Forecasting Time Series with LLMs via Patch-Based Prompting and Decomposition

Anish Gupta; An La; Anshul Vemulapalli; Franck Dernoncourt; Hongjie Chen; Mayank Bumb; Nesreen K. Ahmed; Ryan A. Rossi; Sri Harsha Vardhan Prasad Jella; Yu Wang

REVIEW 6 major objections 8 minor 12 references

Forecasting Time Series with LLMs via Patch-Based Prompting and Decomposition

T0 review · 6 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read PatchInstruct shows that a purely prompt-based method—reversed overlapping patches plus structured instructions—can make a pretrained LLM beat a fine-tuned LLM forecaster on short horizons, at a tenth to a hundredth of the inference cost.

desk verdict PatchInstruct is a plausible prompt-based recipe for LLM forecasting, but the paper's own tables contradict its 'consistent outperformance' claim and the prompt selection appears to have used the test set, so the results are conditional at best. read the letter →

arxiv 2506.12953 v1 pith:3EG7EJT3 submitted 2025-06-15 cs.LG cs.AIcs.CL

Mayank Bumb , Anshul Vemulapalli , Sri Harsha Vardhan Prasad Jella , Anish Gupta , An La , Ryan A. Rossi , Hongjie Chen , Franck Dernoncourt

show 2 more authors

Nesreen K. Ahmed Yu Wang

This is my paper · ORCID

classification cs.LGcs.AIcs.CL

keywords timeseriesforecastinglargelanguagemodelspromptengineeringpatch-basedtokenizationzero-shotneighboraugmentationinferenceefficiencydecomposition

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

This paper tries to establish that a pretrained large language model can forecast time series accurately at short horizons without any fine-tuning or architectural change, if the input history is repackaged as reversed overlapping patches and the model is given explicit instructions to reproduce those patches before predicting. On the Weather and Traffic datasets, the proposed PatchInstruct method beats S2IP-LLM, a fine-tuned semantic-alignment baseline, on most horizons up to 12, with a 10 to 100 times reduction in inference time. The authors read this as evidence that prompt engineering can substitute for some architectural complexity in LLM-based forecasting. If correct, it would make short-horizon forecasting with LLMs dramatically cheaper to deploy, though the evidence is confined to two datasets and prompts selected with knowledge of them.

What carries the argument

The load-bearing object is the reverse-patch tokenization: the 96-value history is split into overlapping windows of three values with stride one, and the resulting patches are listed in reverse chronological order so the most recent window comes first, giving the LLM a recency-first view of local temporal structure. The prompt also fixes an output format that forces the model to produce the patch list before the prediction, acting as a structured scratchpad. Neighbor augmentation optionally adds the five most similar past series, chosen by euclidean distance over previous windows, to supply cross-series context.

What would settle it

Apply the exact PatchInstruct prompt (window 3, stride 1, reversed patches) with no modification to a fresh dataset such as M4 or ETTm2 at horizons 1-12 and compare MSE and MAE against S2IP-LLM; if PatchInstruct does not win on a majority of horizons, the claimed general advantage is refuted.

Watch

Extended reading notes

Core claim

The central claim is that structuring a prompt as a sequence of reversed overlapping patches—window size 3, stride 1, newest patch first—along with instructions to output the patches and then the next h values, lets an off-the-shelf instruction-tuned LLM forecast better than S2IP-LLM, which fine-tunes the LLM with a learned patch encoder and semantic alignment. Across horizons 1, 2, 3, 4, 5, 6 and 12 on Weather and Traffic, PatchInstruct obtains lower MSE and MAE than S2IP-LLM in almost every short-horizon case, and reduces per-forecast wall-clock time from hundreds of seconds to about one second. The paper also shows that adding five nearest-neighbor series to the prompt helps on Weather but can hurt on Traffic, indicating that neighbor context is useful only when the retrieved series are genuinely correlated.

Load-bearing premise

The prompt templates were selected after trying several variants on the same Weather and Traffic datasets used for the final evaluation, so the reported accuracy advantage may reflect tuning to those benchmarks rather than a general property of patch-based prompting.

