Pith. sign in

REVIEW 4 major objections 6 minor 51 references

NSW-EPNews: A News-Augmented Benchmark for Electricity Price Forecasting with LLMs

T0 review · 4 major / 6 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 This paper claims that current large language models cannot yet be trusted to forecast day-ahead electricity prices, even when they are fed news and weather data.

desk verdict The dataset is a genuine resource, but the paper's headline claim that LLMs never beat ARIMA is contradicted by its own Table 2, and the 'marginal news benefit' claim for classical models is untested. read the letter →

arxiv 2506.11050 v1 pith:Y3UK54WV submitted 2025-05-22 cs.LG

Zhaoge Bi , Linghan Huang , Haolin Jin , Qingwen Zeng , Huaming Chen This is my paper
classification cs.LG
keywords electricitypriceforecastingLLMbenchmarkingmultimodaltime-seriesnewsaugmentationhallucinationdetectionpromptengineeringNSWmarketday-aheadprediction
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 introduces NSW-EPNews, a benchmark that fuses ten years of half-hourly New South Wales spot electricity prices with daily temperatures and summarized market news, and asks whether adding text helps anyone predict the next day's 48 prices. Its central claim is that current LLM-based forecasters are not yet practically reliable: GPT-4o and Gemini 1.5 Pro give volatile accuracy across temporal splits and prompt styles, and they frequently emit hallucinated price sequences even when the prompts carry news, weather, and reasoning instructions. Classical tools remain competitive, with ARIMA keeping the lowest or near-lowest MAE on most splits while vectorized news gives linear regression and XGBoost only marginal gains. If the claim is right, operational forecasters should not yet trust prompt-based LLMs for day-ahead price prediction, and the field needs grounding and fine-tuning before text-augmented forecasting pays off.

What carries the argument

The carrying object is the multimodal record: one day's 48 half-hourly prices, the WattClarity news summaries for that day, and the day's temperature range, with the next day's 48 prices as target in a sliding-window design. The news side is compressed by GPT-4o with a four-block prompt (role, classification criteria, key attributes, summary rules) into a fixed schema with a relevance level. For LLMs, the benchmark builds 3.6k prompt-target pairs in four styles — zero-shot, few-shot, zero-shot with chain-of-thought, few-shot with chain-of-thought — plus a price-only ablation, and pairs accuracy metrics with a hallucination detector that flags echoing, trivial transformation, degenerate copying, and format violations. That detector is what turns raw MAE/MSE tables into the paper's reliability conclusion.

What would settle it

Replace the GPT-4o-generated news summaries with human-written or independently produced summaries and rerun the LLM evaluations on the same splits; if GPT-4o and Gemini then beat ARIMA with low hallucination rates, the paper's conclusion that current LLMs cannot deliver reliable news-aware forecasts is overturned.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that multimodal context and prompt engineering do not cure LLM brittleness in numerical forecasting. Across everything from a price-only ablation to few-shot chain-of-thought prompts, GPT-4o and Gemini 1.5 Pro produce forecasts whose accuracy varies widely with the data window and whose outputs are regularly caught in four surface failure modes: echoing the history verbatim, applying a constant offset, repeating one value, or failing the requested format. The paper concludes that regardless of occasional isolated wins on particular metrics and subsets, no LLM configuration delivers dependable news-aware forecasts, and that a simple ARIMA(1,1,1) fit to the price series alone stays competitive or superior over most of the comparison.

Load-bearing premise

The news signal in the benchmark is generated by GPT-4o, the same model family being scored, so any bias or invented detail in those summaries would contaminate both the measured value of news and GPT-4o's apparent performance.

Editorial extensions

If this is right

  • If the benchmark's conclusion is correct, adding news and temperature to LLM prompts does not by itself turn them into trustworthy day-ahead forecasters in volatile electricity markets.
  • Classical baselines such as ARIMA and linear regression remain defensible operational choices, and their relative standing should be reported in any future text-augmented forecasting study.
  • Hallucination rates need to accompany accuracy metrics in LLM forecasting evaluations, because discarding malformed outputs can make a model look better than it is.
  • Temporal splits matter: evaluation over the full ten-year horizon can hide the much larger errors that appear on the volatile recent 10 percent of data.
  • The released prompt-target pairs give later work a concrete testbed for retrieval-augmented generation, adaptive prompting, and fine-tuned open models.

