Pith. sign in

REVIEW 3 major objections 5 minor 75 references

To Trade or Not to Trade: An Agentic Approach to Estimating Market Risk Improves Trading Decisions

T0 review · 3 major / 5 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 An LLM that first discovers a stochastic model of a stock's price path makes better daily trade decisions than an LLM that only reads news.

desk verdict Useful proof-of-concept, but the headline Sharpe gain is confounded by RSI; the SDE-specific claim needs an ablation. read the letter →

arxiv 2507.08584 v1 pith:CLG6XDUZ submitted 2025-07-11 q-fin.ST cs.AIcs.CEcs.MAq-fin.CP

Dimitrios Emmanoulopoulos , Ollie Olby , Justin Lyon , Namid R. Stillman This is my paper
classification q-fin.STcs.AIcs.CEcs.MAq-fin.CP
keywords LLMsagenticfinancemodeldiscoverymarketriskstochasticdifferentialequationsmulti-agenttradingSharperatiobacktesting
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 tries to establish that an agentic trading system benefits from a genuine model-building step, not just news and sentiment. In its framework, an LLM-based builder and critic iterate to discover a stochastic differential equation (SDE) for an equity's price path; the calibrated SDE then produces risk and trend metrics that a trader LLM uses, alongside news, to decide buy, hold, or sell each day. In backtests on four equities over roughly seven months, the average Sharpe ratio rose from 0.88 with news context alone to 1.40 with the model-informed context, a claimed 37% increase; a synthetic-market test with fictional but causally consistent news showed a 22% average increase. The paper also reports that frontier LLMs can recover known target SDEs (CEV, CIR, jump-diffusion) to low calibration loss, with the strongest reasoning models outperforming others at implementation and calibration. If the claims hold, LLM-based trading agents can be improved by letting them construct interpretable stochastic simulators rather than only summarizing text.

What carries the argument

The load-bearing object is the stochastic differential equation discovered by the agent loop, written $dS_t = f(S_t,t;\theta)\,dt + g(S_t,t;\theta)\circ dW_t$. The builder sub-agent turns a candidate equation into calibrated simulator code; the critic sub-agent scores the simulated paths against the historical path using moments, tail statistics, the Hill estimator, Hurst exponent, jump intensity, autocorrelation, and distribution-distance tests, then proposes the next candidate. The loop is bootstrapped with geometric Brownian motion and run monthly, with daily recalibration to the latest 100 trading days. Symbolic similarity between proposed and target SDEs is measured by the Weisfeiler-Lehman graph kernel on drift and diffusion trees. The chosen equation's risk metrics — VaR, CVaR, maximum drawdown, and generalized-Pareto tail parameters from peaks-over-threshold — are passed to the trader along with RSI and drift polarity, and these numbers are what shift the trader's decisions relative to news-only context.

What would settle it

Re-run the pipeline with the model frozen before the evaluation window: after the monthly discovery run on trailing data, lock the chosen SDE and trade the next twenty days without allowing any test-period information to influence selection; repeat for the four equities and compare average Sharpe with and without model metrics. If the average no longer rises from 0.88 to 1.40, the claim fails; a single asset where the chosen SDE's VaR consistently underpredicts realized losses while Sharpe still improves would also signal that something other than risk estimation is driving the trades.

Watch

Extended reading notes

Core claim

The central discovery is that model-informed risk estimation, produced by an LLM-driven discovery loop, improves the trading decisions of an LLM trader over news-only reasoning on the assets and periods tested. The discovery loop represents every candidate as an SDE of the form $dS_t = f(S_t,t;\theta)\,dt + g(S_t,t;\theta)\circ dW_t$; a builder implements and calibrates the equation to the last six months of prices, and a critic scores the simulated paths on summary statistics, tail metrics, and symbolic novelty before suggesting the next candidate. After fifteen discovery runs per asset, the lowest-loss SDE is used to compute value-at-risk, conditional value-at-risk, maximum drawdown, extreme-value-theory tail parameters, RSI, and drift polarity. These metrics are appended to a five-day news summary and given to a trader agent that chooses buy, sell, or hold. Averaged over the seven LLMs that completed the process, the Sharpe ratio rises from 0.88 on news alone to 1.40 on news plus model metrics, and the improvement is largest for LLMs that also performed best at model discovery.

Load-bearing premise

The result depends on the 'most performant SDE' per asset being chosen from the fifteen discovery runs using only trailing calibration data; if test-period performance leaked into that choice, the reported Sharpe improvements would be selection artifacts.

