Pith. sign in

REVIEW 4 major objections 4 minor 32 references

Forecasting Commodity Price Shocks Using Temporal and Semantic Fusion of Prices Signals and Agentic Generative AI Extracted Economic News

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Fusing agentic-LLM news summaries with commodity price history predicts yearly price spikes at mean AUC 0.94 and accuracy 0.91.

desk verdict Useful application, but the headline 0.94 AUC is an artifact of look-ahead leakage in normalization, PCA, and retrospective news generation. read the letter →

arxiv 2508.06497 v1 pith:FTPL3TXZ submitted 2025-07-24 q-fin.CP cs.AIcs.LG

Mohammed-Khalil Ghali , Cecil Pang , Oscar Molina , Carlos Gershenson-Garcia , Daehan Won This is my paper
classification q-fin.CPcs.AIcs.LG
keywords commoditypriceforecastingspikedetectiondual-streamLSTMattentionmechanismnewsembeddingsagenticgenerativeAImultimodalfusiontime-seriesclassification
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's central claim is that yearly commodity price shocks can be forecast by fusing historical price series with semantic embeddings of global economic news produced by an agentic generative-AI pipeline. The model pairs a dual-stream LSTM with attention: one stream reads $z$-scored annual prices, the other reads PCA-compressed news embeddings, and the fused representation is trained to predict whether the next year's average price rises by more than $25\%$. On 64 years of data, the full model reports a mean AUC of 0.94 and an accuracy of 0.91, while logistic regression, random forest, and support vector machines score between 0.34 and 0.57, and removing the news stream drops AUC to 0.46. The authors' point is that unstructured text carries the market-moving context—geopolitics, policy shifts, supply disruptions—that price numbers alone do not, and that a fact-checked autonomous news pipeline can provide that context at a practical operating cost.

What carries the argument

The carrying mechanism is a dual-stream LSTM with scaled dot-product attention. One stream consumes a sliding window of normalized annual commodity prices; the other consumes $d'$-dimensional PCA-compressed embeddings of the year's fact-checked news summary. Attention weights the news-time-step vectors, the two streams are concatenated, and a dense layer with sigmoid outputs the spike probability for the next year. The news summaries themselves come from a manager-worker agent loop: a specialist drafts a summary for the year, a fact-checking agent returns a binary verdict against historical records, and the manager regenerates until the summary passes (with a retry cap of five). This pipeline turns 64 unstructured yearly texts into the semantic embeddings that the ablation results identify as the model's main source of skill.

What would settle it

Generate each year's news summary using only articles published before that year, retrain the identical model, and re-evaluate; if the mean AUC drops well below 0.94 toward the 0.46 no-news baseline, the retrospective fact-checking leaked the outcome it was meant to predict.

Watch

Extended reading notes

Core claim

The paper's central claim is that an agentic generative-AI news pipeline, when fused with price history through a dual-stream LSTM with attention, can detect next-year commodity price spikes across the 64 years 1960-2023 with mean AUC 0.94 and accuracy 0.91. The claim includes a strong attribution statement: the news stream is the decisive input, since removing it collapses AUC to 0.46, whereas removing attention or PCA leaves AUC at 0.83 and 0.76. The paper also reports that the full model maintains AUC above 0.88 for every individual commodity tested, while classical classifiers range widely across commodities, which is offered as evidence of generalization rather than commodity-specific tuning.

Load-bearing premise

The load-bearing premise is that each year's news summary contains only information available at that time; because all 64 summaries are generated retrospectively and fact-checked against the historical record (Section 4.1, Algorithm 1), a summary for a given year could encode knowledge of the very price movement it is trained to predict, and that leakage would make the reported AUC an invalid measure of real-time forecasting skill.

