Pith. sign in

REVIEW 3 major objections 4 minor 33 references

F$^2$Agent: Financial Fusion of Agentic Intelligence for Multimodal Trading

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Fused specialist AI agents out-trade 16 baselines in backtest

desk verdict A coherent agentic fusion design whose headline results are contradicted by its own appendix. read the letter →

arxiv 2608.05668 v1 pith:ALUWFG6S submitted 2026-08-06 cs.MA cs.AIcs.MM

Changshuo Liu , Yanzheng Jin , Shangfeng Cai , Peng Fang , Xiaokui Xiao , Beng Chin Ooi This is my paper · ORCID
classification cs.MAcs.AIcs.MM
keywords multimodaltradingLLMagentsmodalityfusionconsistencyregularizationfinancialAIagenticbacktestingspecialized
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

F2Agent argues that LLM-based trading agents fail not for lack of data but for poor fusion: bolting news text onto price time series via prompt concatenation drowns quantitative cues and leaves decisions vulnerable to market noise. The paper proposes a hierarchy of four specialized agents (market, technical, news, sentiment) whose latent representations are combined by a modality-aware attention mechanism with learned per-modality priors, plus a regularization term that punishes the model for leaning harder on a corrupted modality. On six assets against 16 baselines, the system reports the best annualized return on every asset, the best Sharpe ratio on five, and over 20% average relative improvement in annualized return. A sympathetic reader would take the paper's core claim to be that adaptive, noise-robust fusion, rather than larger general-purpose models, is the key to making multimodal trading agents add value.

What carries the argument

The load-bearing mechanism is the modality-aware adaptive fusion layer coupled with consistency regularization. Four agents produce modality embeddings; each is projected into a shared $d$-dimensional space via Eq. 5. A joint attention block computes modality weights, with learned modality vectors $r_m$ injected into keys and values to add an instance-independent prior, regularized for diversity by $\mathcal{L}_{\text{modality}}$ (Eq. 7). Robustness comes from $\mathcal{L}_{\text{robust}}$ (Eq. 8), which adds Gaussian noise to a randomly chosen modality embedding, enforces logit-level consistency between clean and perturbed samples, and penalizes the model if the attention weight on the perturbed modality rises. The overall objective (Eq. 10) combines cross-entropy with these two regularizers. This design is what distinguishes F2Agent from prompt-concatenation agent baselines and is claimed to explain why performance improves as more modalities are added, whereas baselines degrade.

What would settle it

Inspect the timestamp distribution of the fine-tuning set $\mathcal{D}_{\text{NA}}$ for the news analysis agent: if any simulated application trajectory or prompt example is built from news dated on or after 2025-04-01, rerun the full pipeline with the news agent frozen at the 2025-03-31 validation cut-off and check whether the reported annualized return and Sharpe advantages over FinAgent and other baselines persist.

Watch

Extended reading notes

Core claim

The paper's central claim is that a hierarchy of modality-specialized agents combined with a modality-aware adaptive fusion mechanism and noise-robust consistency regularization yields consistently superior trading performance across diverse assets. Concretely, F2Agent reports the best annualized return on all six test assets (AAPL, AMZN, GOOG, MSFT, TSLA, BTCUSD), the highest Sharpe ratio on five of them, and an average relative improvement of over 20% in annualized return over the best baseline per asset. The design replaces prompt-level concatenation, which the paper argues lets textual bias dominate numerical cues, with learned fused representations: each agent's hidden state is projected to a shared space, attention weights are computed across modalities with a learned per-modality prior injected into keys and values, and a robustness loss aligns logits between clean and perturbed inputs while suppressing attention to the perturbed modality.

Load-bearing premise

The reported backtest returns are only genuine out-of-sample results if no test-period information reaches the model through the fine-tuned news agent or the news retrieval pipeline.