Editorial extensions

If this is right

  • Average Sharpe across the completed runs rises from 0.88 with news only to 1.40 with news plus model metrics on the four equities tested, a 37% improvement.
  • LLMs that are better at model discovery tend to produce better trading results, but the mapping is not one-to-one; a low calibration loss does not guarantee profitable trades, so the full agentic chain matters.
  • Adding the model-informed context helps most on the stock that fell over the period and on the low-return stock, where active decisions are needed to beat buy-and-hold.
  • In the synthetic market simulator with fictional news cycles, most tested LLMs beat buy-and-hold with news alone, and adding model metrics improved average Sharpe by about 22%, suggesting the gain is not simply memorized training-data events.
  • The builder-critic loop recovers the functional form of known target SDEs (CEV, CIR, jump-diffusion) to low calibration loss for most frontier models, indicating that automated model selection for such equations is feasible.

Reading between the lines

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

  • The paper's headline average hides large per-model dispersion: some LLMs trade worse with the model metrics than without them, so the practical claim is conditional on using a capable discovery model.
  • A stronger test of the mechanism would isolate each metric's contribution: holding news fixed and adding only the tail-risk estimates versus only the trend metrics would show which part of the model-informed context drives the improvement.
  • The same loop could be extended to two-factor SDEs with stochastic volatility or stochastic jumps; the paper deliberately restricts to a single price variable, so the risk metrics may understate regime-dependent volatility risk.
  • The synthetic-market result suggests a cheap falsification check: run the same agents with no news at all, pitting model-metrics-only against news-only, to see whether the risk metrics alone can carry the improvement.
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 / 5 minor

Summary. The paper proposes an agentic trading framework in which LLM agents discover stochastic differential equations (SDEs) for historical price paths, calibrate them to data, and compute risk metrics (VaR, CVaR, MDD, EVT) plus trend indicators (RSI, drift polarity). These metrics are passed, together with news summaries, to an LLM trader agent that makes daily buy/hold/sell decisions. The authors evaluate the framework in three settings: (i) discovery of three known SDEs (CEV, CIR, jump-diffusion) with seven LLMs; (ii) a seven-month backtest on AAPL, NVDA, MSFT, and F using news-only context versus news-plus-metrics context; and (iii) a synthetic backtest using the Simudyne Horizon simulator. The central empirical claim is that adding model-informed risk metrics improves mean Sharpe ratios from 0.88 (news-only) to 1.40 (news-plus-model), described as a 37% improvement, and that this result extends to the synthetic backtest.

Significance. If the central claim were fully supported, the paper would be a valuable demonstration that LLM-driven SDE discovery can produce risk metrics that improve agentic trading decisions over a news-only baseline, with a credible attempt to control for LLM pretraining bias via out-of-period data and a synthetic backtest. The symbolic-similarity evaluation of discovered equations (Section 2.1.3) and the explicit comparison of seven LLMs are useful methodological contributions. However, the headline attribution is currently confounded by the inclusion of conventional technical indicators (RSI) in the treatment arm, and the single un-replicated backtest with mixed per-symbol results does not yet support a strong general claim. The paper is honest about several limitations, including the acknowledged inconsistency across LLMs and symbols, which lends credibility to the authors' interpretation, but the missing ablation is a load-bearing gap.