Editorial extensions

If this is right

  • An early-warning system built on yearly data can give import-dependent countries roughly one calendar year of lead time before a budget-affecting commodity spike.
  • The ablation results imply that unstructured news is a core input, not a refinement: dropping it cuts AUC from 0.94 to 0.46.
  • Removing attention keeps precision high at 0.94 but cuts recall to 0.53, so attention is what prevents missed spikes in this setting.
  • PCA compression does substantive work: removing it drops AUC to 0.76 and F1 by 0.47, indicating that raw high-dimensional news embeddings are too noisy to fuse directly.
  • Because the verified summaries are generated once and reused across commodities, the pipeline's ongoing operating cost is low enough for adoption in data-scarce economies.

Reading between the lines

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

  • If the reported AUC is genuine, the same fusion design should transfer to monthly or weekly data with a recalibrated spike threshold and a real-time news ingestion step; the yearly design cannot say how much lead time would survive that shift.
  • The large full-versus-no-news gap suggests a diagnostic use of the framework: for commodities with sparse news coverage the model should drift toward the no-news result, so the gap itself could quantify how much predictive value the text source adds per market.
  • Because the target is a year-over-year average-price jump, a correct forecast identifies the year of a spike, not the month; converting the warning into action still requires higher-frequency monitoring.
  • One could repurpose the fact-checking agent for counterfactual testing: suppress verification for a subset of years and compare AUCs to separate the value of factual checking from the value of raw LLM summaries.
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 / 4 minor

Summary. The paper proposes a hybrid framework for forecasting yearly commodity price spikes (defined as a >25% year-over-year increase) by fusing historical price series with embeddings of LLM-generated, fact-checked news summaries. A dual-stream LSTM with attention combines price and news modalities; evaluations report mean AUC 0.94, accuracy 0.91, and ablations showing news embeddings are crucial. Using 64 years of World Bank data and three baselines, the authors argue the method is suitable for early warning in developing economies.

Significance. If the reported results were leakage-free, the paper would demonstrate a useful application of agentic LLM pipelines to commodity forecasting. The paper provides detailed pseudocode (Algorithms 1 and 2), explicit equations, and a transparent architecture, which are strengths. However, the central evaluation is compromised by multiple forms of look-ahead leakage in preprocessing (z-score normalization), in dimensionality reduction (PCA on full data), and, most severely, in the retrospective construction of news summaries that can encode the exact outcome being predicted. The significance of the empirical claims therefore cannot be assessed from the current manuscript.

major comments (4)
  1. [Section 3, Eq. (2)] The z-score normalization in Eq. (2) uses the full-period mean μ_i and standard deviation σ_i for each commodity. Because the sample spans 1960–2023, the normalized price for any year t is a function of prices in years after t. The label for a sample at time t is y_{t+1} (the spike indicator for the next year, which depends on price at t+1); hence the test inputs contain information about the label through the normalization statistics. This look-ahead bias invalidates the reported AUCs as estimates of forecasting skill.
  2. [Algorithm 2, Step 1] Algorithm 2 computes WPCA on all T embeddings ({e_t}_{t=1}^T) and then applies this transformation to both training and test embeddings. This contradicts the text in Section 4.2, which states that 'WPCA contains the top d′ principal components extracted from the training set.' Fitting PCA on the full dataset, including test years, leaks future information into the test features and further compromises the evaluation.
  3. [Section 4.1, Algorithm 1] The news summaries are generated retrospectively by a manager–worker agent pipeline and accepted only after the fact-checking agent verifies them against 'historical records' (Eqs. (3)–(5) and Algorithm 1). For each year 1960–2023, the summary can therefore contain statements about actual price movements—for example, that a given year saw a major price surge—which is precisely the outcome encoded in the label (Eq. (1)). The news embeddings used as input features thus carry label information, so the 0.94 AUC and the large AUC drop when news embeddings are removed (Table 2) cannot be attributed to genuine semantic forecasting.
  4. [Section 5, Table 3 and Fig. 10] The baseline results are largely below random chance (logistic regression AUC 0.34, SVM 0.47), and the 'Without News Embeddings' ablation falls to AUC 0.46. Such below-chance performance is not a meaningful benchmark and is consistent with a corrupted evaluation protocol (e.g., small sample, class imbalance, or leakage). The paper does not state whether baselines were given the same (leaky) news features or only price data; without this detail and without a corrected protocol, the comparative claims in Section 5 are unsupported.
minor comments (4)
  1. [Section 5] The claim of a 'strict 20% hold-out test set, never seen by the model during training or validation' is misleading because the z-score normalization, PCA, and news-summary generation are all performed on the full dataset before the split; please clarify how the hold-out set avoids these preprocessing steps.
  2. [Section 4.2 / Algorithm 2] The notation for the target variable is inconsistent: Section 4.2 defines the target as y_{t+1}, while Algorithm 2 lists labels as {y_t}_{t=k+1}^T. Please make the temporal indexing consistent.
  3. [Section 6] The limitations paragraph acknowledges yearly granularity, small sample size, and summary-level omissions, but it does not discuss the look-ahead issues in normalization, PCA, or news generation; adding that discussion would be necessary for any revised evaluation.
  4. [Abstract] The phrase 'saving the very high costs of operating a full generative AI agents pipeline' is confusing: the proposed framework itself includes a full agentic pipeline. Please clarify what cost savings are being claimed.

