed to 0.5 (a plain average of the LLM and Transformer representations) on the same datasets, keeping everything else identical. If the fixed 50-50 mixture matches or beats the learned-gate version, then the adaptive gate in Eq. 5 is not the source of the reported accuracy, and the paper's central claim about the fusion mechanism would be refuted.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the strengths of the two model families are complementary rather than competing. A vanilla Transformer trained directly on time series encodes continuous temporal dynamics, while a lightly adapted LLM, prompted with task, feature, and data descriptions, extracts sparse high-level semantic patterns from the same patches. The paper's Eq. 5 defines the fused token representation as $Z_E^{l-1}=g\cdot Z_{\mathrm{LLM}}+(1-g)\cdot Z_{l-1}$, with $g$ produced by a linear-sigmoid gate over the concatenation of the two representations. The reported averaged results across the tested datasets show this adaptive fusion achieving the lowest average MSE and MAE among all compared models, with notable margins on the small ILI dataset and on ETTh2.

Load-bearing premise

The whole method rests on the idea that a single learned number can safely blend the LLM's way of seeing the series with the Transformer's way of seeing the series; if those two ways are too different, the blend can throw away the information the method is supposed to combine.

Editorial extensions

If this is right

If the gated fusion is responsible for the gains, then LLM-based forecasters should be built with the LLM as a side feature extractor, not as the final predictor, at least for numerical series.
The reported result that larger, more linguistically capable LLMs hurt performance implies the semantic side branch does not need to scale; a small model with modest language priors suffices.
Because the gate is a learned interpolation, the same architecture can be dropped into an existing patched-Transformer backbone by choosing which layer receives the LLM representation, without changing the prediction head.
The comparison against direct addition indicates that how the fusion happens matters: adaptive gating is claimed to avoid the confusion that comes from simply adding heterogeneous features.
The method points toward a general fusion strategy for multimodal or multi-encoder forecasting: keep each encoder in its native role and learn a single mixing weight rather than expecting one model to do everything.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper does not test whether the gate itself is doing the work or whether any interpolation would help. A natural next experiment is to fix $g$ at 0.5; if the fixed mixture matches the learned gate, the 'adaptive' part of the mechanism is not essential.
The t-SNE evidence that the two encoders form separate clusters suggests this is a late-space mixing problem. If that separation is real, a per-dimension gate or a low-rank alignment projector before the convex combination might preserve more information than one scalar gate.
The finding that stronger LLMs hurt implies the useful signal may not be world knowledge but the attention sparsity the LLM imposes on patches. Testing an LLM with randomized weights, or a non-LLM transformer with the same sparse attention pattern, would separate those explanations.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a reasonable incremental architecture—use a vanilla Transformer as the time-series backbone, a frozen-ish LLM as a semantic feature extractor, and a gated sum to fuse the two. The internal comparison in Table 2 is the paper's real strength: gated fusion (Ours) beats direct addition (Trans-LLM), which beats each component alone, and this ordering is consistent across all six datasets. That is a clean, self-consistent ablation, and the negative result in Figure 4 (bigger LLMs hurt) is a useful data point. The qualitative attention and t-SNE analysis is honest, even if not conclusive.

The problem is the external claim. Section 4.1 says the approach achieves the lowest average MSE and MAE, and Table 3's caption says 'Avg.' is across eight datasets. Only six datasets appear; Electricity is mentioned in Section 3.1 but never shows up; there is no Avg row. If you average the numbers that are actually printed, Time-LLM wins on both metrics: MSE 0.497 vs. 0.502, MAE 0.425 vs. 0.433. So the central SOTA assertion is not merely weakly supported; it is contradicted by direct arithmetic on the paper's own table. That is a load-bearing flaw that must be fixed before the claim can be evaluated. The small differences against PatchTST on several datasets also cry out for confidence intervals or significance tests, which are absent.

The fusion mechanism itself has a softer weakness: Eq. (5) uses a single scalar gate computed from concatenated features, with no analysis of how well the LLM and Transformer representation spaces are aligned, and no ablation comparing scalar gating to per-dimension gating. In principle a scalar convex combination can corrupt information if the spaces are poorly aligned. This is a minor-to-moderate concern, not a fatal one, because the ablations show the gated version does help in practice.

The citation pattern is fine; self-citations are background and no circularity. The paper ships a code link, which is good practice, though I have not run the code.