major comments (3)
  1. [Section 2.2.2, Table 2] The central comparison between contextN and contextN_M confounds the effect of the discovered SDE with conventional technical indicators. The treatment arm bundles SDE-derived metrics (VaR, CVaR, MDD, EVT) with RSI (Equation 12) and Drift Polarity (Equation 13), where RSI is computed directly from the price path and is not a product of the model-discovery loop. The news-only baseline receives no quantitative price input at all, as the trader example in Appendix A shows. Therefore the reported average Sharpe improvement from 0.88 to 1.40 could in principle be driven entirely by adding RSI or any simple trend signal, and does not demonstrate that the agentically discovered SDE is the effective ingredient. An ablation isolating the SDE-based metrics (e.g., news+RSI, news+SDE metrics without RSI, and the full contextN_M) is required to support the abstract's claim that model-informed trading outperforms because of the discovered model.
  2. [Section 3.2 vs Section 2.2.2] The model-selection protocol is ambiguous and potentially leaky. Section 2.2.2 states that the best SDE is chosen 'by lowest loss,' whereas Section 3.2 says the authors 'repeat the model discovery loop fifteen times to find the most performant SDE model.' If 'most performant' means the model that performed best on the subsequent trading or backtest period, the reported Sharpe improvements would be selection artifacts rather than out-of-sample results. The manuscript must specify unambiguously that the fifteen runs were ranked solely on training-period calibration loss (or another criterion computed before the evaluation window), and ideally report the distribution of metrics across the fifteen runs. This is load-bearing for the validity of the backtest.
  3. [Section 3.2, Table 2] The evidence for 'improving Sharpe ratios across multiple equities' is statistically weak. The result is a single seven-month backtest on four equities with no error bars, no significance tests, and no repeated runs. Per-symbol results are mixed: for NVDA only Sonnet 3.7 improves, while for MSFT the two OpenAI models (4o-mini and o1-mini) deteriorate. The average improvement from 0.88 to 1.40 may be dominated by a few large gains (e.g., Llama 3.3 on NVDA). To support the general claim, the authors should provide uncertainty estimates (e.g., bootstrap over daily returns), per-symbol consistency analysis, or multiple evaluation periods, and should temper the abstract until such evidence is available.
minor comments (5)
  1. [Section 3.2, paragraph 2] The arithmetic for the claimed improvement is wrong: (1.40 − 0.88)/1.40 = 37%, but a relative improvement over the baseline is (1.40 − 0.88)/0.88 ≈ 59%. Please correct the percentage or reword to '0.52 higher Sharpe ratio.'
  2. [Section 3.1, paragraph after Table 1] The sentence 'The loss is calculated from the calibration process, given in Equation 7' cites the wrong equation; the calibration loss is defined in Equation 2, not Equation 7.
  3. [References] Reference [16] is listed under 'LLM Co-authored References' and contains a malformed arXiv identifier ('arXiv:2:2506.09250') and non-standard authorship ('C Opus and A Lawsen'). Please verify that this is a genuine, correctly formatted reference, or remove it. As written, it undermines the reference list's credibility.
  4. [Throughout] Several typos and formatting issues need correction: 'analyzast' in footnote 1, 'Portofolio' in Figure 7, 'has lead to' in the Introduction, 'exceedences' in Section 2.1.4, and the inconsistent use of italics for agent roles (builder, critic, trader) in the first paragraphs.
  5. [Section 3.3, Table 3] The statement 'We see an average 22% increase in SR' compares averages over different model sets (seven models for contextN but six for contextN_M because Llama 3.3 failed), and the average Sharpe is negative in both cases. Please report the calculation explicitly and avoid describing an improvement from −0.93 to −0.54 as a percentage 'increase' without qualification.

Circularity Check

0 steps flagged · score 2.0 of 10

No equation-level circularity: the Sharpe-ratio claim is an empirical backtest, not a derivation. Minor self-citation and a bundled-signal confound exist, but no claimed result is identical to its inputs by construction.

full rationale

The paper's derivation chain is empirical rather than deductive. SDE parameters are calibrated to trailing price paths via MAE (Eq. 2), risk metrics are computed from the calibrated model (Eqs. 7-11), and trading outcomes are measured out-of-sample in Sections 3.2 and 3.3. No equation reduces to an input by construction. The main caveat is a confound, not circularity: contextN_M bundles SDE-derived VaR/CVaR/MDD/EVT with RSI (Eq. 12), a standard price-based indicator, and Drift Polarity (Eq. 13); the news-only baseline lacks all quantitative price signals, so the 0.88-to-1.40 Sharpe gain does not isolate the discovered SDE. Model selection uses lowest calibration loss (Section 2.2.2); the phrase 'most performant SDE model' in Section 3.2 is ambiguous and could imply test-period selection, but the stated method is loss-based, so no fitted-input-called-prediction reduction is established. Reference [25] is a self-citation by co-author N. Stillman and the Simudyne Horizon simulator is affiliated with a co-author, but neither is load-bearing: the model-discovery capability is independently benchmarked in Section 3.1 against known target SDEs with objective loss and symbolic-similarity metrics. The anomalous 'LLM Co-authored References' entry [16] is not used in the argument. Overall: no significant circularity.

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

The central empirical claim rests on several modeling and evaluation assumptions rather than on new mathematical derivation. The main hand-chosen hyperparameters (EVT threshold, RSI lookback, transaction cost, Monte Carlo path count, discovery iterations) all affect the risk and trend metrics that drive trading decisions, but none is fitted to support a derivation. No new physical or financial entities are introduced; the 'agents' are software components.