Circularity Check

3 steps flagged · score 8.0 of 10

The 0.94 AUC is inflated by label leakage: news summaries are fact-checked against historical records describing the price movements that define the spike labels, and PCA/z-score preprocessing uses the full 1960–2023 sample including test years.

  1. self definitional [Section 4.1, Algorithm 1 (news extraction); Eq. (1) label definition]
    "The generated summary ŝ_i is then passed to the fact-checking agent A_fact, which returns a binary verdict v_i ∈ {0, 1} indicating whether the content is factually consistent with historical records. ... Each entry s_i contains the year, a list of referenced commodities, and a concise summary of major economic, geopolitical, and market-related developments impacting commodity prices in that year."

    The spike label s_i (Eq. 1) is defined by the year-over-year change in average price for that year. The news summary s_i is generated and fact-checked after the fact, with the fact-checker explicitly permitted to use 'historical records' that include the price movements determining the label. The embedding e_i = f_embed(s_i) is therefore constructed with access to the outcome being predicted. The paper's own ablation shows this news stream is the dominant signal (removing it drops AUC to 0.46), so the reported 0.94 AUC largely measures label information already embedded in the input rather than forecasting skill.

  2. other [Section 4.2, Algorithm 2, Phase I Step 1; Eq. (8)]
    "Step 1: Dimensionality reduction of news embeddings. Compute PCA transformation: W_PCA ← PCA({e_t}_{t=1}^T, d′) ... for t = 1 to T do: õ_t ← W_PCA^T e_t. The text of Eq. (8) says 'W_PCA contains the top d′ principal components extracted from the training set.'"

    Algorithm 2 fits PCA on all T embeddings before any train/test split, contradicting Eq. (8) and Section 5's claim that the validation 'avoids information leakage.' Test embeddings are projected using principal components estimated on the test set itself, so the held-out years' news features contain test-period statistical information. This is a look-ahead in feature construction that makes the test 'predictions' depend on the test distribution.

1 more flagged steps
  1. other [Section 3, Eq. (2)]
    "z_i,t = (x_i,t − μ_i)/σ_i ... where x_i,t denotes the raw price, μ_i is the historical mean, and σ_i is the standard deviation of commodity i over the full observation period."