Editorial extensions

If this is right

  • If the claim holds, LLM trading agents should be built with explicit fusion layers rather than prompt concatenation; the paper's ablations show concatenation fusion yields far lower annualized return and Sharpe ratio on both AAPL and BTCUSD.
  • Richer multimodal inputs help F2Agent but degrade general and financial LLM baselines such as Llama4-Scout-17B and FinAgent on AAPL and BTCUSD, indicating the fusion mechanism itself makes additional modalities usable.
  • The noise-robust consistency regularizer contributes a large share of the gain: removing it drops AAPL annualized return from 50.08% to 34.77% in Table 2, so robustness training is not incidental.
  • The system reports lower daily token consumption than TradingAgents and FinAgent (Figure 5), suggesting the design is deployable without prohibitive LLM costs.
  • Extended-horizon results through 2026-03-20 claim the advantage persists, with best annualized return and Sharpe on AAPL, GOOG, TSLA, and BTCUSD, so the effect is not confined to the original test window.
  • The fusion layer's attention scores are shown to shift across modalities over time, supporting the claim that the model adapts its reliance on market, technical, news, and sentiment evidence.
  • Full model variations consistently beat ablations without the modality prior or the robustness regularizer, indicating the components complement each other rather than one carrying all the benefit.

Reading between the lines

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

  • The paper's own appendix raises a decisive check: the news analysis agent is fine-tuned on 'simulated application trajectories designed to mimic real usage of F2Agent' generated with GPT-4o-mini, and the appendix's logged prompt examples are dated 2025-11-03, after the stated 2025-03-31 validation cut-off. If that simulated data draws on post-cutoff news, the reported out-of-sample returns would b
  • The same fusion-plus-consistency recipe could transfer to other noisy multimodal forecasting settings such as demand forecasting with social signals or macroeconomic nowcasting, since the consistency loss against a perturbed modality is modality-agnostic.
  • A natural extension would test F2Agent on 'sell-the-news' events by measuring whether it avoids sentiment traps on dates where news sentiment and price momentum disagree, using the attention scores the paper visualizes as a diagnostic.
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

3 major / 4 minor

Summary. The paper proposes F2Agent, a multimodal agentic trading system that combines a hierarchy of specialized agents (market, technical, news, and sentiment) with a modality-aware adaptive fusion mechanism and a noise-robust consistency regularizer. The main claims are that F2Agent consistently outperforms 16 competitive baselines across six assets, achieving the best annualized return on all six assets and the highest Sharpe ratio on five, with an average improvement in annualized return exceeding 20% relative to the best baseline. The evaluation in Table 1 and its appendices cover five stocks and BTCUSD, with additional experiments on a broader asset pool and an extended test horizon, plus random-seed robustness and statistical significance analyses in Appendix G.

Significance. If validated, the architecture-level contributions—specialized agent hierarchy, adaptive cross-modal attention with per-modality priors, and consistency regularization against noisy modalities—would be a meaningful advance for LLM-based multimodal trading systems, going beyond prompt-level concatenation. The paper is commendably detailed in places: it provides full prompt templates, hyperparameters, backtesting assumptions, and raw seed-level results in the appendices, which aid reproducibility. However, the headline claim of consistent state-of-the-art performance is directly contradicted by the paper's own appendix, and the out-of-sample integrity of the fine-tuning pipeline is not established. As it stands, the paper's central claim is not supported by its own evidence.