Reading between the lines

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

  • An implicit consequence is that the four hallucination checks likely undercount the problem: a model producing plausible-looking but causally wrong numbers would pass all four rules, as the paper's own limitations section admits.
  • The data-generation loop is a testable confound: because GPT-4o summarized the news it later forecast with, rerunning the comparison with a different summarizer could reshuffle the ranking of models.
  • The benchmark could be extended to spike-day subsets to test whether LLM failures concentrate in extreme events, which the paper's 10 percent split results suggest.
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.

Referee Report

4 major / 6 minor

Summary. The paper introduces NSW-EPNews, a multimodal benchmark for day-ahead electricity price forecasting in New South Wales, combining 175k half-hourly spot prices (2015-2024), daily temperature readings, and GPT-4o-generated summaries of WattClarity news articles. It compares ARIMA, Linear Regression, and XGBoost with GPT-4o and Gemini 1.5 Pro under zero-shot, few-shot, and chain-of-thought prompt variants plus a news-free ablation, across full, 50%, 30%, and 10% splits. The authors report standard error metrics and a custom hallucination detection framework (echoing, trivial transformation, degenerate copying, format violation). Their headline finding is that news provides only marginal benefit to classical models, and that current LLMs, despite prompting, produce volatile forecasts and frequent hallucinations, making them unreliable relative to statistical baselines.

Significance. The paper addresses a timely and practical question: whether LLMs can incorporate unstructured news into electricity price forecasting. The released dataset and the explicit hallucination metric definitions are potentially useful resources for the community. If the negative result on LLM reliability were robust, it would be a valuable cautionary finding. However, the current manuscript does not yet support those conclusions, because the central comparative claim is contradicted by the paper's own Table 2, the news-ablation claim for classical models is untested, and the data in Table 2 show unexplained identical rows across models. These issues must be resolved before the benchmark can serve as a reliable testbed.

major comments (4)
  1. [Section 5 / Table 2] The Key Findings claim that "on every split the two state-of-the-art LLMs record higher MAE/RMSE than both classical baselines and an ablation that uses historical prices alone" is directly contradicted by Table 2. For example, on Recent 50%, Gemini 1.5 Pro Zeroshot+CoT achieves MAE 47.026 (lower than ARIMA's 54.579 and ChatGPT-4o Ablation's 52.713), and on Recent 10%, the same model has MAE 60.745 versus ARIMA's 89.092 and LR's 90.195. ChatGPT-4o Zeroshot also has MAE 53.646 on Recent 50%, below ARIMA's 54.579. Since the paper's central claim is that LLMs remain inferior to classical baselines, the discrepancy between the text and the table must be resolved by correcting either the data or the claim.
  2. [Abstract, Section 5, Section C.1] The claim that traditional models gain only "marginal benefit from news" is not tested. Section C.1 specifies that Linear Regression and XGBoost use a feature vector with the last ten prices, a 50-dimensional TF-IDF news representation, and temperature; there is no ablation that removes the news component from these models. ARIMA ignores news but is a different model class and cannot serve as the news-free control for LR/XGBoost. Without a news-free LR/XGBoost ablation, the paper cannot support the marginal-benefit conclusion.
  3. [Table 2] Several rows of Table 2 are numerically identical across different models, which implies a data-handling error. ChatGPT-4o Ablation and Gemini Ablation are identical on all three splits, ChatGPT-4o Fewshot and Gemini Fewshot share identical values on Recent 50%, and ChatGPT-4o Fewshot+CoT and Gemini Fewshot+CoT also share identical values on Recent 50%. These coincidences are implausible for independent model runs and must be explained or corrected before any comparative interpretation of the table is possible.
  4. [Section 3.1, Limitation (i)] The news modality is constructed entirely from GPT-4o summaries (Section 3.1, Figure 1), and GPT-4o is also one of the two evaluated forecasters. The paper's own limitation (i) acknowledges that "all news summaries are produced by GPT-4o," leaving the impact of alternative summarisers unknown. This creates a confound: if GPT-4o's price forecasts are systematically influenced by its own summarization style or by details it hallucinated during summarization, the measured effect of news on GPT-4o and any GPT-4o-versus-baseline comparison are not clean. The authors should either add a control using raw or alternatively summarized news, or substantially temper the claims about the value of news for LLMs.
minor comments (6)
  1. [Throughout] The manuscript contains numerous typos and grammatical errors, such as "compresive" (Abstract), "Statiscial" (Section 3.2), "cmoponents" (Section 3.2), "thte" (Appendix E), and "calcualte" (Appendix E); careful proofreading is needed.
  2. [Checklist] Several checklist answers appear to be copied from another paper, for example the theory-proof justification mentions graph initialization and the license question refers to the "MFIDDR dataset"; these should be replaced with answers specific to this paper.
  3. [Section 5] The name "NSW-EPN EWS" is a typo for "NSW-EPNews".
  4. [Table 2] The caption mentions red and blue highlighting, but the highlighting is not visible in the text; please clarify which entries are meant to be highlighted or remove the reference.
  5. [Appendix F] The statement that the prompt limits summaries to a length of at most 30,000 characters seems excessively large for a news summary; please verify the intended limit.
  6. [Section 3.3 / Appendix B] The threshold for degenerate copying ("repeating one number for over five times") may be confounded with the natural occurrence of repeated price values in the dataset; please justify the threshold.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical benchmark comparisons, not derivations from fitted inputs or self-citations.