free parameters (6)
  • EVT threshold u = 95th percentile of empirical loss distribution
    Set by hand in Section 2.1.4; controls which exceedances are used to fit the generalized Pareto distribution, directly shaping the EVT-based risk metrics passed to the trader.
  • RSI lookback period = 14 days
    Fixed to 14 days per cited literature (Section 2.2.1); affects the trend metric included in trader context.
  • Transaction cost kappa = 8.8 x 10^-4
    Assumed constant from NYSE average (Section 2.2.2); affects PnL and Sharpe comparisons.
  • Monte Carlo path count N_sims
    Not specified in the paper; all VaR, CVaR, and MDD estimates are Monte Carlo estimates whose precision depends on this number.
  • Model discovery iterations N
    Not specified; the final SDE model and hence the risk metrics depend on when the builder-critic loop stops.
  • SDE parameters theta = varies by asset and window
    Calibrated by gradient descent to trailing price data; risk metrics and drift polarity are functions of these fitted parameters.
assumptions (4)
  • domain assumption Financial price paths can be adequately represented by a one-dimensional SDE in price alone, without stochastic volatility or multi-asset coupling.
    Section 2.1.1 explicitly restricts the model space to a single independent variable (St) and notes this excludes stochastic volatility models. If volatility clustering is material, the calibrated risk metrics are misspecified.
  • domain assumption LLM-generated novelty scores and model suggestions are informative for discovering useful SDEs.
    Section 2.1.2 uses an LLM scorer and suggestor without validation against random or exhaustive search; the framework's success depends on these LLM judgments being better than chance.
  • domain assumption The trading period (Sept 2024-Apr 2025) is outside the LLMs' training data, so no futuristic information leaks through the models.
    Section 3.2 asserts this based on published knowledge-cutoff dates ([59]); this cannot be fully verified for proprietary models.
  • domain assumption Simudyne Horizon's synthetic paths and news are a valid proxy for real causally plausible market evolution.
    Section 2.2.4 uses one synthetic S&P 500 scenario; the general conclusion drawn from it assumes the simulator's price-news alignment is representative.

how reviews work

0 comments
Cite this review

Pith. "Pith review of To Trade or Not to Trade: An Agentic Approach to Estimating Market Risk Improves Trading Decisions." pith.science (2026). https://pith.science/paper/CLG6XDUZ

@misc{pith2026250708584,
  author       = {Pith},
  title        = {Pith review of: To Trade or Not to Trade: An Agentic Approach to Estimating Market Risk Improves Trading Decisions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CLG6XDUZ}},
  note         = {Machine review of arXiv:2507.08584}
}
read the original abstract

Large language models (LLMs) are increasingly deployed in agentic frameworks, in which prompts trigger complex tool-based analysis in pursuit of a goal. While these frameworks have shown promise across multiple domains including in finance, they typically lack a principled model-building step, relying instead on sentiment- or trend-based analysis. We address this gap by developing an agentic system that uses LLMs to iteratively discover stochastic differential equations for financial time series. These models generate risk metrics which inform daily trading decisions. We evaluate our system in both traditional backtests and using a market simulator, which introduces synthetic but causally plausible price paths and news events. We find that model-informed trading strategies outperform standard LLM-based agents, improving Sharpe ratios across multiple equities. Our results show that combining LLMs with agentic model discovery enhances market risk estimation and enables more profitable trading decisions.

Figures

Figures reproduced from arXiv: 2507.08584 by the authors.

Figure 1
Figure 1. Agentic method for estimating market risk. We identify a stochastic model of historical price paths using an agentic approach. Here, we follow a builder -critic pattern for model discovery, in which AI agents implement, simulate, test, and score proposed stochastic models. These models are generated as the output of our agents. We use this to construct and calibrate stochastic models of price paths which provide mod… view at source ↗
Figure 2
Figure 2. Risk analyst agentic workflow. The risk analyst uses a builder (dashed boxes) and critic (dotted boxes) framework for model discovery, in which the builder is tasked with implementation and simulation of the time series. Meanwhile, the critic calculates evaluation metrics on the simulated data and uses this to proposed new models to be implemented by the builder. methods, scales to an arbitrary number of parameters,… view at source ↗
Figure 3
Figure 3. Symbolic trees. We convert the stochastic differential equation, used to simulate the financial time series, into a directed acyclic symbolic tree which we then use to calculate symbolic similarity and diversity scores. Here, we show (a) the CIR model, (b) the CEV model, and (c) the JD model, which are described in Section 3.1. derived from an additional agent, using the model summary to determine a score between 1-… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Overview of the trading strategy pipeline. Each month, the model discovery phase identifies a suitable SDE based on the past six months of price data. On each trading day, the SDE is recalibrated to the latest market conditions, recent news is fetched via the FinnHub A…
Figure 5
Figure 5. Figure 5: Synthetic news generation. To ensure that the tested LLMs do not have relevant context in their pretrained weights to improve trading results, we use the Simudyne Horizon simulator to generate synthetic price and news data. This provides a more advanced form of context…
Figure 6
Figure 6. Figure 6: Model discovery loop examples. To ex￾amine how effective the model discovery loop is at identifying suitable candidate models of the SDE we examine the time series of the ground truth (black) and an array of Monte Carlo simulations of the candidate model. All time, t, …
Figure 7
Figure 7. Figure 7: Examples of trading results for each of the symbols considered. Backtesting results from agentic trading strategies shown in orange for the period covering 17th September 2024 to 24th April 2025. We show results for Llama 3.3 trading on AAPL, Sonnet 3.7 trading on MSFT…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