major comments (3)
  1. [Section 3.2, Appendix D, and Appendix F.1] The claim that F2Agent obtains the best ARR on all six assets and consistently outperforms 16 competitive baselines is contradicted by Appendix F.5. Table 10 lists an unexplained method, PMRL_Finance, with CR/ARR/SR values exceeding F2Agent on AAPL (51.28/102.56/2.77 vs. 25.04/50.08/1.22), AMZN (27.77/55.54/1.73 vs. 20.44/40.87/1.18), and GOOG (64.07/127.80/2.39 vs. 60.24/120.48/2.55). Tables 11 and 12 further show PMRL_Finance with higher ACC and MCC than F2Agent on all three assets. PMRL_Finance is never described in the methodology or in the baseline list of Section F.4, and the Improvement rows in Table 10 are unchanged despite its presence. If PMRL_Finance is a legitimate comparison method, the headline results are misreported; if it is a labeling error, the appendix is internally inconsistent and must be corrected.
  2. [Section 3.2, Appendix D, and Appendix F.1] The out-of-sample integrity of the reported results is not established. The supervised trading model uses a chronological split (Appendix F.1), but the fine-tuning data D_NA for the news analysis agent in Eq. (4) is described only as containing 'simulated application trajectories' generated with GPT-4o-mini, with no temporal cutoff specified. Appendix D's logged examples in Tables 6 and 7 carry a reference date of 2025-11-03, which falls after the main test period (2025-04-01 to 2025-09-30) and inside the extended horizon used in Tables 14-15. If such post-cutoff cases are included in D_NA, the fine-tuned news agent has seen test-window information, contaminating the reported backtest returns. The authors must state the exact date range of D_NA and verify that no test-period data were used in fine-tuning.
  3. [Section 5.1, Table 1, and Appendix G] The main results in Table 1 are reported for a single random seed per method, and the multi-seed analysis in Appendix G covers only F2Agent, Transformer, and Concat. Fusion on AAPL and BTCUSD. There are no multi-seed results for the 16 baselines or for the six main assets, so the central claim of consistent outperformance is not accompanied by any variance or significance assessment over the full comparison set. The seed experiments are useful, but they do not substantiate the headline improvements over FinAgent, DeepFund, or the other baselines.
minor comments (4)
  1. [Appendix F.5, Tables 9 and 10] Tables 9 and 10 appear to be near-duplicate performance tables, with the only visible difference being the presence of PMRL_Finance in Table 10; please remove the duplicate or explain why both are needed.
  2. [Appendix F.6, Tables 14-15] The extended test horizon (2025-04-01 to 2026-03-20) is used in Tables 14 and 15, but the main test period reported in Appendix F.1 ends on 2025-09-30; the paper should clearly state which horizon corresponds to the headline numbers in Table 1.
  3. [Table 1, Improvement row] The Improvement row contains dashes for SR and MDD on several assets; please specify how the relative improvement is computed and why certain entries are omitted.
  4. [Appendix A and Section 5.2] There are minor typographical and grammatical issues, such as 'heuritics' (should be 'heuristics') and 'the most closest competitor' in Section 5.2.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the fusion head is trained on a chronological split and tested out-of-sample; PMRL_Finance and LLM fine-tuning date issues are correctness risks, not circular steps.

full rationale

F2Agent's derivation chain is not circular. The market/technical encoders and the fusion head are trained on a fixed chronological split (training 2023-10-01–2024-09-30, validation 2024-10-01–2025-03-31, testing 2025-04-01–2025-09-30; Appendix F.1) and evaluated with a unified backtesting protocol that executes signals on day t+1, so the reported ARR/SR/MDD are out-of-sample quantities. The prediction label (Eq. 2) is defined directly from adjusted closing prices and used consistently in training and evaluation; HOLD is explicitly an execution-state mapping (Appendix E.3, Table 8), not a third class smuggled into the supervised target. The loss (Eq. 10) is a standard cross-entropy plus two regularizers; no test-period performance metric is used as a fitting target. Citations to FinMEM, FinAgent, TradingAgents, and DeepFund are external baselines, and none of the paper's load-bearing methodological choices (modality attention, Eqs. 6-7; robustness loss, Eq. 8) is imported from the authors' own prior work, so there is no self-citation load-bearing chain. Two manuscript-level concerns remain, but they are correctness/leakage issues rather than circularity: (i) Appendix F.5 Tables 10-12 introduce an unexplained method 'PMRL_Finance' that beats F2Agent on AAPL, AMZN, and GOOG in ARR/ACC/MCC, contradicting the headline 'best ARR on all six assets' and the baseline count of 16; (ii) the fine-tuning set D_NA for the news agent (Section 3.2, Eq. 4) is not given an explicit temporal cutoff, and Appendix D logs prompts dated 2025-11-03, which is inside the extended test horizon (2025-04-01 to 2026-03-20, Table 14) even though it is after the main test window. Neither issue makes a prediction equivalent to its input by construction; they affect validity and reproducibility, not circularity. Accordingly, the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central performance claim rests on a standard supervised-learning pipeline with several hand-chosen hyperparameters (loss weights, threshold, transaction cost) whose values are not all disclosed, plus implicit domain assumptions about predictability, temporal cleanliness, and backtest fidelity. No invented entities are introduced.