full rationale

This paper is an empirical benchmark study rather than a derivation chain. The central claims—that classical models gain little from vectorized news and that LLM forecasters show volatile accuracy and hallucination-like failures—are supported by direct measurements (MAE, RMSE, MAPE, and hand-specified output checks) against ground-truth half-hourly prices. There is no equation in which an output is defined as an input, no fitted parameter renamed as a prediction, and no load-bearing self-citation: the authors' prior work is not used to justify any premise, and the benchmark design is described in the paper itself. The most plausible self-reference concern is that GPT-4o generates the news summaries and is also one of the evaluated forecasters; however, the paper explicitly discloses this in its Limitations ('all news summaries are produced by GPT-4o, so the impact of alternative LLM summarisers on data fidelity and downstream accuracy remains unknown'). That is a data-generation confound, not a circular reduction: the accuracy and hallucination metrics are computed from model outputs against external ground truth and could in principle have favored GPT-4o or any other model. In addition, the review-level observation that Table 2 contains rows numerically identical across different models and entries where LLMs beat ARIMA (e.g., Gemini 1.5 Pro Zeroshot+CoT MAE 47.026 vs ARIMA 54.579 on Recent 50%) contradicts the paper's summary text. This is a serious correctness and reproducibility issue, but it is not circularity. Accordingly, no circular step is exhibited and the circularity score is 0.

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

The benchmark rests on three hand-set evaluation parameters (hallucination thresholds, TF-IDF dimension, ARIMA order) and several domain assumptions about the news signal and the summarizer. No new physical entities are introduced. The most consequential free choices are the hallucination thresholds, since they determine the headline failure rates.

