Pith. sign in

REVIEW 4 major objections 4 minor 5 cited by

The paper argues that LLM-based time series forecasting improves when predictions are conditioned on a hierarchical memory of accumulated experience — historical patterns, distilled reasoning wisdom, and induced general laws — rather than t

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 ↗

MemCast claims LLM time-series forecasting improves when retrieval from a hierarchical memory of patterns, wisdom, and laws conditions reasoning, but the reported gains depend on a test-label-rewarded confidence update.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection The hierarchical memory design is worth a look, but the main empirical claim collapses under test-label leakage in the confidence-adaptation ablation. the 4 major comments →

arxiv 2602.03164 v2 pith:V6H673OK submitted 2026-02-03 cs.LG cs.AI

MemCast: Memory-Driven Time Series Forecasting with Experience-Conditioned Reasoning

Xiaoyu Tao , Mingyue Cheng , Ze Guo , Shuo Yu , Yaguo Liu , Qi Liu , Shijin Wang This is my paper
classification cs.LG cs.AI
keywords time series forecastinglarge language modelsexperience-conditioned reasoninghierarchical memoryretrieval-augmented generationdynamic confidence adaptationtrajectory selectiongeneral laws
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.

The reading

The paper tries to establish that large language model forecasters can accumulate and reuse experience across forecasting instances, turning time series forecasting into an experience-conditioned reasoning task. It proposes MemCast, which builds a hierarchical memory from the training set: historical patterns summarize past forecasting outcomes, reasoning wisdom distills successful and failed inference trajectories, and general laws capture extracted temporal regularities as rules. At inference, these memory modules guide case retrieval, trajectory selection, and reflective correction, while a dynamic confidence adaptation scheme re-weights memory entries during testing. The authors report that MemCast consistently outperforms statistical, deep learning, and other LLM-based baselines across ten benchmarks. If correct, this offers a route to lifelong forecasting improvement without retraining the LLM.

Core claim

The central claim is that forecasting should be reformulated as experience-conditioned reasoning: the LLM's output distribution becomes P(Y|X, C_sim), where C_sim is a set of retrieved historical cases, instead of P(Y|X). MemCast constructs three memory modules — M_his (paired lookback-window and outcome summaries), M_rea (distilled success and failure wisdom from reasoning trajectories), and M_gen (laws induced by having an LLM summarize extracted temporal features). Retrieval uses a composite similarity that blends feature-space cosine similarity with dynamic-time-warping-based structural similarity. A dynamic confidence adaptation updates a per-entry confidence score r_j during inference,

What carries the argument

The load-bearing mechanism is the hierarchical memory with experience-conditioned retrieval and confidence re-weighting. Historical patterns are semantic summaries of past windows; reasoning wisdom partitions generated trajectories into successful and failed cases and distills them; general laws are induced from extracted features and used as reflection criteria. At inference, top-k retrieval supplies cases via a composite similarity score S(x_q, x_k) = α·CosSim(f(x_q), f(x_k)) + (1−α)·exp(−DTW(x_q, x_k)/τ), blending semantic and structural match. The dynamic confidence adaptation is the component meant to provide continual evolution: it re-weights memory entries based on whether the current

Load-bearing premise

The reported continual-evolution gains rest on the legitimacy of updating each memory entry's confidence based on whether the current prediction beats a moving-average baseline in error, where both errors are computed against the true future value of the test instance.

What would settle it

Recompute MemCast's dynamic confidence adaptation while freezing all confidence scores at their training-derived values, or replacing the moving-average baseline with a reference that has no access to the test target; if the reported performance gap over the no-adaptation variant closes, the claimed gains are attributable to test-label access rather than to experience accumulation.

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

If this is right

  • LLM forecasters could improve with each encountered instance without retraining, since memory accumulates across queries.
  • Retrieval-augmented reasoning with explicit failure cases may reduce hallucinations under distribution shift, as the paper's error-mode analysis suggests.
  • The generate-then-select trajectory protocol implies active selection of reasoning paths can outperform simple ensembling for LLM forecasts.
  • General laws acting as reflection criteria provide a mechanism to enforce physical constraints on predictions, potentially transferable to other constrained generation tasks.