75 extracted references · 41 canonical work pages

  1. [1]

    Application of natural language processing in financial risk detection

    Liyang Wang, Yu Cheng, Ao Xiang, Jingyu Zhang, and Haowei Yang. “Application of natural language processing in financial risk detection”. In: arXiv preprint arXiv:2406.09765 (2024)

    arXiv 2024

  2. [3]

    From llm reasoning to au- tonomous ai agents: A comprehensive review

    Mohamed Amine Ferrag, Norbert Tihanyi, and Merouane Debbah. “From llm reasoning to au- tonomous ai agents: A comprehensive review”. In: arXiv preprint arXiv:2504.19678 (2025)

    arXiv 2025

  3. [4]

    FinMem: A performance-enhanced LLM trading agent with layered memory and character design

    Yangyang Yu et al. “FinMem: A performance-enhanced LLM trading agent with layered memory and character design”. In: Proceedings of the AAAI Symposium Series . Vol. 3. 1. 2024, pp. 595– 597

    work page 2024

  4. [5]

    Large language model agent in financial trading: A survey

    Han Ding, Yinheng Li, Junhao Wang, and Hang Chen. “Large language model agent in financial trading: A survey”. In: arXiv preprint arXiv:2408.06361 (2024)

    arXiv 2024

  5. [6]

    Can Large Language Models Trade? Testing Financial Theories with LLM Agents in Market Simulations

    Alejandro Lopez-Lira. “Can Large Language Models Trade? Testing Financial Theories with LLM Agents in Market Simulations”. In: arXiv preprint arXiv:2504.10789 (2025). 21

    arXiv 2025

  6. [7]

    Revolutionizing finance with llms: An overview of applications and insights

    Huaqin Zhao et al. “Revolutionizing finance with llms: An overview of applications and insights”. In: arXiv preprint arXiv:2401.11641 (2024)

    arXiv 2024

  7. [8]

    A survey of large language models for financial applications: Progress, prospects and challenges

    Yuqi Nie et al. “A survey of large language models for financial applications: Progress, prospects and challenges”. In: arXiv preprint arXiv:2406.11903 (2024)

    arXiv 2024

  8. [9]

    Large language models in finance: A survey

    Yinheng Li, Shaofei Wang, Han Ding, and Hang Chen. “Large language models in finance: A survey”. In: Proceedings of the fourth ACM international conference on AI in finance . 2023, pp. 374–382

    work page 2023

Show all 75 references
  1. [10]

    From Deep Learning to LLMs: A survey of AI in Quantitative Investment

    Bokai Cao et al. “From Deep Learning to LLMs: A survey of AI in Quantitative Investment”. In: arXiv preprint arXiv:2503.21422 (2025)

    2025 arXiv

  2. [11]

    Agentic Large Language Models, a survey

    Aske Plaat et al. “Agentic Large Language Models, a survey”. In: arXiv preprint arXiv:2503.23037 (2025)

    2025

  3. [12]

    A Survey of Frontiers in LLM Reasoning: Inference Scaling, Learning to Reason, and Agentic Systems

    Zixuan Ke et al. “A Survey of Frontiers in LLM Reasoning: Inference Scaling, Learning to Reason, and Agentic Systems”. In: arXiv preprint arXiv:2504.09037 (2025)

    2025

  4. [13]

    Openai o1 system card

    Aaron Jaech et al. “Openai o1 system card”. In: arXiv preprint arXiv:2412.16720 (2024)

    2024 arXiv

  5. [14]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learn- ing

    Daya Guo et al. “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learn- ing”. In: arXiv preprint arXiv:2501.12948 (2025)

    2025 arXiv

  6. [15]

    The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity

    Parshin Shojaee et al. “The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity”. In: arXiv preprint arXiv:2506.06941 (2025)

    2025 arXiv

  7. [16]

    The Illusion of the Illusion of Thinking

    C Opus and A Lawsen. “The Illusion of the Illusion of Thinking”. In: arXiv preprint arXiv:2:2506.09250 (2025). 24 Appendix A Example Agent Output Summarising Agent This is a non-linear SDE with interesting properties: Looking at the model code, we can see: • The drift term is ...