free parameters (6)
  • lambda_mod (modality regularizer weight) = not reported
    Coefficient of L_modality in Eq. 10, controlling diversity of modality prior vectors; tuned on validation but value not stated.
  • lambda_rob (robust regularizer weight) = not reported
    Coefficient of L_robust in Eq. 10, controlling logit-consistency and attention-suppression strength; tuned on validation, value not stated.
  • gamma (attention suppression weight) = not reported
    Weight inside L_robust (Eq. 8) for the max(0, alpha'_m* - alpha_m*) term; value not reported.
  • lambda (auxiliary classification weight for LLM fine-tuning) = not reported
    Weight of the auxiliary UP/DOWN classification loss in Eq. 4; value not reported.
  • News summarizer rating threshold = 5
    Appendix E.1: articles with rating R_i >= 5 retained, top 3 selected; ad hoc filtering choice affecting textual inputs.
  • transaction cost = 0.003
    Uniform transaction cost applied to all methods in backtesting (Appendix E.3); a modeling choice, not fitted, but it affects absolute returns.
assumptions (4)
  • domain assumption Next-day price direction is predictable from the four modalities (OHLCV, technical indicators, news summaries, sentiment) within a 30-day lookback.
    The entire supervised learning setup (Eq. 2, Section 3.2) presupposes this predictability; if prices are efficient to these features, the central claim collapses.
  • domain assumption The chronological train/validation/test split is not violated by the LLM fine-tuning data or news retrieval (no temporal leakage).
    Section 3.2 fine-tunes the news agent on GPT-4o-mini-generated 'simulated application trajectories' without a stated cutoff; Appendix F.1 asserts the split only for the supervised trading model. This is the weakest assumption reviewed.
  • domain assumption The backtesting protocol (next-day execution, long-or-flat, 0.3% transaction cost, no slippage/market impact) is a faithful proxy for real trading performance.
    Appendix E.3 explicitly states slippage and market impact are not modeled; the headline returns (GOOG 120.48% ARR) are produced under this protocol.
  • domain assumption Modalities are available and informative for all six assets on all trading days; missing modalities are handled by the robustness regularizer at training only.
    The fusion attention in Eq. 6-7 assumes all M embeddings present; the paper does not specify a test-time mask for missing modalities, though the robustness study varies modality subsets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of F$^2$Agent: Financial Fusion of Agentic Intelligence for Multimodal Trading." pith.science (2026). https://pith.science/paper/ALUWFG6S

@misc{pith2026260805668,
  author       = {Pith},
  title        = {Pith review of: F$^2$Agent: Financial Fusion of Agentic Intelligence for Multimodal Trading},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ALUWFG6S}},
  note         = {Machine review of arXiv:2608.05668}
}
abstract

With increasingly diverse and heterogeneous information sources, effectively leveraging multimodal data is becoming pivotal for high-quality financial trading. Although recent advancements in Large Language Model (LLM)-based agents have enabled the ingestion of multimodal inputs, existing methods fail to capture nuanced cross-modal dependencies and remain vulnerable to market noise, due to limited multimodal modeling, ineffective fusion mechanisms, and inadequate robustness. To address these challenges, we propose F$^2$Agent, a novel multimodal agentic paradigm driven by the Financial Fusion of Agentic Intelligence. F$^2$Agent first deploys a hierarchy of specialized agents to comprehensively extract modality-specific signals. It further introduces a modality-aware adaptive fusion mechanism coupled with noise-robust consistency regularization to dynamically capture fine-grained inter-modality dependencies and generate noise-resilient trading signals. Extensive experiments on six stocks and cryptocurrency assets demonstrate that F$^2$Agent consistently outperforms 16 competitive baselines across multiple trading metrics, with over 20% relative improvement in annualized return on average. Notably, F$^2$Agent delivers returns of 120.48% on GOOG and 148.41% on TSLA, demonstrating its efficacy and robustness in varying market dynamics.

Figures

Figures reproduced from arXiv: 2608.05668 by the authors.