Where Pith is reading between the lines

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

  • The same confidence-adaptation idea could be lifted from forecasting to other retrieval-augmented LLM tasks, where per-entry weights are tuned on successful retrievals at inference time — but the paper does not discuss that transfer.
  • A testable extension: compare the reported moving-average baseline against a baseline that never sees the test target (e.g., a simple seasonal-naive forecast); if gains shrink, the adaptation is effectively using the test label rather than a legitimate reference.
  • The composite similarity S combining cosine and DTW could itself be evaluated on retrieval quality directly (e.g., hit rate of relevant cases), a metric the paper does not report; that would isolate whether gains come from retrieval or from the downstream selection.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 4 minor

Summary. MemCast is a memory-augmented LLM forecasting framework. It builds a hierarchical memory from training data (historical pattern summaries, distilled reasoning wisdom, and induced general laws), retrieves relevant entries during inference, samples multiple reasoning trajectories, selects among them with a semantic-consistency score, and performs rule-based reflection. A dynamic confidence adaptation mechanism updates per-entry confidence scores during inference. The paper claims consistent state-of-the-art results on ten benchmarks, with ablations showing each component helps.

Significance. The idea of accumulating reusable experience across forecasting instances, and explicitly storing failure wisdom, is timely and potentially useful. If the empirical claims held, the framework would be a practical way to improve LLM forecasters without retraining. The paper also ships a code link and covers diverse benchmarks. However, the central empirical support is invalidated by the use of test labels in the confidence-adaptation loop, and the main tables contain internal inconsistencies. The contribution as evaluated is therefore not established.

major comments (4)
  1. [§3.4 (Dynamic Confidence Adaptation)] This module updates confidence scores r_j using the rule 'A prediction is considered successful only if it demonstrates superior reasoning ability, i.e., L_LLM < L_MA.' Both losses are errors against the ground-truth target Y for the test instance. The updated r_j then modulates retrieval and trajectory selection (r_j > 0). The memory-selection policy is therefore being fit to test labels during inference, contradicting the same section's claim of 'strict train-test separation' and 'without leaking the test set distribution.' Figure 3's ablation does not isolate continual evolution; it partly measures the benefit of reweighting memory by test outcomes. The comparison against static baselines is thus unfair.
  2. [Table 2 / Table 3] Table 2, ETTm column, TimeReasoner row reports MSE=1.574 with MAE=3.625. For any forecast, MSE = mean(e_i^2) >= (mean|e_i|)^2 = MAE^2, so this pair is impossible (1.574 >= 13.14 would be required). In addition, MemCast's DE result is 210.445/9.195 in Table 2 but 282.801/10.903 in the 'Ours' row of Table 3. Such inconsistencies in the principal results make it difficult to treat the reported numbers as reliable.
  3. [§4.1 and Figures 3-6] The experiments use GPT-5 with temperature 0.6 and top-p 0.7, so the LLM outputs are stochastic. Yet Tables 2-6 and Figures 3, 4, and 6 report single deterministic-looking numbers with no standard deviations, error bars, or significance tests. Some reported differences are small (e.g., Figure 3 for long horizons); without repeated sampling it is impossible to know whether the proposed components give robust gains or merely reflect sampling noise.
  4. [Abstract and §4.2] The abstract claims MemCast 'consistently outperforms previous methods,' but even the paper's own Table 2 shows counterexamples: on MOPEX, TimeReasoner reports 5.465/1.442 versus MemCast's 6.193/1.525, and on SP TimeReasoner reports 17.599/1.847 versus MemCast's 22.998/1.876. The body text carefully says 'best or second-best,' but the abstract's blanket claim is unsupported by the reported numbers and should be qualified or corrected.
minor comments (4)
  1. [§3.4] The scoring function φ(·) used for trajectory selection is never defined. 'Semantic consistency of the trajectory against retrieved reasoning wisdom' is too vague to replicate; please provide the exact computation or pseudocode.
  2. [§3.3 and Eq. (1)] The symbol τ is used both as the success/failure threshold in reasoning wisdom distillation and as the temperature-like scaling in the DTW similarity S_str = exp(-DTW(x_q,x_k)/τ). This notational clash confuses two different quantities.
  3. [Appendix A.1] The MOPEX benchmark is described as a hydrological streamflow dataset, but the cited reference (Makovoz & Marleau, 2005) is an astronomical point-source extraction paper. Please verify and replace with the correct MOPEX/hydrology reference.
  4. [Appendix A.1 / Table 1] Minor typos: 'Sundy Power' should be 'Sunny Power'; also the MOPEX entry in Table 1 lists '6' variables while the appendix text names four meteorological drivers plus streamflow, which should be reconciled.