    2025 arXiv

  8. [17]

    tau-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains

    Shunyu Yao, Noah Shinn, Pedram Razavi, and Karthik Narasimhan. “tau-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains”. In: arXiv preprint arXiv:2406.12045 (2024)

    2024 arXiv

  9. [18]

    TradingAgents: Multi-Agents LLM Financial Trading Framework

    Yijia Xiao, Edward Sun, Di Luo, and Wei Wang. “TradingAgents: Multi-Agents LLM Financial Trading Framework”. In: arXiv preprint arXiv:2412.20138 (2024)

    2024 arXiv

  10. [19]

    Alpha-gpt: Human-ai interactive alpha mining for quantitative investment

    Saizhuo Wang et al. “Alpha-gpt: Human-ai interactive alpha mining for quantitative investment”. In: arXiv preprint arXiv:2308.00016 (2023)

    2023

  11. [20]

    Automate strategy finding with llm in quant investment

    Zhizhuo Kou et al. “Automate strategy finding with llm in quant investment”. In: arXiv preprint arXiv:2409.06289 (2024)

    2024

  12. [21]

    Tradexpert: Revolutionizing trading with mixture of expert llms

    Qianggang Ding, Haochen Shi, Jiadong Guo, and Bang Liu. “Tradexpert: Revolutionizing trading with mixture of expert llms”. In: arXiv preprint arXiv:2411.00782 (2024)

    2024 arXiv

  13. [22]

    Tradinggpt: Multi- agent system with layered memory and distinct characters for enhanced financial trading perfor- mance

    Yang Li, Yangyang Yu, Haohang Li, Zhi Chen, and Khaldoun Khashanah. “Tradinggpt: Multi- agent system with layered memory and distinct characters for enhanced financial trading perfor- mance”. In: arXiv preprint arXiv:2309.03736 (2023)

    2023 arXiv

  14. [23]

    Quantifying Qualitative Insights: Leveraging LLMs to Market Predict

    Hoyoung Lee, Youngsoo Choi, and Yuhee Kwon. “Quantifying Qualitative Insights: Leveraging LLMs to Market Predict”. In: arXiv preprint arXiv:2411.08404 (2024)

    2024 arXiv

  15. [24]

    Automated statistical model discovery with language models

    Michael Y Li, Emily B Fox, and Noah D Goodman. “Automated statistical model discovery with language models”. In: arXiv preprint arXiv:2402.17879 (2024)

    2024 arXiv

  16. [25]

    Neuro-Symbolic Traders: Assessing the Wisdom of AI Crowds in Markets

    Namid R Stillman and Rory Baggott. “Neuro-Symbolic Traders: Assessing the Wisdom of AI Crowds in Markets”. In: arXiv preprint arXiv:2410.14587 (2024)

    2024 arXiv

  17. [26]

    A Useful Method for Model-Building

    G. E. P. Box and William G. Hunter. “A Useful Method for Model-Building”. In: Technometrics 4.3 (1962), pp. 301–318. issn: 00401706

    1962

  18. [27]

    On neural differential equations

    Patrick Kidger. “On neural differential equations”. In: arXiv preprint arXiv:2202.02435 (2022)

    2022 arXiv

  19. [28]

    DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines

    Omar Khattab et al. “DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines”. In: 2024. 22

    2024

  20. [29]

    Understanding uncertainty shocks and the role of black swans

    Anna Orlik and Laura Veldkamp. “Understanding uncertainty shocks and the role of black swans”. In: Journal of Economic Theory 222 (2024), p. 105905. issn: 0022-0531

    2024

  21. [30]

    Text summarization using large language models: a compara- tive study of mpt-7b-instruct, falcon-7b-instruct, and openai chat-gpt models

    Lochan Basyal and Mihir Sanghvi. “Text summarization using large language models: a compara- tive study of mpt-7b-instruct, falcon-7b-instruct, and openai chat-gpt models”. In: arXiv preprint arXiv:2310.10449 (2023)

    2023 arXiv

  22. [31]

    Ledex: Training LLMs to better self-debug and explain code

    Nan Jiang et al. “Ledex: Training LLMs to better self-debug and explain code”. In: Advances in Neural Information Processing Systems 37 (2024), pp. 35517–35543

    2024

  23. [32]

    OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code

    Maxence Faldor, Jenny Zhang, Antoine Cully, and Jeff Clune. “OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code”. In: arXiv preprint arXiv:2405.15568 (2024)

    2024 arXiv

  24. [33]

    Open-endedness is essential for artificial superhuman intelligence

    Edward Hughes et al. “Open-endedness is essential for artificial superhuman intelligence”. In: arXiv preprint arXiv:2406.04268 (2024)

    2024 arXiv

  25. [34]