Who is this for? Researchers working on LLM-plus-Transformer fusion for forecasting who want a simple, testable way to inject LLM representations into a temporal encoder. The paper deserves a serious referee, but only after the authors resolve the dataset discrepancy and either provide the missing Avg row or retract the SOTA claim. This is a conditional accept with major revisions, not a desk reject.

Recommendation: send it to peer review, but the reviewers should be instructed to check the tables against the prose. The arithmetic contradiction alone is enough to block acceptance until fixed.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes a hybrid time series forecasting architecture that fuses semantic representations extracted by a pre-trained LLM (GPT-2, tuned with LoRA) into a standard patch-based Transformer encoder through a scalar-gated convex combination at an intermediate layer. The method is evaluated on six tabulated datasets (Weather, four ETT variants, ILI) across four prediction horizons, with comparisons against Transformer-, MLP-, and LLM-based baselines. The paper's central claims are (i) that the gated fusion in Eq. (5) yields a representation that inherently contains both temporal dynamics and semantic understanding, and (ii) that the approach achieves the lowest average MSE and MAE among all compared methods, i.e., state-of-the-art forecasting accuracy.

Significance. If validated, the proposed architecture would be a simple and parameter-efficient recipe for injecting LLM-derived semantic priors into conventional time series Transformers, and the internal ablation (LLM-only, Trans-only, Trans-LLM, Ours) gives some initial support for the value of adaptive fusion. However, the headline empirical claim as stated in Section 4.1 is contradicted by the authors' own table: a direct arithmetic check of Table 3 over the six reported datasets shows that Time-LLM has a lower average MSE (~0.497 vs ~0.502) and lower average MAE (~0.425 vs ~0.433) than the proposed model. The table captions promise an aggregate over eight datasets, but only six appear, and the Electricity dataset mentioned in Section 3.1 is absent from the results. The significance of the contribution is therefore currently unestablished; the internal comparison is suggestive, but the external state-of-the-art claim needs corrected evidence before the contribution can be assessed.

major comments (3)

[Section 4.1, Table 3] The claim that 'the proposed approach achieves the lowest average MSE and average MAE error values' is contradicted by the numbers printed in Table 3. Across the six datasets actually reported (Weather, ETTh1, ETTh2, ETTm1, ETTm2, ILI), Time-LLM has a lower average MSE (approximately 0.497 versus 0.502 for Ours) and a lower average MAE (approximately 0.425 versus 0.433 for Ours). The table caption and Table 2's caption say 'Avg.' aggregates over eight datasets, but no Avg row is printed and only six datasets appear; Electricity is described in Section 3.1 but has no results in either table. The aggregate claim therefore needs either the missing Electricity results or a corrected statement, and the conclusion of superior accuracy is not supported by the data as reported.
[Eq. (5), Sections 2.3 and 4.3] The core fusion mechanism, Z_E = g * Z_LLM + (1-g) * Z_{l-1} with a single scalar gate, assumes that the LLM and Transformer representation spaces are commensurate enough for a convex combination to be meaningful. The t-SNE visualization in Section 4.3 shows that the two token types form distinct clusters, which is consistent with representational misalignment rather than complementarity, and the paper offers no analysis of the geometry between the two spaces. The manuscript also does not ablate the choice of fusion layer, does not compare scalar gating with per-dimension or gated concatenation, and does not report how sensitive the results are to the gate's behavior. Given that the internal Trans-LLM baseline (simple addition) is within 0.005-0.02 MSE of the gated model on several datasets (Table 2), the superiority of the scalar-gated fusion as the proposed mechanism is not convincingly established.
[Section 3.3 and Table 3] The paper states that all reported results are averaged over three independent runs, but it does not report standard deviations, confidence intervals, or significance tests. Several comparisons against the strongest baselines are very small (e.g., Weather MSE 0.237 for Ours versus 0.225 for both PatchTST and Time-LLM), so without variance information the per-dataset and aggregate ordering cannot be distinguished from noise. This issue compounds the contradiction noted above and should be addressed if any comparative claim is to be made.

minor comments (5)