Circularity Check

1 steps flagged

Dynamic confidence adaptation in §3.4 updates memory confidence r_j using the test-target-based rule L_LLM < L_MA, so Figure 3's claimed gains are contaminated by test-label leakage.

specific steps
  1. fitted input called prediction [Section 3.4, Dynamic Confidence Adaptation; Figure 3 (cf. Section 3.3 for definition of L)]
    "During inference, the generated prediction Y is evaluated against a moving average (MA) baseline. A prediction is considered successful only if it demonstrates superior reasoning ability, i.e., LLLM < LMA. In such cases, the confidence scores of the contributing experiences are increased to reinforce future selection... When rj > 0, experience utilization considers both the confidence score and the retrieval similarity score S. This design updates only confidence weights rather than memory content, ensuring strict train–test separation and continual memory evolution."

    In this paper, L denotes prediction error against the ground-truth target (Section 3.3 defines e_i = L(ˆy(c_i), y_i)). Therefore evaluating LLLM < LMA requires the true Y of the current test instance. Increasing r_j for memories that contributed to a successful test prediction, and then using r_j in retrieval for later test instances, fits the selection policy to test labels. Figure 3 then credits the resulting lower errors to 'dynamic confidence adaptation'. This is fitting a parameter to the test set and presenting the fitted behavior as an improvement; the paper's own no-leakage claim ('strict train–test separation') is contradicted by the update rule.

full rationale

The offline experience-accumulation modules (historical patterns, reasoning wisdom, general laws) are built from training-set targets and are not circular by themselves. However, the paper's load-bearing claim of 'continual memory evolution' 'without leaking the test set distribution' rests on the Dynamic Confidence Adaptation step in §3.4, and that step is fit to the test set. The rule LLLM < LMA requires the ground-truth value of the current test instance; updating r_j on that basis and then using r_j in retrieval for subsequent test instances adapts the selector to test labels. Figure 3 presents this adaptation as a model gain, but the gain is contaminated by test-label leakage. The paper even states that 'naively incorporating test-time outcomes can bias the model toward the test distribution,' which is exactly what its own update does. The central empirical support for the adaptation component therefore reduces to fitting confidence weights to test outcomes, so the no-leakage conclusion is unsupported and the adaptation ablation is not independent evidence.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 4 invented entities

The central claim rests on a set of LLM-oracle behaviors and on a test-label-driven confidence update. The only numerical parameters in the method are hyperparameters; no formal derivations are offered. The invented memory components are conceptual entities with no independent falsifiable evidence.