free parameters (3)
  • hallucination detection thresholds = echo >=10, offset match >=20, repeat frequency >5
    Hand-set thresholds in Algorithm 4 define the reported echoing, trivial transformation, and degenerate copying rates; changing them changes the headline hallucination statistics.
  • TF-IDF feature dimension = 50
    50-dimensional TF-IDF representation of news used as input to LR and XGBoost; chosen without reported tuning.
  • ARIMA order = (1,1,1)
    Fixed ARIMA(1,1,1) per day; a weak baseline that affects the headline comparison against LLMs.
assumptions (3)
  • domain assumption GPT-4o-generated news summaries faithfully represent the original WattClarity articles.
    The entire news modality is derived from GPT-4o summaries (Section 3.1 and Figure 1). If summaries are biased or hallucinated, all news-based results are affected.
  • domain assumption News posted on the historical day is a relevant signal for predicting the next day's prices.
    The task design feeds only the history day's news; contemporaneous news on the prediction day is not included, so the benchmark tests a specific, possibly narrow definition of news value.
  • domain assumption Median downsampling from 5-minute to 30-minute prices preserves the target forecasting quantity.
    Algorithm 2 aggregates each six 5-minute records with the median. This choice affects the ground truth and all subsequent comparisons, and its effect on forecast errors is not analyzed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NSW-EPNews: A News-Augmented Benchmark for Electricity Price Forecasting with LLMs." pith.science (2026). https://pith.science/paper/Y3UK54WV

@misc{pith2026250611050,
  author       = {Pith},
  title        = {Pith review of: NSW-EPNews: A News-Augmented Benchmark for Electricity Price Forecasting with LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y3UK54WV}},
  note         = {Machine review of arXiv:2506.11050}
}
read the original abstract

Electricity price forecasting is a critical component of modern energy-management systems, yet existing approaches heavily rely on numerical histories and ignore contemporaneous textual signals. We introduce NSW-EPNews, the first benchmark that jointly evaluates time-series models and large language models (LLMs) on real-world electricity-price prediction. The dataset includes over 175,000 half-hourly spot prices from New South Wales, Australia (2015-2024), daily temperature readings, and curated market-news summaries from WattClarity. We frame the task as 48-step-ahead forecasting, using multimodal input, including lagged prices, vectorized news and weather features for classical models, and prompt-engineered structured contexts for LLMs. Our datasets yields 3.6k multimodal prompt-output pairs for LLM evaluation using specific templates. Through compresive benchmark design, we identify that for traditional statistical and machine learning models, the benefits gain is marginal from news feature. For state-of-the-art LLMs, such as GPT-4o and Gemini 1.5 Pro, we observe modest performance increase while it also produce frequent hallucinations such as fabricated and malformed price sequences. NSW-EPNews provides a rigorous testbed for evaluating grounded numerical reasoning in multimodal settings, and highlights a critical gap between current LLM capabilities and the demands of high-stakes energy forecasting.

Figures

Figures reproduced from arXiv: 2506.11050 by the authors.

Figure 1
Figure 1. Prompt used for news classification Price data preprocess. Our benchmark includes ten years of Australia New South Whales’ elec￾tricity price data collected from the Australian National Electricity Market (NEM)[27]. The data frequency of electricity price data recording were thirty-minutes until 1st October 2021, NEM changed the frequency to five-minutes. To ensure consistency across the entire dataset, we applied a… view at source ↗
Figure 2
Figure 2. Example record in the raw dataset The raw dataset will eventually have more than 3600 pairs of records. It primarily serves two purposes: testing the performance of baseline forecasting mod￾els and constructing prompt–target pairs for LLM￾based evaluation. There is an example record shown in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Zero-shot 14 [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Few-shot [ { "prompt": "The historical load data is: 68.885, 61.675, 63.4, 63.19, ……(more prices),\nBased on the historical load data, please predict the load consumption in the next day,then output **exactly 48 numbers**, comma-separated.\nThe region for prediction is…
Figure 5
Figure 5. Figure 5: Zero-shot + CoT B Hallucination and error examples The hallucination and error outputed by LLMs in the experiment are defined by us as below: • Echoing Failure Although our prompts explicitly instruct the model to generate forecasts based on an analysis of historical e…
Figure 6
Figure 6. Figure 6: Few-shot + CoT tive number from the historical data, that generation is identified as a trivial transformation case. • Degenerate Copying In other instances, the model produces the same value repeatedly sometimes for the entire forecast horizon. Although the official e…
Figure 7
Figure 7. Figure 7: Echoing Failure Prompt: The historical load data is: 57.98, 57.98, 54.95, 48.91, 41.52, 49.78, 52.22, 46.205, 51.06, 42.26, 36.06, 35.88, 26.485, 4.82, 13.445, 40.395, 0.0, 36.06, 36.06, 44.915, 39.58, 37.2, 54.95, 54.95, 48.575, 28.235, 24.07, 37.125, 30.685, 36.305, …
Figure 8
Figure 8. Figure 8: Trival Transformation 17 [PITH_FULL_IMAGE:figures/full_fig_p017_8.png]
Figure 9
Figure 9. Figure 9: Degenerate Copy Prompt: The historical load data is: 57.98, 57.98, 54.95, 48.91, 41.52, 49.78, 52.22, 46.205, 51.06, 42.26, 36.06, 35.88, 26.485, 4.82, 13.445, 40.395, 0.0, 36.06, 36.06, 44.915, 39.58, 37.2, 54.95, 54.95, 48.575, 28.235, 24.07, 37.125, 30.685, 36.305, …
Figure 10
Figure 10. Figure 10: Format Violation 18 [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 37 canonical work pages