[Section 3.2] The subsection heading 'For the MLP-based models' appears twice, and the second instance incorrectly lists GPT4TS and Time-LLM, which are LLM-based models, under the MLP category.
[Section 5.1 References] The references Huang et al. 2019a and 2019b are identical (both cite 'Dsanet: Dual self-attention network for multivariate time series forecasting'); one should be removed or corrected.
[Section 4.5 and Figure 5] The text says Figure 5 presents results on the ILI dataset, but the figure caption reads 'Case study on the ETTh1 dataset'; this inconsistency should be resolved.
[Section 2.1] There is a typo: 'the output of the last HMA layer' should be 'the output of the last MHA layer'.
[Section 3.2 footnote] The footnote 'We employ PatchTST/42' is unclear; the notation '/42' is not defined and its connection to the historical window of 336 is not explained.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the gated-fusion definition and the background self-citations do not make the empirical accuracy claim circular, though the reported Table 3 averages are internally inconsistent.

full rationale

The paper's central claim is empirical: an architecture combining a Transformer encoder, an LLM encoder, and a gated fusion layer (Eq. 5) is trained end-to-end and evaluated on test sets. The statement that the fused representation ZE = g*ZLLM + (1-g)*Z_{l-1} 'inherently possesses both temporal dynamics and semantic understanding' is a definitional paraphrase of the convex combination chosen, not a derivation of forecasting accuracy from that definition; the paper does not claim to predict the error reduction from Eq. 5 alone. Self-citations (Su et al., 2025; Tian et al., 2025) appear only in Related Work and background context, and no load-bearing premise or uniqueness constraint is imported from those works. The most serious issue found is not circularity: Table 3's caption promises an 'Avg.' row across eight datasets, but only six datasets appear, no Avg row is printed, and the Section 4.1 assertion that 'the proposed approach achieves the lowest average MSE and average MAE error values' is contradicted by direct arithmetic on the six reported rows (Time-LLM has a lower average MSE and MAE). That is an internal-consistency and reproducibility problem, not a case where a prediction reduces to its inputs by construction. Accordingly, no circular step is identified and the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim is empirical and depends on the assumptions listed above. No new physical entities are introduced. The labeled free parameters are architecture choices that are not ablated, so the reported effect may not be robust.

free parameters (3)

Fusion layer index = 2
The LLM output is injected into the second layer of the 3-layer Transformer (Section 3.3). No ablation over fusion layer positions is provided, so the reported improvement may depend on this choice.
Gate formulation = Concatenation + linear + sigmoid, producing a scalar g
Eq. 5 uses a single scalar mixing weight. The paper does not test per-dimension gates or multi-layer fusion, leaving open whether this specific gate design is load-bearing.
Patch length and stride = T=16, S=8 (T=24, S=2 for ILI)
These hyperparameters follow PatchTST (Nie et al., 2022) and are not tuned or ablated in this paper, yet they define the token sequence seen by the LLM.

assumptions (3)

domain assumption Text-pretrained LLMs encode useful semantic patterns for linearly projected numerical time series tokens.
The entire method presupposes that GPT-2's pretrained representations, after linear projection and English prompts, carry information that helps forecasting. This is asserted in Sections 1, 2.2, and 4.3, with only qualitative attention maps as evidence.
ad hoc to paper A single scalar gate is sufficient to fuse the LLM and Transformer representation spaces.
Eq. 5 assumes a convex combination of two heterogeneous representation vectors. The paper provides no alignment analysis and no ablation of gate dimensionality.
domain assumption The standard TSF evaluation protocol transfers directly to this architecture.
The paper relies on standard benchmarks, MSE/MAE, and chronological splits without demonstrating that baselines were tuned to the same effort or that the patch/stride choices are appropriate for the LLM component.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Fusing Large Language Models with Temporal Transformers for Time Series Forecasting." pith.science (2026). https://pith.science/paper/CKVH3LDO

@misc{pith2026250710098,
  author       = {Pith},
  title        = {Pith review of: Fusing Large Language Models with Temporal Transformers for Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CKVH3LDO}},
  note         = {Machine review of arXiv:2507.10098}
}

read the original abstract

Recently, large language models (LLMs) have demonstrated powerful capabilities in performing various tasks and thus are applied by recent studies to time series forecasting (TSF) tasks, which predict future values with the given historical time series. Existing LLM-based approaches transfer knowledge learned from text data to time series prediction using prompting or fine-tuning strategies. However, LLMs are proficient at reasoning over discrete tokens and semantic patterns but are not initially designed to model continuous numerical time series data. The gaps between text and time series data lead LLMs to achieve inferior performance to a vanilla Transformer model that is directly trained on TSF data. However, the vanilla Transformers often struggle to learn high-level semantic patterns. In this paper, we design a novel Transformer-based architecture that complementarily leverages LLMs and vanilla Transformers, so as to integrate the high-level semantic representations learned by LLMs into the temporal information encoded by time series Transformers, where a hybrid representation is obtained by fusing the representations from the LLM and the Transformer. The resulting fused representation contains both historical temporal dynamics and semantic variation patterns, allowing our model to predict more accurate future values. Experiments on benchmark datasets demonstrate the effectiveness of the proposed approach.