Figure 1
Figure 1. The overall architecture of F2Agent, which orchestrates a hierarchy of specialized agents to extract latent features from heterogeneous multimodal inputs and synthesizes these representations via a Multi-agent Fusion module to generate robust trading signals. is characterized by a temporal sequence of T days, Xm = {x m 1 , xm 2 , . . . , xm T }. The multimodal input can be compactly represented as X = {Xm}m∈M. Each … view at source ↗
Figure 2
Figure 2. Modality-aware adaptive fu￾sion mechanism. over M modalities, each modality sequence X(m) is pro￾cessed by a modality-specific agent fm to produce a se￾quence representation Hm(x) = fm(X(m) ) ∈ R Tm×d˜m, where Tm denotes the sequence length and ˜dm is the hid￾den dimension of agent m. We then extract a fixed-length summary hm(x) ∈ R d˜m. For LLM-based agents, we use last-token pooling BehnamGhader et al. (2024) to s… view at source ↗
Figure 3
Figure 3. System workflow of F2Agent. News Summarizer. To reduce noise and control token budgets, the news summarizer serves as a preprocessing filter. It retrieves ticker-relevant articles via semantic vector search, ranks them by estimated market im￾pact using DeepSeek-R1, and compresses the most relevant items into concise daily sum￾maries for downstream reasoning. Modality Processing Module. F 2Agent en￾codes four modalit… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Cumulative return comparison of F2Agent and competing benchmarks on AAPL, AMZN, and BTCUSD. Results for the remaining three assets are provided in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Token consumption analysis. for signal generation across varying assets. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: The visualization provides empirical support for [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Case study of F2Agent vs. DeepFund when performing single asset trading. DeepFund is the strongest baseline on AAPL among the compared alternatives. 6 CONCLUSION F 2Agent is a multimodal agentic paradigm for financial trading that improves cross-modal under￾standing an…
Figure 8
Figure 8. Figure 8: Case studies of our stock prediction system, we only display partial prompt for brevity. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Illustration of News Summarizer. The News Summarizer, depicted in [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Performance comparison on cumulative return over time between F [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 11
Figure 11. Figure 11: Attention scores of our proposed method. [PITH_FULL_IMAGE:figures/full_fig_p031_11.png]
Figure 12
Figure 12. Figure 12: Attention scores of our proposed method (after standardized). [PITH_FULL_IMAGE:figures/full_fig_p032_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 10 canonical work pages

  1. [1]

    Llm2vec: Large language models are secretly powerful text encoders.arXiv preprint arXiv:2404.05961,

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. Llm2vec: Large language models are secretly powerful text encoders.arXiv preprint arXiv:2404.05961,

    arXiv

  2. [4]

    C DETAILS OFPROBLEMFORMULATION In addition to numerical modalities, we incorporate textual news

    for the definitions of all symbols used. C DETAILS OFPROBLEMFORMULATION In addition to numerical modalities, we incorporate textual news. For stocks at date t, the summarized news set is N s t ={n s t,1, . . . , ns t,|N s t |}, with |N s t | the number of daily news items. The model outputs a binary prediction ˆys t ∈ {0,1}together with a human-readable e...

    work page 2001

  3. [5]

    Gpt-4o system card.arXiv preprint arXiv:2410.21276,

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv preprint arXiv:2410.21276,

    arXiv

  4. [8]

    Time travel is cheating: Going live with deepfund for real-time fund investment benchmarking.arXiv preprint arXiv:2505.11065,

    Changlun Li, Yao Shi, Chen Wang, Qiqi Duan, Runke Ruan, Weijie Huang, Haonan Long, Lijun Huang, Nan Tang, and Yuyu Luo. Time travel is cheating: Going live with deepfund for real-time fund investment benchmarking.arXiv preprint arXiv:2505.11065,

    arXiv

  5. [12]

    Can chatgpt forecast stock price movements? return predictability and large language models.arXiv preprint arXiv:2304.07619,

    Alejandro Lopez-Lira and Yuehua Tang. Can chatgpt forecast stock price movements? return predictability and large language models.arXiv preprint arXiv:2304.07619,

    arXiv

  6. [13]

    Playing atari with deep reinforcement learning.arXiv preprint arXiv:1312.5602,

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, Alex Graves, Ioannis Antonoglou, Daan Wierstra, and Martin Riedmiller. Playing atari with deep reinforcement learning.arXiv preprint arXiv:1312.5602,

    arXiv

  7. [17]