  1. [1]

    A probabilistic forecast methodology for volatile electricity prices in the Australian National Electricity Market

    C. Cornell, N. T. Dinh, and S. A. Pourmousavi. A probabilistic forecast methodology for volatile electricity prices in the australian national electricity market. arXiv preprint arXiv:2311.07289, 2023

    work page Pith review arXiv 2023

  2. [2]

    G. E. P. Box and G. M. Jenkins. Time Series Analysis: Forecasting and Control. Prentice Hall PTR, USA, 3rd edition, 1994

    work page 1994

  3. [3]

    S. M. Gonzales, H. Iftikhar, and J. L. López-Gonzales. Analysis and forecasting of electricity prices using an improved time series ensemble approach: an application to the peruvian electricity market. AIMS Mathematics, 9(8):21952–21971, 2024

    work page 2024

  4. [4]

    S. Ballı. Data analysis of covid-19 pandemic and short-term cumulative case forecasting using machine learning time series methods. Chaos, Solitons & Fractals, 142:110512, 2021

    work page 2021

  5. [5]

    A. A. A. Ahmed, A. Aljabouh, P. K. Donepudi, and M. S. Choi. Detecting fake news using machine learning: A systematic literature review. arXiv preprint arXiv:2102.04458, 2021

    arXiv 2021

  6. [6]

    Junqué de Fortuny, T

    E. Junqué de Fortuny, T. De Smedt, D. Martens, and W. Daelemans. Evaluating and understand- ing text-based stock price prediction models. Information Processing & Management, 50(2): 426–441, 2014

    work page 2014

  7. [7]

    C. Liu, Q. Xu, H. Miao, S. Yang, L. Zhang, C. Long, Z. Li, and R. Zhao. Timecma: Towards llm- empowered multivariate time series forecasting via cross-modality alignment. arXiv preprint arXiv:2406.01638, 2025

    arXiv 2025

  8. [8]

    Lazarczyk

    E. Lazarczyk. Market-specific news and its impact on forward premia on electricity markets. Energy Economics, 54:326–336, 2016

    work page 2016

Show all 51 references
  1. [9]

    Rogmann, J

    J. Rogmann, J. Beckmann, R. Gaschler, and H. Landmann. Media sentiment emotions and consumer energy prices. Energy Economics, 130:107278, 2024

    2024

  2. [10]

    K. Wei, Z. J. Zhang, and B. Lin. Does news propaganda really affect residents’ electricity rebound effect: New evidence of non-price information. Energy, 300:131589, 2024

    2024

  3. [11]

    A. K. Dubey, A. Kumar, V . García-Díaz, A. K. Sharma, and K. Kanhaiya. Study and analysis of sarima and lstm in forecasting time series data. Sustainable Energy Technologies and Assessments, 47:101474, 2021

    2021

  4. [12]

    Mosquera-López, J

    S. Mosquera-López, J. M. Uribe, and O. Joaqui-Barandica. Weather conditions, climate change, and the price of electricity. Energy Economics, 137:107789, 2024

    2024

  5. [13]

    Are transformers effective for time series forecasting?, 2022

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting?, 2022. URL https://arxiv.org/abs/2205.13504

    2022 arXiv

  6. [14]

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

    Jesus Lago, Grzegorz Marcjasz, Bart De Schutter, and Rafał Weron. Forecasting day-ahead electricity prices: A review of state-of-the-art algorithms, best practices and an open-access benchmark. Applied Energy, 293:116983, July 2021. ISSN 0306-2619. doi: 10.1016/j.apenergy. 202...