Figures

Figures reproduced from arXiv: 2507.10098 by the authors.

**Figure 2.** Comparison of our approach and the model [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** (a) Attention visualization from the last layer of our approach’s Transformer encoder and (b) the second [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Comparison of different types and sizes of LLMs as time series encoders for forecasting. We report MSE [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: Case study on the ETTh1 dataset. The his [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Differencing the Diffusion Trajectory toward Uncertain Components for Time Series Forecasting
cs.AI 2026-06 conditional novelty 6.0 of 10

DiffDiff rewires diffusion forecasting so corruption gradually emphasizes second-order differences, concentrating generation on history-uncertain parts and improving forecasts on seven benchmarks.
Text Reinforcement for Multimodal Time Series Forecasting
cs.CL 2025-08 conditional novelty 6.0 of 10

Reinforcement learning trains an LLM to generate improved text from time series, improving multimodal forecasting on Time-MMD.
Diffusion Models for Time Series Forecasting: A Survey
stat.ML 2025-07 conditional novelty 4.0 of 10

A survey classifies diffusion-based time series forecasting models into a two-axis taxonomy by conditioning source and integration method.

Reference graph

Works this paper leans on

13 extracted references · 5 canonical work pages · cited by 3 Pith papers

[3]

arXiv preprint arXiv:2310.01728

Time-llm: Time series forecasting by reprogramming large language models. arXiv preprint arXiv:2310.01728. Luka Jovanovic, Nemanja Milutinovic, Masa Gajevic, Jelena Krstovic, Tarik A Rashid, and Aleksandar Petrovic. 2022. Sine cosine algo- rithm for simple recurrent neural network tun- ing for stock market prediction. In 2022 30th Telecommunications Forum...

arXiv 2022
[6]

ACM computing surveys, 55(9):1–35

Pre-train, prompt, and predict: A sys- tematic survey of prompting methods in natural language processing. ACM computing surveys, 55(9):1–35. Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. 2024b. Timer: Generative pre-trained trans- formers are large time series models. arXiv preprint arXiv:2402.02368. Zeyuan Liu, Ji...

arXiv 2021
[8]

arXiv preprint arXiv:1704.02971

A dual-stage attention-based recurrent neural network for time series prediction. arXiv preprint arXiv:1704.02971. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al

arXiv
[10]

Journal of machine learning research, 21(140):1–67

Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1–67. David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. 2020. Deepar: Prob- abilistic forecasting with autoregressive recur- rent networks. International journal of fore- casting, 36(3):1181–1191. Monir Yahya...

work page 2020
[11]

Advances in neural informa- tion processing systems, 32

Think globally, act locally: A deep neu- ral network approach to high-dimensional time series forecasting. Advances in neural informa- tion processing systems, 32. Chen Su, Yuanhe Tian, and Yan Song. 2025. Mul- timodal conditioned diffusive time series fore- casting. arXiv preprint arXiv:2504.19669. Chenxi Sun, Hongyan Li, Yaliang Li, and Shenda Hong. 202...

arXiv 2025
[2017]

Engineering applications of artifi- cial intelligence, 66:49–59

A hybrid ets–ann model for time series forecasting. Engineering applications of artifi- cial intelligence, 66:49–59. Kumar Prakhar, S Sountharrajan, E Suganya, M Karthiga, and Sathis Kumar. 2022. Effec- tive stock price prediction using time series forecasting. In 2022 6th International Confer- ence on Trends in Electronics and Informatics (ICOEI), pages ...

work page 2022
[2019]

OpenAI blog, 1(8):9

Language models are unsupervised mul- titask learners. OpenAI blog, 1(8):9. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu

work page
[2020]

applied sciences, 10(7):2322

Temporal convolutional networks ap- plied to energy-related time series forecasting. applied sciences, 10(7):2322. Brian Lester, Rami Al-Rfou, and Noah Con- stant. 2021. The power of scale for parameter- efficient prompt tuning. arXiv preprint arXiv:2104.08691. Chenxi Liu, Qianxiong Xu, Hao Miao, Sun Yang, Lingzheng Zhang, Cheng Long, Ziyue Li, and Rui Zh...

arXiv 2021

Show all 13 references

[2021]

In International conference on learn- ing representations

Reversible instance normalization for ac- curate time-series forecasting against distribu- tion shift. In International conference on learn- ing representations. Sushmitha Kothapalli and SG Totad. 2017. A real- time weather forecasting and analysis. In 2017 IEEE International ...