free parameters (7)
  • α (semantic-structural weight) = not reported
    In Eq. (1), S = αSsem + (1−α)Sstr; no value or tuning procedure is given.
  • threshold τ for success/failure split = not reported
    Trajectories are partitioned into successful/failed by ei < τ; τ is never specified.
  • retrieval top-k = k=3
    Sensitivity analysis in Table 6 selects k=3; this is a tuned hyperparameter.
  • number of sampled trajectories M = not fixed; Fig 6(a) shows m>4 saturates
    The sampling count is a hyperparameter that changes results, tuned per dataset presumably.
  • similarity thresholds for dedup/merge = 0.95 / 0.8
    Hard-coded thresholds for replacing redundant and fusing overlapping cases in wisdom distillation.
  • temperature / top-p = 0.6 / 0.7
    Decoding hyperparameters; Figure 6(b) shows temperature matters and 0.6 was chosen.
  • moving-average baseline (MA) = unspecified
    The baseline used for LMA is not defined (window, method), yet it determines which memories get rewarded.
axioms (5)
  • domain assumption LLM summarization S preserves forecasting-relevant information when converting numerical windows to text.
    Section 3.3 uses S(y) to create memory summaries; no evidence that text captures enough numerical detail.
  • domain assumption Retrieved training-set patterns transfer to test instances.
    The retrieval assumes similarity in input space implies similarity in future dynamics; standard but unverified here.
  • ad hoc to paper Ground-truth labels are available at inference time for updating confidence scores.
    Dynamic confidence adaptation requires LLLM and LMA, both computed against the true y; this is the leakage premise.
  • domain assumption Induced general laws are correct domain constraints.
    LLM-induced rules such as 'Peak > 0.98 → Risk' are taken as truth for reflection; no validation is provided.
  • domain assumption The LLM's generated reasoning trajectories are meaningful and scoreable by semantic consistency.
    The selection function φ measures consistency but its definition is not given.
invented entities (4)
  • Historical Pattern memory (Mhis) no independent evidence
    purpose: Store textual summaries of past input/future pairs for analogical retrieval.
    No external validation that these summaries improve generalization beyond this paper's own results.
  • Reasoning Wisdom memory (Wpos, Wneg) no independent evidence
    purpose: Distill successful/failed reasoning trajectories to guide trajectory selection.
    No metric shows wisdom is valid outside the paper's own evaluation.
  • General Law memory (Mgen) no independent evidence
    purpose: Induce rules from training features to constrain predictions.
    Rules are LLM-generated and unverified; no external benchmark.
  • Confidence score r_j no independent evidence
    purpose: Per-entry mutable weight that adapts memory selection during inference.
    Updated with ground-truth test outcomes, so its 'evidence' is the leakage itself.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of MemCast: Memory-Driven Time Series Forecasting with Experience-Conditioned Reasoning." pith.science (2026). https://pith.science/paper/V6H673OK

@misc{pith2026260203164,
  author       = {Pith},
  title        = {Pith review of: MemCast: Memory-Driven Time Series Forecasting with Experience-Conditioned Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V6H673OK}},
  note         = {Machine review of arXiv:2602.03164}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Time series forecasting (TSF) plays a critical role in decision-making for many real-world applications. Recently, large language model (LLM)- based forecasters have made promising advancements. Despite their effectiveness, existing methods often lack explicit experience accumulation and continual evolution. In this work, we propose MemCast, a learning-to-memory framework that reformulates TSF as an experience-conditioned reasoning task. Specifically, we learn experience from the training set and organize it into a hierarchical memory. This is achieved by summarizing prediction results into historical patterns, distilling inference trajectories into reasoning wisdom, and inducing extracted temporal features into general laws. Furthermore, during inference, we leverage historical patterns to guide the reasoning process and utilize reasoning wisdom to select better trajectories, while general laws serve as criteria for reflective iteration. Additionally, to enable continual evolution, we design a dynamic confidence adaptation strategy that updates the confidence of individual entries without leaking the test set distribution. Extensive experiments on multiple datasets demonstrate that MemCast consistently outperforms previous methods, validating the effectiveness of our approach. Our code is available at https://github.com/Xiaoyu-Tao/MemCast-TS.

Figures

Figures reproduced from arXiv: 2602.03164 by Mingyue Cheng, Qi Liu, Shijin Wang, Shuo Yu, Xiaoyu Tao, Yaguo Liu, Ze Guo.