    2021

  7. [15]

    Energy price modelling: A comparative evaluation of four generations of forecasting methods, 2024

    Alexandru-Victor Andrei, Georg Velev, Filip-Mihai Toma, Daniel Traian Pele, and Stefan Lessmann. Energy price modelling: A comparative evaluation of four generations of forecasting methods, 2024. URL https://arxiv.org/abs/2411.03372

    2024 arXiv

  8. [16]

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

    Furong Jia, Kevin Wang, Yixiang Zheng, Defu Cao, and Yan Liu. Gpt4mts: prompt-based large language model for multimodal time-series forecasting. In Proceedings of the Thirty- Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applicatio...

    2024 doi

  9. [17]

    Position: What can large language models tell us about time series analysis, 2024

    Ming Jin, Yifan Zhang, Wei Chen, Kexin Zhang, Yuxuan Liang, Bin Yang, Jindong Wang, Shirui Pan, and Qingsong Wen. Position: What can large language models tell us about time series analysis, 2024. URL https://arxiv.org/abs/2402.02713

    2024 arXiv

  10. [18]

    S. Chen, T. C. Green, H. Gulen, and D. Zhou. What does chatgpt make of historical stock returns? extrapolation and miscalibration in llm stock return forecasts. arXiv preprint arXiv:2409.11540, 2024

    2024 arXiv

  11. [19]

    Timerag: Boosting llm time series forecasting via retrieval-augmented generation, 2024

    Silin Yang, Dong Wang, Haoqi Zheng, and Ruochun Jin. Timerag: Boosting llm time series forecasting via retrieval-augmented generation, 2024. URL https://arxiv.org/abs/2412. 16643

    2024

  12. [20]

    Ruibal and Mainak Mazumdar

    Claudio M. Ruibal and Mainak Mazumdar. Forecasting the mean and the variance of electricity prices in deregulated markets. IEEE Transactions on Power Systems, 23(1):25–32, 2008. doi: 10.1109/TPWRS.2007.913195

    2008

  13. [21]

    Using generative pre-trained transformers (gpt) for electricity price trend forecasting in the spanish market

    Alberto Menéndez Medina and José Antonio Heredia Álvaro. Using generative pre-trained transformers (gpt) for electricity price trend forecasting in the spanish market. Energies, 17 (10), 2024. ISSN 1996-1073. doi: 10.3390/en17102338. URL https://www.mdpi.com/ 1996-1073/17/10/2338

    2024 doi

  14. [22]

    Do, Yan Xu, and Pascale Fung

    Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, Quyet V . Do, Yan Xu, and Pascale Fung. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity, ...

    2023 arXiv

  15. [23]

    Retrieval-augmented generation for knowledge-intensive nlp tasks, 2021

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks, 2021. URL https://arx...

    2021 arXiv

  16. [24]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff...

    2020 arXiv

  17. [26]

    Wattclarity: Commentary and analysis of australia’s national electricity market

    Global-Roam Pty Ltd. Wattclarity: Commentary and analysis of australia’s national electricity market. https://wattclarity.com.au/, 2025. Accessed: May 2025

    2025

  18. [27]

    National electricity market (nem)

    Australian Energy Market Operator. National electricity market (nem). https://aemo.com. au/energy-systems/electricity/national-electricity-market-nem , 2025. Ac- cessed: May 2025

    2025

  19. [28]

    Z. Ji, N. Lee, R. Frieske, T. Yu, D. Su, Y . Xu, E. Ishii, Y . J. Bang, A. Madotto, and P. Fung. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):248, 2023