    Each normalized price for year t uses μ_i and σ_i computed over the entire 1960–2023 sample, including the test years. The inputs presented to the model at a historical time t therefore contain statistics from future years, so the price stream is not causally available at forecast time. Section 5 states the evaluation 'avoids information leakage,' but Eq. (2) violates that for the structured input stream.

full rationale

The paper's central claim—that fusing fact-checked news embeddings with price series yields real-time spike forecasting with AUC 0.94—rests on an evaluation protocol in which the supposedly predictive inputs are constructed using outcome-period information. Algorithm 1's fact-checking agent verifies summaries against 'historical records' for the same year whose price change defines the label, and the LLM summarizer can draw on world knowledge of events such as the 1973 oil shock. Algorithm 2 further fits PCA on all T embeddings, including test years, and Eq. (2) normalizes prices with full-period moments. These are not mere methodological quibbles: the ablation shows that removing the news stream collapses AUC to 0.46, so the dominant signal is exactly the stream with the clearest label leakage. The below-chance baseline AUCs (logistic regression 0.34) are consistent with a corrupted test protocol. There is one self-citation (ref. [25]) but it is not load-bearing. The finding is therefore high circularity in the data-construction sense: the reported skill is not an independent out-of-sample forecast.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The model has no new physical entities. The load-bearing free parameters are the spike threshold and undisclosed architecture hyperparameters. The key assumptions are about data quality and, critically, the temporal validity of the retrospectively generated news summaries.

free parameters (5)
  • Spike threshold = 25%
    Defines the binary label in Eq. 1; results depend on this arbitrary threshold.
  • Window size k = not reported
    Sliding window length in Eqs. 9-10; no value given, so results cannot be reproduced.
  • PCA target dimension d' = not reported
    Dimensionality of reduced news embeddings in Eq. 8; chosen without justification.
  • LSTM hidden size h and attention dimension h_a = not reported
    Architecture capacities in Eqs. 11-14; not specified.
  • Training hyperparameters = not reported
    Learning rate, batch size, epochs, dropout, and regularization weight are listed only as symbols.
assumptions (4)
  • domain assumption World Bank yearly average prices for 1960-2023 are accurate and consistently measured.
    Section 3 relies entirely on this dataset; any revisions would change labels and features.
  • domain assumption The LLM-generated yearly news summaries are factually correct and temporally localized.
    Section 4.1 assumes the fact-checking agent prevents hallucinations and hindsight, but the summaries are generated retrospectively.
  • domain assumption A 25% year-over-year price increase is a meaningful spike for all commodities.
    Eq. 1 applies one threshold uniformly across heterogeneous commodities without economic justification.
  • ad hoc to paper PCA and normalization fitted on the full dataset do not leak future information.
    Algorithm 2 fits PCA on all data and Eq. 2 uses full-period statistics, violating this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Forecasting Commodity Price Shocks Using Temporal and Semantic Fusion of Prices Signals and Agentic Generative AI Extracted Economic News." pith.science (2026). https://pith.science/paper/FTPL3TXZ

@misc{pith2026250806497,
  author       = {Pith},
  title        = {Pith review of: Forecasting Commodity Price Shocks Using Temporal and Semantic Fusion of Prices Signals and Agentic Generative AI Extracted Economic News},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FTPL3TXZ}},
  note         = {Machine review of arXiv:2508.06497}
}
read the original abstract

Accurate forecasting of commodity price spikes is vital for countries with limited economic buffers, where sudden increases can strain national budgets, disrupt import-reliant sectors, and undermine food and energy security. This paper introduces a hybrid forecasting framework that combines historical commodity price data with semantic signals derived from global economic news, using an agentic generative AI pipeline. The architecture integrates dual-stream Long Short-Term Memory (LSTM) networks with attention mechanisms to fuse structured time-series inputs with semantically embedded, fact-checked news summaries collected from 1960 to 2023. The model is evaluated on a 64-year dataset comprising normalized commodity price series and temporally aligned news embeddings. Results show that the proposed approach achieves a mean AUC of 0.94 and an overall accuracy of 0.91 substantially outperforming traditional baselines such as logistic regression (AUC = 0.34), random forest (AUC = 0.57), and support vector machines (AUC = 0.47). Additional ablation studies reveal that the removal of attention or dimensionality reduction leads to moderate declines in performance, while eliminating the news component causes a steep drop in AUC to 0.46, underscoring the critical value of incorporating real-world context through unstructured text. These findings demonstrate that integrating agentic generative AI with deep learning can meaningfully improve early detection of commodity price shocks, offering a practical tool for economic planning and risk mitigation in volatile market environments while saving the very high costs of operating a full generative AI agents pipeline.