2017
[2022]

Sustain- ability, 14(3):1703

Analysis of construction cost and invest- ment planning using time series data. Sustain- ability, 14(3):1703. Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, et al
[2023]

arXiv preprint arXiv:2310.04948

Tempo: Prompt-based generative pre- trained transformer for time series forecasting. arXiv preprint arXiv:2310.04948. Anureet Chhabra, Sunil K Singh, Akash Sharma, Sudhakar Kumar, Brij B Gupta, Varsha Arya, and Kwok Tai Chui. 2024. Sustainable and intelligent time-series model...

arXiv
[2024]

arXiv preprint arXiv:2412.15115

Qwen2.5 technical report. arXiv preprint arXiv:2412.15115. Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. 2023. Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 11121–11128. Haoyi Zhou,...

2023 arXiv
[2025]

arXiv preprint arXiv:2505.09388

Qwen3 technical report. arXiv preprint arXiv:2505.09388. An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al

arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [3]

arXiv preprint arXiv:2310.01728

Time-llm: Time series forecasting by reprogramming large language models. arXiv preprint arXiv:2310.01728. Luka Jovanovic, Nemanja Milutinovic, Masa Gajevic, Jelena Krstovic, Tarik A Rashid, and Aleksandar Petrovic. 2022. Sine cosine algo- rithm for simple recurrent neural network tun- ing for stock market prediction. In 2022 30th Telecommunications Forum...

arXiv 2022

[2] [6]

ACM computing surveys, 55(9):1–35

Pre-train, prompt, and predict: A sys- tematic survey of prompting methods in natural language processing. ACM computing surveys, 55(9):1–35. Yong Liu, Haoran Zhang, Chenyu Li, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. 2024b. Timer: Generative pre-trained trans- formers are large time series models. arXiv preprint arXiv:2402.02368. Zeyuan Liu, Ji...

arXiv 2021

[3] [8]

arXiv preprint arXiv:1704.02971

A dual-stage attention-based recurrent neural network for time series prediction. arXiv preprint arXiv:1704.02971. Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al

arXiv

[4] [10]

Journal of machine learning research, 21(140):1–67

Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1–67. David Salinas, Valentin Flunkert, Jan Gasthaus, and Tim Januschowski. 2020. Deepar: Prob- abilistic forecasting with autoregressive recur- rent networks. International journal of fore- casting, 36(3):1181–1191. Monir Yahya...

work page 2020

[5] [11]

Advances in neural informa- tion processing systems, 32

Think globally, act locally: A deep neu- ral network approach to high-dimensional time series forecasting. Advances in neural informa- tion processing systems, 32. Chen Su, Yuanhe Tian, and Yan Song. 2025. Mul- timodal conditioned diffusive time series fore- casting. arXiv preprint arXiv:2504.19669. Chenxi Sun, Hongyan Li, Yaliang Li, and Shenda Hong. 202...

arXiv 2025

[6] [2017]

Engineering applications of artifi- cial intelligence, 66:49–59

A hybrid ets–ann model for time series forecasting. Engineering applications of artifi- cial intelligence, 66:49–59. Kumar Prakhar, S Sountharrajan, E Suganya, M Karthiga, and Sathis Kumar. 2022. Effec- tive stock price prediction using time series forecasting. In 2022 6th International Confer- ence on Trends in Electronics and Informatics (ICOEI), pages ...

work page 2022

[7] [2019]

OpenAI blog, 1(8):9

Language models are unsupervised mul- titask learners. OpenAI blog, 1(8):9. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu

work page

[8] [2020]

applied sciences, 10(7):2322

Temporal convolutional networks ap- plied to energy-related time series forecasting. applied sciences, 10(7):2322. Brian Lester, Rami Al-Rfou, and Noah Con- stant. 2021. The power of scale for parameter- efficient prompt tuning. arXiv preprint arXiv:2104.08691. Chenxi Liu, Qianxiong Xu, Hao Miao, Sun Yang, Lingzheng Zhang, Cheng Long, Ziyue Li, and Rui Zh...

arXiv 2021