Figure 1
Figure 1. Figure 1: Comparison of training-based, training-free, and memory-enhanced LLM forecasting approaches. a mapping from historical time series and associated con￾textual features, including dynamic features that vary over time (e.g., weather information) and static features that re￾main invariant across the forecasting horizon (e.g., location attributes) , to future outcomes (Cheng et al., 2025a). Building upon this f… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of MemCast as an LLM-driven time series forecasting framework that constructs hierarchical memory from the training set via experience accumulation for experience-conditioned reasoning on the testing set. 3.2. Framework Overview [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation on dynamic confidence adaptation. Dynamic confidence adaptation leads to consistently lower errors compared with the variant without adaptation. 4.3. Ablation Studies Ablation on Hierarchical Memory [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Exploration on aggregation strategies. The full model outperforms baselines, confirming that active selection via seman￾tic consistency surpasses passive aggregation [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: A detailed case study on the oil temperature (OT) forecasting task, illustrating experience-conditioned reasoning via constructed memory to effectively handle out-of-distribution thermal shifts. MSE 45 40 35 30 7 6 5 8 1 2 3 4 5 (a) Sampled Trajectories NP PJM ETTh (b) Temperature MSE 45 40 35 30 7 6 5 8 0.2 0.4 0.6 0.8 1.0 NP PJM ETTh [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Hyperparameter sensitivity analysis regarding the num￾ber of sampled trajectories (a) and sampling temperature (b). consistently yields the most favorable performance. At higher temperatures, the accuracy deteriorates, likely due to the injection of excessive randomness leading to unsta￾ble reasoning and hallucinations inconsistent with temporal constraints. Conversely, an extremely low temperature also re… view at source ↗
Figure 7
Figure 7. Figure 7 [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of typical failure modes: (a) Copy-Paste Repeat (naive history replication); (b) Wrong Trend (significant trajectory divergence); (c) Peak-Overestimation (amplitude exaggeration); and (d) Constant Collapse (degeneration into uninformative flat lines). • TimeXer (Wang et al., 2024): An advanced Transformer framework designed to effectively empower time series forecasting by incorporating and a… view at source ↗

discussion (0)

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

Forward citations

Cited by 5 Pith papers

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

  1. TSSM: Triaxial State Space Model for Global Station Weather Forecasting with Temporal-Variable-Historical Modeling

    cs.LG 2026-07 conditional novelty 6.0

    A weather forecasting model that scans time, variables, and aligned past years jointly beats prior station-forecasting baselines on Weather-5K, with particularly large reported gains on extreme-event and long-horizon metrics.

  2. GeoDecider: A Coarse-to-Fine Agentic Workflow for Explainable Lithology Classification

    cs.AI 2026-05 unverdicted novelty 6.0

    GeoDecider introduces a coarse-to-fine agentic workflow using LLMs for explainable lithology classification from well logs, combining a base classifier, tool-augmented reasoning, and geological refinement to outperfor...

  3. GeoMind: An Agentic Workflow for Lithology Classification with Reasoned Tool Invocation

    cs.AI 2026-04 unverdicted novelty 6.0

    GeoMind applies an agentic workflow with tool-augmented modules and process supervision to outperform static models on lithology classification from well logs while producing traceable decisions.

  4. TimeRFT: Stimulating Generalizable Time Series Forecasting for TSFMs via Reinforcement Finetuning

    eess.SP 2026-04 conditional novelty 6.0

    TimeRFT fine-tunes time-series foundation models with step-wise reward signals and difficulty-filtered data, beating supervised fine-tuning on eight benchmarks across data regimes.

  5. TimeRFT: Stimulating Generalizable Time Series Forecasting for TSFMs via Reinforcement Finetuning

    eess.SP 2026-04 unverdicted novelty 5.0

    TimeRFT applies reinforcement learning with multi-faceted step-wise rewards and informative sample selection to improve generalization and accuracy in TSFM adaptation beyond supervised fine-tuning.

Reference graph

Works this paper leans on

44 extracted references · 9 linked inside Pith · cited by 4 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Multi-patch prediction: Adapting llms for time series representation learning

    Bian, Y., Ju, X., Li, J., Xu, Z., Cheng, D., and Xu, Q. Multi-patch prediction: Adapting llms for time series representation learning. International Conference on Machine Learning, 2024

    2024

  3. [3]

    Remember me, refine me: A dynamic procedural memory framework for experience-driven agent evolution

    Cao, Z., Deng, J., Yu, L., Zhou, W., Liu, Z., Ding, B., and Zhao, H. Remember me, refine me: A dynamic procedural memory framework for experience-driven agent evolution. arXiv preprint arXiv:2512.10696, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    A comprehensive survey of time series forecasting: Concepts, challenges, and future directions