Editorial extensions

If this is right

At horizons up to 12, PatchInstruct beats the fine-tuned S2IP-LLM baseline on MSE and MAE for both Weather and Traffic, with the largest gain at horizon 1 (a 97.7% MSE reduction on Weather).
Inference cost drops from minutes to roughly one second per forecast, a 10-100x speedup, at the price of about 800-1000 input tokens per prediction.
Neighbor-augmented prompts improve Weather forecasts at several horizons but degrade Traffic forecasts, so the value of cross-series context depends on neighbor relevance rather than on the prompting alone.
Ablations over five patch variants show reverse-ordered overlapping patches is the most consistent choice at horizons 1, 3 and 6 across both datasets.
Because no training is required, the method can be applied to any instruction-tuned LLM by changing only the prompt, lowering the engineering barrier for deploying LLM forecasters.

Reading between the lines

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

One reading the paper leaves implicit: the reversed-patch ordering may matter mainly because it forces the model to emit a long structured scratchpad before the prediction; a controlled comparison against natural-order patches with the same scratchpad would separate the ordering effect from the output-format effect.
The prompt-selection procedure on the evaluation datasets means the accuracy numbers are optimistic; a fair test would fix the prompt template on a development split and only then touch the test data, or apply the exact Weather/Traffic prompt to a third dataset.
The method's dependence on a fixed 96-point window and three-value patches ties it to the evaluation setup; the Meta Patches variant, which appends slot indices, hints at an untested route to irregular and multi-frequency series that the paper does not develop.
If the scratchpad interpretation is right, PatchInstruct should transfer to other sequence tasks such as anomaly detection or imputation with minimal prompt edits, an extension the authors mention only as future work.

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

You should read this paper for the prompting recipe, not for the headline claim. PatchInstruct is a specific combination of overlapping patches, reversed order, and neighbor augmentation for zero-shot LLM time series forecasting. That combination is new enough to be a legitimate incremental contribution, and the prompts are described in enough detail to re-implement. The runtime comparison is also useful: the method is orders of magnitude faster than the fine-tuned S2IP-LLM baseline, which matters if you're deploying LLM forecasters.

That said, the paper's central assertion—that PatchInstruct consistently outperforms S2IP-LLM on short horizons—is not supported by its own tables. Traffic H=12 is a direct counterexample: PatchInstruct's MSE is 235.75 versus S2IP-LLM's 68.71. So "consistently" is false. The internal numbers are also inconsistent: Section 4.3 cites a 97.7% MSE reduction on Weather H=1, but Table 2 shows a reduction from 0.0095 to 0.0014 (about 85%), and the Traffic H=1 claim of an 85% reduction doesn't match Table 2 at all (roughly 5%). Appendix A's Table 5 reports MSE 0.0005 for the same Reverse Patches setting that Table 2 reports as 0.0014. That's not an evaluation pipeline you can trust.

The bigger problem is methodological. Section 3.2 says the prompts were "designed through rigorous empirical testing," and Appendix A selects Reverse Patches as the best variant after comparing five strategies on the same Weather and Traffic horizons used in Table 2. There's no held-out split for prompt selection. So the reported PatchInstruct numbers are the best of several prompt designs on the test set. That inflates the apparent advantage. The paper itself acknowledges in the Limitations section that the framework is "heavily contingent upon carefully engineered prompts" and risks overfitting—that's honest, but it undercuts the main claim.

Here's my take: the idea is plausible and worth testing, but the current manuscript overstates its results. A serious revision would need (1) a held-out validation set for prompt choice, (2) consistent numbers across tables and text, and (3) error bars or significance tests. Without those, I can't endorse the claim that prompt engineering can replace architectural complexity.

For peer review: I'd send it. The topic is relevant, the prompt recipe is concrete, and the flaws are correctable. A good referee could help the authors get this into a solid empirical paper. But as it stands, it's a conditional result at best.

Referee Report

6 major / 8 minor

Summary. The paper proposes PatchInstruct, a prompt-only framework for LLM-based time series forecasting that tokenizes the input window into overlapping patches with reverse ordering, optionally augmented with nearest-neighbor series. The method is evaluated without fine-tuning on Weather and Traffic datasets for horizons 1 through 12, with GPT-4/GPT-4o as backbones, and compared against S2IP-LLM and zero-shot baselines. The paper claims that PatchInstruct consistently outperforms baselines on short horizons while reducing inference time by one to two orders of magnitude.