    Analysis of financial time series

    Ruey S Tsay. Analysis of financial time series . John wiley & sons, 2005

    2005

  26. [35]

    Financial modelling with jump processes

    Rama Cont and Peter Tankov. Financial modelling with jump processes. Chapman and Hall/CRC, 2003

    2003

  27. [36]

    International Financial Markets Through 150 Years: Evaluating Stylized Facts

    Sara A Safari, Maximilian Janisch, and Thomas Leh´ ericy. “International Financial Markets Through 150 Years: Evaluating Stylized Facts”. In: arXiv preprint arXiv:2504.08611 (2025)

    2025 arXiv

  28. [37]

    Revisiting stylized facts for modern stock markets

    Ethan Ratliff-Crain, Colin M Van Oort, James Bagrow, Matthew TK Koehler, and Brian F Tiv- nan. “Revisiting stylized facts for modern stock markets”. In:2023 IEEE International Conference on Big Data (BigData) . IEEE. 2023, pp. 1814–1823

    2023

  29. [38]

    Forecasting: principles and practice

    Rob J Hyndman and George Athanasopoulos. Forecasting: principles and practice. OTexts, 2018

    2018

  30. [39]

    Optimal transport: old and new

    C´ edric Villani et al. Optimal transport: old and new . Vol. 338. Springer, 2008

    2008

  31. [40]

    Nonparametric statistical methods

    Myles Hollander, Douglas A Wolfe, and Eric Chicken. Nonparametric statistical methods . John Wiley & Sons, 2013

    2013

  32. [41]

    Can Large Language Models Be an Alternative to Human Evaluations?

    Cheng-Han Chiang and Hung-yi Lee. “Can Large Language Models Be an Alternative to Human Evaluations?” In: Annual Meeting of the Association for Computational Linguistics . 2023

    2023

  33. [42]

    Human Creativity in the Age of LLMs: Randomized Experiments on Divergent and Convergent Thinking

    Harsh Kumar, Jonathan Vincentius, Ewan Jordan, and Ashton Anderson. “Human Creativity in the Age of LLMs: Randomized Experiments on Divergent and Convergent Thinking”. In: arXiv preprint arXiv:2410.03703 (2024)

    2024 arXiv

  34. [43]

    MacGyver: Are Large Language Models Creative Problem Solvers?

    Yufei Tian et al. “MacGyver: Are Large Language Models Creative Problem Solvers?” In: arXiv preprint arXiv:2311.09682 (2023)

    2023 arXiv

  35. [44]

    Demonstrate-Search-Predict: Composing Retrieval and Language Models for Knowledge-Intensive NLP

    Omar Khattab et al. “Demonstrate-Search-Predict: Composing Retrieval and Language Models for Knowledge-Intensive NLP”. In: arXiv preprint arXiv:2212.14024 (2022)

    2022 arXiv

  36. [45]

    Weisfeiler-lehman graph kernels

    Nino Shervashidze, Pascal Schweitzer, Erik Jan Van Leeuwen, Kurt Mehlhorn, and Karsten M Borgwardt. “Weisfeiler-lehman graph kernels.” In: Journal of Machine Learning Research 12.9 (2011)

    2011

  37. [46]

    Stochastic processes for insurance and finance

    Tomasz Rolski, Hanspeter Schmidli, Volker Schmidt, and Jozef L Teugels. Stochastic processes for insurance and finance . John Wiley & Sons, 2009

    2009

  38. [47]

    Stochastic processes with applications to finance

    Masaaki Kijima. Stochastic processes with applications to finance. Chapman and Hall/CRC, 2002

    2002

  39. [48]

    Stochastic modeling and fair valuation of drawdown insurance

    Hongzhong Zhang, Tim Leung, and Olympia Hadjiliadis. “Stochastic modeling and fair valuation of drawdown insurance”. In: Insurance: Mathematics and Economics 53.3 (Nov. 2013), pp. 840–

    2013

  40. [49]

    Optimization of conditional value-at risk

    R. Tyrrell Rockafellar and Stanislav Uryasev. “Optimization of conditional value-at risk”. In: Journal of Risk 3 (2000), pp. 21–41. 23

    2000

  41. [50]

    Quantitative Risk Management

    Paul Embrechts, Rdiger Frey, and Alexander McNeil. Quantitative Risk Management. 2011

    2011

  42. [51]

    An introduction to statistical mod- eling of extreme values

    Stuart Coles, Joanna Bawa, Lesley Trenner, and Pat Dorazio. An introduction to statistical mod- eling of extreme values . Vol. 208. Springer, 2001

    2001

  43. [52]

    Princeton university press, 2015

    Alexander J McNeil, R¨ udiger Frey, and Paul Embrechts.Quantitative risk management: concepts, techniques and tools-revised edition. Princeton university press, 2015