    2023

  20. [29]

    B. Meskó. Prompt engineering as an important emerging skill for medical professionals: Tutorial. Journal of Medical Internet Research, 25:e50638, 2023

    2023

  21. [30]

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, and D. Zhou. Chain-of-thought prompting elicits reasoning in large language models. arXiv preprint arXiv:2201.11903, 2023

    2023 arXiv

  22. [31]

    H. Qu, Q. Duan, and M. Niu. Modeling the volatility of realized volatility to improve volatility forecasts in electricity markets. Energy Economics, 74:767–776, 2018. 12

    2018

  23. [32]

    Bégin, F

    J.-F. Bégin, F. Gómez, K. Ignatieva, and H. Li. The stochastic behavior of electricity prices under scrutiny: Evidence from spot and futures markets. Energy Economics, 144:108296, 2025

    2025

  24. [33]

    Gpt-4o technical report

    OpenAI. Gpt-4o technical report. https://openai.com/index/gpt-4o, 2024. Accessed: May 2025

    2024

  25. [34]

    M. Reid, N. Savinov, D. Teplyashin, D. Lepikhin, T. Lillicrap, J.-B. Alayrac, R. Soricut, A. Lazaridou, O. Firat, J. Schrittwieser, I. Antonoglou, R. Anil, S. Borgeaud, A. Dai, K. Millican, E. Dyer, M. Glaese, T. Sottiaux, B. Lee, F. Viola, M. Reynolds, Y . Xu, et al. Gemini 1...

    2024 arXiv

  26. [35]

    A. Kent. Nsw prices spike early ahead of forecasts on afternoon of 27th november. https://wattclarity.com.au/articles/2024/11/ nsw-prices-spike-early-ahead-of-forecasts-on-afternoon-of-27th-november/ ,

    2024

  27. [36]

    prompt":

    P. McArdle. A much shorter run of evening volatility in qld and nsw on friday 8th november 2024. https://wattclarity.com.au/articles/2024/11/ 08nov-volatility-qldandnsw/, 2024. Accessed: May 2025. 13 A Prompt settings Zero-shot. The zero-shot style prompt includes key contextu...

    2024

  28. [37]

    Guidelines: • The answer NA means that the abstract and introduction do not include the claims made in the paper

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The abstract has comprehensively reflected the paper’s contributions and scope. Guidelines: • The answer NA means that...

  29. [38]

    Limitations

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: Please refer to the 5 for the limitations of our paper. 23 Guidelines: • The answer NA means that the paper has no limitation while the answer No mea...

  30. [39]

    Guidelines: • The answer NA means that the paper does not include theoretical results

    Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [Yes] Justification: To prove the frequency-domain structure of the learnable DCT anchors for graph initializatio...

  31. [40]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and da...

  32. [41]

    Guidelines: • The answer NA means that paper does not include experiments requiring code

    Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The source code and dataset...

  33. [42]

    • The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them

    Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: All relevant information are ...

  34. [43]

    Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] Justification: Guidelines: • The answer NA means that the paper d...

  35. [44]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [No] Justification: Did not include the comp...

  36. [45]

    • If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics

    Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: NA Guidelines: • The answer NA means that the authors have not reviewed the Neu...

  37. [46]

    Guidelines: • The answer NA means that there is no societal impact of the work performed

    Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [NA] Justification: Positive impacts include improving DR detection accuracy and supporting clinical workflows. Guidelines: • ...

  38. [47]

    Guidelines: 27 • The answer NA means that the paper poses no such risks

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: Our paper poses...

  39. [48]

    Guidelines: • The answer NA means that the paper does not use existing assets

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: The original pap...

  40. [49]

    The newly generated dataset is provided via an anonymized URL in Appendix ??

    New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: The code for the experiments are provided in the supplementary materials. The newly generated dataset is provided vi...

  41. [50]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? A...

  42. [51]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

  43. [52]

    Answer: [NA] Justification: The core method development in this paper does not involve LLMs as any important, original, or non-standard components

    Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the ...

    2025

Pith tools

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