Significance. If substantiated, the result would support the practically interesting claim that prompt engineering can replace some architectural complexity in LLM forecasting, avoiding expensive fine-tuning. The paper contributes explicitly specified prompting templates and an ablation of patch-based variants, and it uses public datasets. However, the central claim is not currently supported: the reported numbers contradict the headline consistency, the prompt design appears to have been selected on the same data used for evaluation, and several quantitative statements in the text do not match the tables.

major comments (6)

[§4.2, Table 2] The headline claim that PatchInstruct consistently outperforms S2IP-LLM is contradicted by Table 2: at Traffic H=12, PatchInstruct has MSE 235.75 versus 68.71 for S2IP-LLM, and at Weather H=3 its MAE is 0.085 versus 0.068 for S2IP-LLM. The paper should either qualify the claim to the horizons and datasets where it holds or provide an explanation for the failure at longer horizons.
[§4.3] The cost–performance analysis is inconsistent with Table 2. The text reports a 97.7% MSE reduction on Weather H=1 (from 1.15e-2 to 2.6e-4) and an 85% reduction on Traffic H=1, but Table 2 shows a reduction from 0.0095 to 0.0014 (approximately 85%) and from 21.08 to 20.05 (approximately 5%). The same section quotes a runtime of 0.86s for 'Reverse Patch' while Table 3 reports 1.24s for PatchInstruct at Weather H=1. These discrepancies invalidate the quantitative efficiency argument as written.
[§3.2 and Appendix A] The prompt selection procedure appears to have used the evaluation data. Section 3.2 states that the prompts were 'designed through rigorous empirical testing,' and Appendix A (Table 5) selects 'Reverse Patches' as the best of five variants on the same Weather and Traffic horizons reported in Table 2. No train/validation/test split for prompt choice is described, so the reported accuracy may be the result of test-set selection. The authors should either adopt a held-out prompt selection protocol or present the results as exploratory.
[Tables 2 and 5] The numerical results for the same configuration are inconsistent across tables. For Weather H=1, Table 2 reports PatchInstruct MSE 0.0014, while Table 5 reports 0.0005 for Reverse Patches; for Traffic H=1, the values are 20.05 and 3.69, respectively. Unless the two tables use different test sets or normalization, this indicates an unstable evaluation pipeline and needs to be resolved before the accuracy claims can be assessed.
[§4.2, Table 3] The baseline setup is under-specified. The text says S2IP-LLM requires extensive fine-tuning, but also says all methods are evaluated 'in a consistent zero-shot setting without model retraining.' The reader cannot tell whether the S2IP-LLM numbers come from a fine-tuned model or a pretrained checkpoint without dataset-specific training, which changes the interpretation of the comparison substantially.
[Throughout] No error bars, significance tests, or number of evaluation samples are reported. Since LLM outputs are stochastic, single-run MSE/MAE values are not sufficient to support the claim of consistent improvement; the authors should report means and standard deviations over multiple runs and, ideally, a paired significance test.

minor comments (8)

[Abstract and §1] The abstract and introduction claim 'consistently outperform' but Table 2 contains counterexamples; consider qualifying the claim to the specific horizons and datasets where it holds.
[§4.3] The text says a single prediction consumes about 800–1000 input tokens, but Table 3 reports input token counts of 8500 (Weather) and 7950 (Traffic) for PatchInstruct.
[§5.1] The statement that all methods achieve over 80% MSE improvement on Weather is not supported by Table 2; for example, at H=4 PatchInstruct improves over S2IP-LLM by about 28%.
[§5.3] The text gives PatchInstruct Traffic MSE range as 6.47–20.05, omitting H=6 (25.59) and H=12 (235.75) from Table 2.
[§6] The conclusion ends with an incomplete sentence: 'This suggests that while prompt-based strategies offer a lightweight and effective alternative for time series forecasting.'
[Appendix A, Table 5] The timing columns in Table 5 contain inconsistent decimal precision (e.g., 1.61510, 1.2553, 1.2290); these should be formatted uniformly.
[Appendix A vs. main text] The name 'Reverse Patches' in Appendix A and 'PatchInstruct' in the main text should be explicitly identified as the same configuration to avoid confusion.
[Table 1] The single 'Value Range' for Weather (0.5–18.13) is implausible for all 14 meteorological variables unless the data are normalized; please specify the preprocessing.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: PatchInstruct's empirical claims are direct measurements against baselines; prompt selection on the same benchmarks is a generalization risk, not a by-construction reduction.