    Cheng, M., Liu, Z., Tao, X., Liu, Q., Zhang, J., Pan, T., Zhang, S., He, P., Zhang, X., Wang, D., et al. A comprehensive survey of time series forecasting: Concepts, challenges, and future directions. Authorea Preprints, 2025 a

    2025

  5. [5]

    Can slow-thinking llms reason over time? empirical studies in time series forecasting

    Cheng, M., Wang, J., Wang, D., Tao, X., Liu, Q., and Chen, E. Can slow-thinking llms reason over time? empirical studies in time series forecasting. arXiv preprint arXiv:2505.24511, 2025 b

    arXiv 2025

  6. [6]

    Convtimenet: A deep hierarchical fully convolutional model for multivariate time series analysis

    Cheng, M., Yang, J., Pan, T., Liu, Q., Li, Z., and Wang, S. Convtimenet: A deep hierarchical fully convolutional model for multivariate time series analysis. In Companion Proceedings of the ACM on Web Conference 2025, pp.\ 171--180, 2025 c

    2025

  7. [7]

    T., Xu, M

    Chow, W., Gardiner, L., Hallgr \' msson, H. T., Xu, M. A., and Ren, S. Y. Towards time series reasoning with llms. arXiv preprint arXiv:2409.11376, 2024

    Pith/arXiv arXiv 2024

  8. [8]

    Temporal relational ranking for stock prediction

    Feng, F., He, X., Wang, X., Luo, C., Liu, Y., and Chua, T.-S. Temporal relational ranking for stock prediction. ACM Transactions on Information Systems (TOIS), 37 0 (2): 0 1--30, 2019

    2019

  9. [9]

    Gruver, N., Finzi, M., Qiu, S., and Wilson, A. G. Large language models are zero-shot time series forecasters. Advances in Neural Information Processing Systems, 36: 0 19622--19635, 2023

    2023

  10. [10]

    Exploiting language power for time series forecasting with exogenous variables

    Huang, Q., Zhou, Z., Yang, K., and Wang, Y. Exploiting language power for time series forecasting with exogenous variables. In Proceedings of the ACM on Web Conference 2025, pp.\ 4043--4052, 2025 a

    2025

  11. [11]

    Timebase: The power of minimalism in efficient long-term time series forecasting

    Huang, Q., Zhou, Z., Yang, K., Yi, Z., Wang, X., and Wang, Y. Timebase: The power of minimalism in efficient long-term time series forecasting. In Forty-second International Conference on Machine Learning, 2025 b

    2025

  12. [12]

    Hyndman, R. J. and Khandakar, Y. Automatic time series forecasting: the forecast package for r. Journal of statistical software, 27: 0 1--22, 2008

    2008

  13. [13]

    2025 iflytek renewable power forecasting challenge (wind & solar)

    iFLYTEK AI Challenge . 2025 iflytek renewable power forecasting challenge (wind & solar). https://challenge.xfyun.cn/topic/info?type=renewable-power-forecast&option=ssgy&ch=dwsf259, 2025. Accessed: 2026-01

    2025

  14. [14]

    Gpt4mts: Prompt-based large language model for multimodal time-series forecasting

    Jia, F., Wang, K., Zheng, Y., Cao, D., and Liu, Y. Gpt4mts: Prompt-based large language model for multimodal time-series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 23343--23351, 2024

    2024

  15. [15]

    Timexl: Explainable multi-modal time series prediction with llm-in-the-loop

    Jiang, Y., Yu, W., Lee, G., Song, D., Shin, K., Cheng, W., Liu, Y., and Chen, H. Timexl: Explainable multi-modal time series prediction with llm-in-the-loop. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  16. [16]

    Y., Shi, X., Chen, P.-Y., Liang, Y., Li, Y.-F., Pan, S., et al

    Jin, M., Wang, S., Ma, L., Chu, Z., Zhang, J. Y., Shi, X., Chen, P.-Y., Liang, Y., Li, Y.-F., Pan, S., et al. Time-llm: Time series forecasting by reprogramming large language models. In The Twelfth International Conference on Learning Representations, 2024