    Financial time series forecasting with deep learning: A systematic literature review: 2005–2019.Applied soft computing, 90:106181,

    Omer Berat Sezer, Mehmet Ugur Gudelek, and Ahmet Murat Ozbayoglu. Financial time series forecasting with deep learning: A systematic literature review: 2005–2019.Applied soft computing, 90:106181,

    2005

  8. [19]

    Bloomberggpt: A large language model for finance.arXiv preprint arXiv:2303.17564,

    12 Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhan- jan Kambadur, David Rosenberg, and Gideon Mann. Bloomberggpt: A large language model for finance.arXiv preprint arXiv:2303.17564,

    arXiv

Show all 33 references
  1. [20]

    Tradingagents: Multi-agents llm financial trading framework.arXiv preprint arXiv:2412.20138,

    Yijia Xiao, Edward Sun, Di Luo, and Wei Wang. Tradingagents: Multi-agents llm financial trading framework.arXiv preprint arXiv:2412.20138,

    arXiv

  2. [21]

    Stock movement prediction from tweets and historical prices

    Yumo Xu and Shay B Cohen. Stock movement prediction from tweets and historical prices. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 1970–1979,

    1970

  3. [22]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388,

    arXiv

  4. [23]

    Qlib: An ai-oriented quantitative investment platform.arXiv preprint arXiv:2009.11189,

    Xiao Yang, Weiqing Liu, Dong Zhou, Jiang Bian, and Tie-Yan Liu. Qlib: An ai-oriented quantitative investment platform.arXiv preprint arXiv:2009.11189,

    2009 arXiv

  5. [24]

    Magiclens: Self-supervised image retrieval with open-ended instructions.arXiv preprint arXiv:2403.19651, 2024a

    Kai Zhang, Yi Luan, Hexiang Hu, Kenton Lee, Siyuan Qiao, Wenhu Chen, Yu Su, and Ming-Wei Chang. Magiclens: Self-supervised image retrieval with open-ended instructions.arXiv preprint arXiv:2403.19651, 2024a. Wentao Zhang, Lingxuan Zhao, Haochong Xia, Shuo Sun, Jiaze Sun, Molei...

    arXiv

  6. [25]

    Findeepresearch: Evaluating deep research agents in rigorous financial analysis.arXiv preprint arXiv:2510.13936,

    Fengbin Zhu, Xiang Yao Ng, Ziyang Liu, Chang Liu, Xianwei Zeng, Chao Wang, Tianhui Tan, Xuan Yao, Pengyang Shao, Min Xu, et al. Findeepresearch: Evaluating deep research agents in rigorous financial analysis.arXiv preprint arXiv:2510.13936,

  7. [26]

    Related Work

    13 Appendix •A. Related Work. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .15 •B. Notations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ...

    1992

  8. [28]

    We adopt DeepSeek-R1 Huang & Wang (2025) as the backbone model

    Sentiment Analysis Agent.The sentiment analysis agent processes financial news and social media signals to capture sentiment information that may influence market dynamics. We adopt DeepSeek-R1 Huang & Wang (2025) as the backbone model. Tool-augmented News Retrieval.Given a st...

    2025

  9. [29]

    trading decisions under different cases

    to maph SA s,t into the shared fusion space. News Analysis Agent.As illustrated in Fig. 1, the news analysis agent serves two objectives. First, it retrieves and consolidates news evidence relevant to a queried stock-date pair (s, t). Second, it produces transparent, human-rea...

    2022

  10. [30]

    (2025), while the execution is performed in a single summarization-reasoning step

    draw inspiration from FinMEM Yu et al. (2025), while the execution is performed in a single summarization-reasoning step. Beyond the textual explanation, we extract a fixed-length representation from the last-layer hidden states of the LLM and expose it to the fusion module. L...

    2025

  11. [31]

    Sentiment Analysis Agent.The Sentiment Analysis Agent assesses market sentiment by analyzing news and social media streams (e.g., FinHub)

    and used as the news modality input for downstream fusion. Sentiment Analysis Agent.The Sentiment Analysis Agent assesses market sentiment by analyzing news and social media streams (e.g., FinHub). Given a stock s and date t, it constructs a set of summarized textual items XSA...