full rationale

PatchInstruct is presented as an empirical prompting framework, not as a result derived from equations or from prior work by the same authors. The central comparison in Tables 2-4 is a direct measurement of MSE/MAE against the S2IP-LLM baseline on the Weather and Traffic datasets; no fitted parameter is renamed as a prediction. The only selection step is the choice among prompt variants described in Section 3.2 ('The prompts were designed through rigorous empirical testing') and Appendix A (Table 5), where Reverse Patches is chosen after evaluating five strategies on the same benchmarks. This is a real risk of benchmark overfitting, and the paper itself concedes in Section 7 that the framework 'risks overfitting to specific tasks or datasets.' However, overfitting is not circularity: selecting the best of five prompt templates on the evaluation data does not make the reported PatchInstruct numbers equal to the selection criterion by construction, and the ranking against S2IP-LLM remains an empirical outcome. No load-bearing self-citation, imported uniqueness theorem, or ansatz-by-citation appears; the 'reverse patches' idea is presented as a design choice, not as a forced consequence of a prior result. Internal numeric inconsistencies (e.g., Weather H=1 MSE 0.0014 in Table 2 vs 0.0005 in Table 5) are reproducibility concerns, not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several hand-chosen design decisions and assumptions about LLM behavior. Most are reasonable for an empirical prompt-engineering paper, but the selection of the prompt variant on the evaluation data is the main weak point.

free parameters (4)

Patch window size = 3
Chosen by hand; the paper does not justify why 3 is optimal, and main experiments fix it at 3 without sensitivity analysis.
Patch stride = 1
Chosen by hand; overlapping patches with stride 1 are a design decision.
Number of neighbors = 5
Selected for the Neighs variant without a sensitivity analysis.
Prompt variant selection = Reverse Patches
Selected after comparing five prompt strategies in Appendix A on the same datasets, i.e., selected on the evaluation data.

assumptions (4)

domain assumption GPT-4/GPT-4o can reliably parse and forecast from string-formatted numeric patches.
The paper relies on the LLM's zero-shot numerical reasoning from prompted sequences without proving this capability.
domain assumption The Weather and Traffic datasets are representative enough to support generalizable conclusions.
The paper generalizes from only two datasets, which limits external validity.
ad hoc to paper Euclidean distance over past windows is a useful similarity measure for choosing neighbor series.
Introduced without justification as a general principle; no comparison to alternative similarity measures.
standard math Standard statistical background such as MSE and MAE definitions.
Implicitly assumed for evaluation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Forecasting Time Series with LLMs via Patch-Based Prompting and Decomposition." pith.science (2026). https://pith.science/paper/3EG7EJT3

@misc{pith2026250612953,
  author       = {Pith},
  title        = {Pith review of: Forecasting Time Series with LLMs via Patch-Based Prompting and Decomposition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3EG7EJT3}},
  note         = {Machine review of arXiv:2506.12953}
}

read the original abstract

Recent advances in Large Language Models (LLMs) have demonstrated new possibilities for accurate and efficient time series analysis, but prior work often required heavy fine-tuning and/or ignored inter-series correlations. In this work, we explore simple and flexible prompt-based strategies that enable LLMs to perform time series forecasting without extensive retraining or the use of a complex external architecture. Through the exploration of specialized prompting methods that leverage time series decomposition, patch-based tokenization, and similarity-based neighbor augmentation, we find that it is possible to enhance LLM forecasting quality while maintaining simplicity and requiring minimal preprocessing of data. To this end, we propose our own method, PatchInstruct, which enables LLMs to make precise and effective predictions.

Figures

Figures reproduced from arXiv: 2506.12953 by the authors.