Figures

Figures reproduced from arXiv: 2508.06497 by the authors.

Figure 1
Figure 1. Normalized commodities prices over time with spikes marked The overall preprocessing workflow is illustrated in fig. 2, which summa￾rizes the data engineering pipeline used to prepare the model-ready dataset. The pipeline begins by extracting yearly average prices from the raw World Bank commodities table. These values are then used to compute yearly percentage changes and assign binary spike labels, which are store… view at source ↗
Figure 2
Figure 2. Data preprocessing pipeline with construction of average price time series, spike labels, and embedding-aligned features. The output of this process is a consolidated dataset consisting of three components: (i) a vector representation of news for each year, (ii) the corre￾sponding average commodity price, and (iii) the binary spike indicator. This structured input is used to train and evaluate the forecasting model … view at source ↗
Figure 3
Figure 3. Double phase commodities price spike forecast graphical abstract 4.1. Agentic Generative AI News Extraction To construct a temporally aligned dataset of economic news summaries relevant to commodity price movements, a hierarchical agent-based generative framework is employed. OpenAI Agents SDK is used to implement the needed agentic framework[32]. The architecture follows a manager–worker paradigm, where a high-leve… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: News extraction using an agentic generative AI approach Let Y = {y1, y2, . . . , yN } denote the set of years under analysis, with N = 64 in the current study. For each year yi ∈ Y, the orchestrator agent Amgr initiates the generation of a draft news summary sˆi by que…
Figure 5
Figure 5. Figure 5: Price forecast component 16 [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: News context component From this output, query, key, and value matrices are computed: Qt = Hnews t WQ, Kt = Hnews t WK, Vt = Hnews t WV (13) where WQ,WK,WV ∈ R h×ha are trainable parameters. Attention weights are computed as: At = softmax  QtK⊤ √ t ha  (14) The atten…
Figure 7
Figure 7. Figure 7: Price spikes forecast where σ(·) is the sigmoid activation function, and yˆt+1 represents the proba￾bility of a price spike. The model is trained using the binary cross-entropy loss: L = − 1 N X N t=1 [yt+1 log ˆyt+1 + (1 − yt+1) log(1 − yˆt+1)] (19) Optimization is pe…
Figure 8
Figure 8. Figure 8: Performance of the full model. Left: ROC curve showing strong classification ability (AUC = 0.9444). Right: Confusion matrix reflecting low false positives and false negatives on testing set. The ROC curve in fig. 8 (left) illustrates the classification capability of t…
Figure 9
Figure 9. Figure 9: Ablation results. Left: Fold-wise AUC for each model variant. Right: Boxplot showing distribution of AUC scores. To understand the relative contribution of each model component, ab￾lation experiments were performed. The left plot of fig. 9 shows the AUC scores across c…
Figure 10
Figure 10. Figure 10: AUC(left) and Accuracy(right) comparison of baseline models vs. proposed approach The bar charts in fig. 10 provide a comparative analysis of the proposed model against classical machine learning baselines including Logistic Regres￾sion, Random Forest, and SVM evaluat…
Figure 11
Figure 11. Figure 11: AUC comparison of forecasting models across a subset of individual commodities The individual commodity forecasting experiment is conducted to rigor￾ously evaluate the proposed model’s capacity to generalize across heteroge￾neous market structures by measuring its pre…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 26 canonical work pages

  1. [1]