    2025

  12. [32]

    Assumptions used in the backtesting setup.Details of the backtesting protocol are specified below

    Thus, HOLD is an execution-state outcome induced by consecutive predicted signals, rather than a third class in the binary price-direction prediction task. Assumptions used in the backtesting setup.Details of the backtesting protocol are specified below. Our experiments use a ...

    2024

  13. [33]

    F2Agent w/o MaF

    Lower values indicate lesser risk and higher strategy robustness. MDD=max( Ppeak −P trough Ppeak ).(22) F.4 DETAILS OFBASELINES We compare the trading performance of F 2Agent against four widely used rule-based strategies (B&H,MACD,ZMR,SMA), two ML&DL-based price prediction mo...

    2020

  14. [1998]

    Qwen2 technical report.arXiv preprint arXiv:2407.10671, 2(3),

    Qwen Team et al. Qwen2 technical report.arXiv preprint arXiv:2407.10671, 2(3),

    arXiv

  15. [2013]

    A comprehensive survey of mixture-of-experts: Algorithms, theory, and applications.arXiv preprint arXiv:2503.07137,

    Siyuan Mu and Sen Lin. A comprehensive survey of mixture-of-experts: Algorithms, theory, and applications.arXiv preprint arXiv:2503.07137,

  16. [2015]

    How green are neural language models? analyzing energy consumption in text summarization fine-tuning.arXiv preprint arXiv:2501.15398,

    Tohida Rehman, Debarshi Kumar Sanyal, and Samiran Chattopadhyay. How green are neural language models? analyzing energy consumption in text summarization fine-tuning.arXiv preprint arXiv:2501.15398,

    arXiv

  17. [2016]

    Chatgpt informed graph neural network for stock movement prediction.arXiv preprint arXiv:2306.03763,

    Zihan Chen, Lei Nico Zheng, Cheng Lu, Jialu Yuan, and Di Zhu. Chatgpt informed graph neural network for stock movement prediction.arXiv preprint arXiv:2306.03763,

    arXiv

  18. [2017]

    101 formulaic alphas.Wilmott, 2016(84):72–81,

    Zura Kakushadze. 101 formulaic alphas.Wilmott, 2016(84):72–81,

    2016

  19. [2018]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    arXiv

  20. [2020]

    Fingpt: Democratizing internet- scale data for financial large language models.arXiv preprint arXiv:2307.10485,

    Xiao-Yang Liu, Guoxuan Wang, Hongyang Yang, and Daochen Zha. Fingpt: Democratizing internet- scale data for financial large language models.arXiv preprint arXiv:2307.10485,

    arXiv

  21. [2021]

    High-modality multimodal transformer: Quantifying modality & interaction heterogeneity for high-modality representation learning.arXiv preprint arXiv:2203.01311,

    Paul Pu Liang, Yiwei Lyu, Xiang Fan, Jeffrey Tsaw, Yudong Liu, Shentong Mo, Dani Yogatama, Louis-Philippe Morency, and Ruslan Salakhutdinov. High-modality multimodal transformer: Quantifying modality & interaction heterogeneity for high-modality representation learning.arXiv p...

    arXiv

  22. [2022]

    Finrl: A deep reinforcement learning library for automated stock trading in quantitative finance.arXiv preprint arXiv:2011.09607,

    11 Xiao-Yang Liu, Hongyang Yang, Qian Chen, Runjia Zhang, Liuqing Yang, Bowen Xiao, and Christina Dan Wang. Finrl: A deep reinforcement learning library for automated stock trading in quantitative finance.arXiv preprint arXiv:2011.09607,

    2011 arXiv

  23. [2023]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948,

    arXiv

  24. [2024]

    A deep reinforcement learning framework for the financial portfolio management problem.arXiv preprint arXiv:1706.10059,

    Zhengyao Jiang, Dixing Xu, and Jinjun Liang. A deep reinforcement learning framework for the financial portfolio management problem.arXiv preprint arXiv:1706.10059,

    arXiv

  25. [2025]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347,

    arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.