    2024

  17. [17]

    Forecasting day-ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark

    Lago, J., Marcjasz, G., De Schutter, B., and Weron, R. Forecasting day-ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark. Applied Energy, 293: 0 116983, 2021

    2021

  18. [18]

    Temporal query network for efficient multivariate time series forecasting

    Lin, S., Chen, H., Wu, H., Qiu, C., and Lin, W. Temporal query network for efficient multivariate time series forecasting. In Forty-second International Conference on Machine Learning, 2025

    2025

  19. [19]

    Timecma: Towards llm-empowered multivariate time series forecasting via cross-modality alignment

    Liu, C., Xu, Q., Miao, H., Yang, S., Zhang, L., Long, C., Li, Z., and Zhao, R. Timecma: Towards llm-empowered multivariate time series forecasting via cross-modality alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 18780--18788, 2025 a

    2025

  20. [20]

    Liu, H., Zhao, Z., Wang, J., Kamarthi, H., and Prakash, B. A. Lstprompt: Large language models as zero-shot time series forecasters by long-short-term prompting. arXiv preprint arXiv:2402.16132, 2024 a

    Pith/arXiv arXiv 2024

  21. [21]

    Calf: Aligning llms for time series forecasting via cross-modal fine-tuning

    Liu, P., Guo, H., Dai, T., Li, N., Bao, J., Ren, X., Jiang, Y., and Xia, S.-T. Calf: Aligning llms for time series forecasting via cross-modal fine-tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 18915--18923, 2025 b

    2025

  22. [22]

    Non-stationary transformers: Exploring the stationarity in time series forecasting

    Liu, Y., Wu, H., Wang, J., and Long, M. Non-stationary transformers: Exploring the stationarity in time series forecasting. Advances in neural information processing systems, 35: 0 9881--9893, 2022

    2022

  23. [23]

    itransformer: Inverted transformers are effective for time series forecasting

    Liu, Y., Hu, T., Zhang, H., Wu, H., Wang, S., Ma, L., and Long, M. itransformer: Inverted transformers are effective for time series forecasting. In The Twelfth International Conference on Learning Representations, 2024 b

    2024

  24. [24]

    Time series forecasting as reasoning: A slow-thinking approach with reinforced llms

    Luo, Y., Zhou, Y., Cheng, M., Wang, J., Wang, D., Pan, T., and Zhang, J. Time series forecasting as reasoning: A slow-thinking approach with reinforced llms. arXiv preprint arXiv:2506.10630, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    and Marleau, F

    Makovoz, D. and Marleau, F. R. Point-source extraction with mopex. Publications of the Astronomical Society of the Pacific, 117 0 (836): 0 1113, 2005

    2005

  26. [26]

    Nguyen, N., Sinthong, P., and Kalagnanam, J

    Nie, Y., H. Nguyen, N., Sinthong, P., and Kalagnanam, J. A time series is worth 64 words: Long-term forecasting with transformers. In International Conference on Learning Representations, 2023

    2023

  27. [27]

    S 2 ip-llm: Semantic space informed prompt learning with llm for time series forecasting

    Pan, Z., Jiang, Y., Garg, S., Schneider, A., Nevmyvaka, Y., and Song, D. S 2 ip-llm: Semantic space informed prompt learning with llm for time series forecasting. In Forty-first International Conference on Machine Learning, 2024

    2024

  28. [28]

    S., Sheng, Z., et al

    Qiu, X., Hu, J., Zhou, L., Wu, X., Du, J., Zhang, B., Guo, C., Zhou, A., Jensen, C. S., Sheng, Z., et al. Tfb: Towards comprehensive and fair benchmarking of time series forecasting methods. arXiv preprint arXiv:2403.20150, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Large language models for time series analysis: Techniques, applications, and challenges

    Shi, F., Yin, X., Wang, K., Tu, W., Sun, Q., and Ning, H. Large language models for time series analysis: Techniques, applications, and challenges. arXiv preprint arXiv:2506.11040, 2025 a

    Pith/arXiv arXiv 2025

  30. [30]