**Figure 2.** PatchInstruct Forecasting Pipeline 3.1 Overview of Framework Our framework is designed to adapt large language models (LLMs) for time series forecasting without any fine-tuning, using carefully structured prompts that condition the model with temporal data and forecasting instructions. The pipeline is modular and supports multiple prompting strategies, including PatchInstruct, Neighbors, and PatchInstruct + Neighbo… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 11 canonical work pages

[1]

Decompose the series into trendt andresidual t =series t −trend t

work page
[2]

For each time-step create a pair token: (trendt ,residual t)

work page
[3]

Split the 96 composite tokens into overlap- ping patches (window = 3, stride = 1)

work page
[4]

series": 96 raw numbers (Humidity, 10-min cadence) -

Use those patches to forecast the next 3 raw values. Output exactly [[T1,R1], [T2,R2], [T3,R3]] [[T2,R2], [T3,R3], [T4,R4]] . . . [[T94,R94], [T95,R95], [T96,R96]] Prediction: [y1, y2, y3] No headings or extra words. Decimals ≤ 4 places; keep leading zeros (e.g.,0.8032). Reverse Ordered Patches PatchInstruct You are a forecasting assistant that sees time ...

work page
[7]

Split the series into overlapping patches (window = 3, stride = 1)

work page
[8]

Generate them in natural order, then reverse the list so the most recent patch appears first

work page
[9]

series": 96 raw numbers (Humidity, 10-min cadence) -

Use those patch tokens to forecast the next 3 normalised values. Output format: Patches: [[latest_patch], ... , [oldest_patch]] Prediction: [y1, y2, y3] No headings or extra words. Decimals ≤ 4 places; keep leading zeros (e.g.,0.8032). Meta tokens Patches PatchInstruct You are a forecasting assistant that sees time series data, where each datapoint is pai...

work page
[10]

Convert the 96-point series into 96 two- element tokens as above

work page

Show all 12 references

[11]

Split the token stream into overlapping patches (window = 3, stride = 1)
[12]

Output format: [(v1;slot1), (v2;slot2), (v3;slot3)] [(v2;slot2), (v3;slot3), (v4;slot4)]

Use those patches to forecast the next 3 raw demand values. Output format: [(v1;slot1), (v2;slot2), (v3;slot3)] [(v2;slot2), (v3;slot3), (v4;slot4)] . . . [(v94;slot94), (v95;slot95), (v96;slot96)] Prediction: [y1, y2, y3] No headings or extra words. Decimals ≤ 4 places; keep ...
[2023]

Xinyu Zhou, Zhengyuan Ding, Shuo Ren, Yutao Chen, Xinhui Huang, Jianhao Shi, and Wayne Xin Zhao

One fits all: Power general time series analysis by pretrained lm.Advances in neural information processing systems, 36:43322–43355. Xinyu Zhou, Zhengyuan Ding, Shuo Ren, Yutao Chen, Xinhui Huang, Jianhao Shi, and Wayne Xin Zhao
[2024]

recency-first

Ditto: A survey on fine-grained align- ments of large language models.arXiv preprint arXiv:2411.05793. A Summary of Forecasting Results Across different datasets Table 5 represents evaluating five prompting strate- gies—Basic, Non-Overlapping, STR Decompose, Reverse Patches, a...

2022 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Decompose the series into trendt andresidual t =series t −trend t

work page

[2] [2]

For each time-step create a pair token: (trendt ,residual t)

work page

[3] [3]

Split the 96 composite tokens into overlap- ping patches (window = 3, stride = 1)

work page

[4] [4]

series": 96 raw numbers (Humidity, 10-min cadence) -

Use those patches to forecast the next 3 raw values. Output exactly [[T1,R1], [T2,R2], [T3,R3]] [[T2,R2], [T3,R3], [T4,R4]] . . . [[T94,R94], [T95,R95], [T96,R96]] Prediction: [y1, y2, y3] No headings or extra words. Decimals ≤ 4 places; keep leading zeros (e.g.,0.8032). Reverse Ordered Patches PatchInstruct You are a forecasting assistant that sees time ...

work page

[5] [7]

Split the series into overlapping patches (window = 3, stride = 1)

work page

[6] [8]

Generate them in natural order, then reverse the list so the most recent patch appears first

work page

[7] [9]

series": 96 raw numbers (Humidity, 10-min cadence) -

Use those patch tokens to forecast the next 3 normalised values. Output format: Patches: [[latest_patch], ... , [oldest_patch]] Prediction: [y1, y2, y3] No headings or extra words. Decimals ≤ 4 places; keep leading zeros (e.g.,0.8032). Meta tokens Patches PatchInstruct You are a forecasting assistant that sees time series data, where each datapoint is pai...

work page

[8] [10]

Convert the 96-point series into 96 two- element tokens as above

work page