    C. M. Wheeler, J. Baffes, A. N. Kabundi, G. Kindberg-Hanlon, P. S. O. Nagle, F. L. Ohnsorge, Adding fuel to the fire: cheap oil during the covid-19 pandemic, Tech. rep., The World Bank (2020)

    work page 2020

  2. [2]

    De Janvry, E

    A. De Janvry, E. Sadoulet, Agricultural growth and poverty reduction: Additional evidence, The World bank research observer 25 (1) (2010) 1–20

    work page 2010

  3. [3]

    P. C. Abbott, C. Hurt, W. E. Tyner, What’s driving food prices in 2011? (2011)

    work page 2011

  4. [4]

    Mamun, D

    A. Mamun, D. Laborde, Role of international price and domestic infla- tion in triggering export restrictions on food commodities, Agricultural Economics (2024)

    work page 2024

  5. [5]

    Ghoshray, M

    A. Ghoshray, M. Kejriwal, M. Wohar, Breaks, trends and unit roots in commodity prices: a robust investigation, Studies in Nonlinear Dynamics and Econometrics 18 (1) (2014) 23–40

    work page 2014

  6. [6]

    Zhang, Y

    Q. Zhang, Y. Hu, J. Jiao, S. Wang, Exploring the trend of commodity prices: A review and bibliometric analysis, Sustainability 14 (15) (2022) 9536

    work page 2022

  7. [7]

    S. Kaur, S. Aggarwal, S. Sarwar, Trade balance, monetary supply, com- modity prices, and greener energy growth: Contextual evidence from brics economies in the lens of sustainability, Environmental Science and Pollution Research 30 (29) (2023) 73928–73940

    work page 2023

  8. [8]

    N. B. Behmiri, J. R. P. Manso, Crude oil price forecasting techniques: A comprehensive review of literature, Alternative Investment Analyst Review 2013 (2) (2013) 29–48

    work page 2013

Show all 32 references
  1. [9]

    Lasheras, F

    F. Lasheras, F. Juez, A. S. Sánchez, A. Krzemień, Forecasting the comex copper spot price by means of neural networks and arima models, Resources Policy 45 (2015) 37–43

    2015

  2. [10]

    M. Kwas, M. Rubaszek, Forecasting commodity prices: Looking for a benchmark, Forecasting 3 (2) (2021) 447–459. 27

    2021

  3. [11]

    Baumeister, L

    C. Baumeister, L. Kilian, Forecasting the real price of oil in a chang- ing world: A forecast combination approach, Journal of Business and Economic Statistics 33 (3) (2015) 338–351

    2015

  4. [12]

    O. B. Sezer, M. U. Gudelek, A. M. Ozbayoglu, Financial time series forecasting with deep learning : A systematic literature review: 2005– 2019, Applied Soft Computing 90 (2020) 106181

    2020

  5. [13]

    Y. Zhao, J. Li, L. Yu, A deep learning ensemble approach for crude oil price forecasting, Energy Economics 66 (2017) 9–16

    2017

  6. [14]

    Chiroma, S

    H. Chiroma, S. Abdulkareem, T. Herawan, Evolutionary neural network model for west texas intermediate crude oil price prediction, Applied Energy 142 (2015) 266–273

    2015

  7. [15]

    P. Chen, Z. Boukouvalas, R. Corizzo, A deep fusion model for stock market prediction with news headlines and time series data, Neural Computing and Applications 36 (2024) 21229–21271

    2024

  8. [16]

    P. C. Tetlock, Giving content to investor sentiment: The role of media in the stock market, Journal of Finance 62 (3) (2007) 1139–1168

    2007

  9. [17]

    Y. Li, S. Jiang, X. Li, S. Wang, The role of news sentiment in oil futures returns and volatility forecasting: A data-decomposition based deep learning approach, Energy Economics 95 (2021) 105008

    2021

  10. [18]

    A. K. Banerjee, A. Sensoy, C. K. M. Lau, P. Pathak, Impact of media hype and fake news on commodity futures prices: A deep learning approach over the covid-19 period, Finance Research Letters 59 (2024) 104917