    Time-moe: Billion-scale time series foundation models with mixture of experts

    Shi, X., Wang, S., Nie, Y., Li, D., Ye, Z., Wen, Q., and Jin, M. Time-moe: Billion-scale time series foundation models with mixture of experts. In The Thirteenth International Conference on Learning Representations, 2025 b

    2025

  31. [31]

    Openai gpt-5 system card

    Singh, A., Fry, A., Perelman, A., Tart, A., Ganesh, A., El-Kishky, A., McLaughlin, A., Low, A., Ostrow, A., Ananthram, A., et al. Openai gpt-5 system card. arXiv preprint arXiv:2601.03267, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    Time series forecasting with llms: Understanding and enhancing model capabilities

    Tang, H., Zhang, C., Jin, M., Yu, Q., Wang, Z., Jin, X., Zhang, Y., and Du, M. Time series forecasting with llms: Understanding and enhancing model capabilities. ACM SIGKDD Explorations Newsletter, 26 0 (2): 0 109--118, 2025

    2025

  33. [33]

    Taylor, S. J. and Letham, B. Forecasting at scale. The American Statistician, 72 0 (1): 0 37--45, 2018

    2018

  34. [34]

    O., Ildiz, M

    Tire, K., Taga, E. O., Ildiz, M. E., and Oymak, S. Retrieval augmented time series forecasting. arXiv preprint arXiv:2411.08249, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Timedart: A diffusion autoregressive transformer for self-supervised time series representation, 2025

    Wang, D., Cheng, M., Liu, Z., and Liu, Q. Timedart: A diffusion autoregressive transformer for self-supervised time series representation, 2025

    2025

  36. [36]

    Deep factors for forecasting

    Wang, Y., Smola, A., Maddix, D., Gasthaus, J., Foster, D., and Januschowski, T. Deep factors for forecasting. In International conference on machine learning, pp.\ 6607--6617. PMLR, 2019

    2019

  37. [37]

    Timexer: Empowering transformers for time series forecasting with exogenous variables

    Wang, Y., Wu, H., Dong, J., Qin, G., Zhang, H., Liu, Y., Qiu, Y., Wang, J., and Long, M. Timexer: Empowering transformers for time series forecasting with exogenous variables. Advances in Neural Information Processing Systems, 37: 0 469--498, 2024

    2024

  38. [38]

    Winters, P. R. Forecasting sales by exponentially weighted moving averages. Management science, 6 0 (3): 0 324--342, 1960

    1960

  39. [39]

    and Salim, F

    Xue, H. and Salim, F. D. Promptcast: A new prompt-based learning paradigm for time series forecasting. IEEE Transactions on Knowledge and Data Engineering, 36 0 (11): 0 6851--6864, 2023

    2023

  40. [40]

    Timerag: Boosting llm time series forecasting via retrieval-augmented generation

    Yang, S., Wang, D., Zheng, H., and Jin, R. Timerag: Boosting llm time series forecasting via retrieval-augmented generation. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pp.\ 1--5. IEEE, 2025

    2025

  41. [41]

    Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pp.\ 11121--11128, 2023

    Zeng, A., Chen, M., Zhang, L., and Xu, Q. Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pp.\ 11121--11128, 2023

    2023

  42. [42]

    Irregular multivariate time series forecasting: A transformable patching graph neural networks approach

    Zhang, W., Yin, C., Liu, H., Zhou, X., and Xiong, H. Irregular multivariate time series forecasting: A transformable patching graph neural networks approach. In Forty-first International Conference on Machine Learning, 2024

    2024

  43. [43]

    X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al

    Zhao, W. X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., Min, Y., Zhang, B., Zhang, J., Dong, Z., et al. A survey of large language models. arXiv preprint arXiv:2303.18223, 1 0 (2), 2023

    Pith/arXiv arXiv 2023

  44. [44]

    Informer: Beyond efficient transformer for long sequence time-series forecasting

    Zhou, H., Zhang, S., Peng, J., Zhang, S., Li, J., Xiong, H., and Zhang, W. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pp.\ 11106--11115, 2021

    2021

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.