    2024

  11. [19]

    Bollen, H

    J. Bollen, H. Mao, X. Zeng, Twitter mood predicts the stock market, Journal of Computational Science 2 (1) (2011) 1–8. doi:https://doi.org/10.1016/j.jocs.2010.12.007. URL https://www.sciencedirect.com/science/article/pii/ S187775031100007X

    2011 doi

  12. [20]

    X. Ding, Y. Zhang, T. Liu, J. Duan, Deep learning for event-driven stock prediction, in: Proceedings of the 24th International Joint Conference on Artificial Intelligence (IJCAI), 2015, pp. 2327–2333. 28

    2015

  13. [21]

    Nassirtoussi, S

    A. Nassirtoussi, S. Aghabozorgi, T. Y. Wah, D. C. L. Ngo, Text min- ing for market prediction: A systematic review, Expert Systems with Applications 41 (16) (2014) 7653–7670

    2014

  14. [22]

    Z. Liu, D. Huang, K. Huang, Z. Li, J. Zhao, Finbert: A pre-trained financial language representation model for financial text mining, in: Proceedings of the 29th International Joint Conference on Artificial Intelligence (IJCAI), 2020, pp. 4513–4519

    2020

  15. [23]

    Lopez-Lira, Y

    A. Lopez-Lira, Y. Tang, Can chatgpt forecast stock price movements? return predictability and large language models, SSRN Electronic Jour- nalWorking Paper, available at SSRN 4392644 (2023)

    2023

  16. [24]

    Wu, et al., Bloomberggpt: A large language model for finance, arXiv preprint arXiv:2303.17564 (2023)

    S.-W. Wu, et al., Bloomberggpt: A large language model for finance, arXiv preprint arXiv:2303.17564 (2023)

    2023 arXiv

  17. [25]

    Ghali, A

    M.-K. Ghali, A. Farrag, S. Lam, D. Won, Beyondwords is all you need: Agentic generative ai based social media themes extractor (2025).arXiv: 2503.01880. URLhttps://arxiv.org/abs/2503.01880

    2025 arXiv

  18. [26]

    Zhang, Y

    W. Zhang, Y. Li, J. Wang, H. Chen, Finagent: A multimodal foundation agent for financial trading, arXiv preprint arXiv:2402.18485 (2024)

    2024 arXiv

  19. [27]

    Y. Li, Y. Yu, H. Li, Z. Chen, K. Khashanah, Tradinggpt: Multi-agent system with layered memory and distinct characters for enhanced finan- cial trading performance (2023).arXiv:2309.03736. URLhttps://arxiv.org/abs/2309.03736

    2023 arXiv

  20. [28]

    M. Wang, K. Izumi, H. Sakaji, Llmfactor: Extracting profitable factors through prompts for explainable stock movement prediction (2024). arXiv:2406.10811. URLhttps://arxiv.org/abs/2406.10811

    2024 arXiv

  21. [29]

    Ramyar, F

    S. Ramyar, F. Kianfar, Forecasting crude oil prices: A comparison between artificial neural networks and vector autoregressive models, Computational Economics 53 (3) (2019) 743–761

    2019

  22. [30]

    Jiang, K

    Y. Jiang, K. Ning, Z. Pan, et al., Multi-modal time series analysis: A tutorial and survey, arXiv preprint arXiv:2503.13709 (2024). 29

    2024 arXiv

  23. [31]

    URL https://www.worldbank.org/en/research/ commodity-markets

    World Bank, Commodity markets, accessed: 2025-05-29 (2024). URL https://www.worldbank.org/en/research/ commodity-markets

    2024

  24. [32]

    URLhttps://github.com/openai/openai-agents-python 30

    Openai, Openai/openai-agents-python: A lightweight, powerful frame- work for multi-agent workflows. URLhttps://github.com/openai/openai-agents-python 30

Pith tools

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