    2015

  44. [53]

    A comparative study of the MACD-base trading strategies: evidence from the US stock market

    Pat Tong Chio. “A comparative study of the MACD-base trading strategies: evidence from the US stock market”. In: arXiv preprint arXiv:2206.12282 (2022)

    2022 arXiv

  45. [54]

    Longrope: Extending llm context window beyond 2 million tokens

    Yiran Ding et al. “Longrope: Extending llm context window beyond 2 million tokens”. In: arXiv preprint arXiv:2402.13753 (2024)

    2024 arXiv

  46. [55]

    Measuring and Modeling Execution Cost and Risk

    Robert F. Engle, R. Ferstenberg, and Jeffrey R. Russell. “Measuring and Modeling Execution Cost and Risk”. In: IO: Firm Structure (2006)

    2006

  47. [56]

    The constant elasticity of variance model and its implications for option pricing

    Stan Beckers. “The constant elasticity of variance model and its implications for option pricing”. In: the Journal of Finance 35.3 (1980), pp. 661–673

    1980

  48. [57]

    Estimation in the cox-ingersoll-ross model

    Ludger Overbeck and Tobias Ryden. “Estimation in the cox-ingersoll-ross model”. In: Econometric Theory 13.3 (1997), pp. 430–461

    1997

  49. [58]

    A modern view on Merton’s jump-diffusion model

    Gerald HL Cheang and Carl Chiarella. “A modern view on Merton’s jump-diffusion model”. In: Stochastic Processes, Finance and Control: A Festschrift in Honor of Robert J Elliott . World Scientific, 2012, pp. 217–234

    2012

  50. [59]

    LLM Knowledge Cutoff Dates

    Hao Wang. LLM Knowledge Cutoff Dates . GitHub repository. 2024. url: https://github.com/ HaoooWang/llm-knowledge-cutoff-dates

    2024

  51. [60]

    The ai scientist: Towards fully automated open-ended scientific discovery

    Chris Lu et al. “The ai scientist: Towards fully automated open-ended scientific discovery”. In: arXiv preprint arXiv:2408.06292 (2024)

    2024 arXiv

  52. [61]

    Towards an AI co-scientist

    Juraj Gottweis et al. “Towards an AI co-scientist”. In: arXiv preprint arXiv:2502.18864 (2025). LLM Co-authored References

    2025 arXiv

  53. [62]

    Novelty (50% of score): • The model uses a sine function which is a novel approach for SDEs • The structure dS = a*sin(S/b)dt + cSdW is relatively simple but introduces non- linear behavior • The sine function could capture cyclical patterns in the data • Score for novelty: 35...

  54. [63]

    The news suggests this is affecting various tech companies, particularly those involved in AI and semiconductors

    Metrics Accuracy (50% of score): • Delta metrics show good alignment in mean (0.001) and jump intensity (0.005) • Volatility is reasonably close (delta of 0.278) • ACF values are fairly close to target (deltas around 0.02-0.04) • Kurtosis and skewness show moderate differences...

  55. [64]

    Direct competitors in the GPU/AI chip space (AMD, INTC)

  56. [65]

    Key semiconductor manufacturing partners (TSM)

  57. [66]

    Major AI technology companies affected by DeepSeek developments (MSFT, GOOGL)

  58. [67]

    News Analyst Agent The emergence of DeepSeek and other competitive threats suggests NVIDIA’s dominant position in AI chips could face increasing pressure

    Companies involved in AI infrastructure (ORCL) These companies will provide a complete picture of the competitive landscape, supply chain, and broader AI market dynamics affecting NVDA. News Analyst Agent The emergence of DeepSeek and other competitive threats suggests NVIDIA’...

    2025

  59. [68]

    patently absurd

    V aluation Concerns: Multiple sources indicate AI chip stocks, including NVIDIA, are trading at “patently absurd” valuations with triple-digit P/E ratios that appear discon- nected from fundamentals, suggesting an overvalued position

  60. [69]

    Emerging Competitive Threats : The development of China’s DeepSeek AI tool rep- resents a tangible competitive threat that has already caused an 8% pre-market drop in NVIDIA’s stock, indicating the market’s sensitivity to competitive pressures in the AI chip space

  61. [70]

    inevitable market correction

    AI Spending Bubble Risk: There are substantial concerns about an AI spending bubble, with analysts warning of an “inevitable market correction” for AI tech stocks. This suggests NVIDIA’s current price may be inflated by unsustainable investment patterns

  62. [71]

    Project Stargate

    Market Speculation Over F undamentals: Multiple articles indicate that NVIDIA and other AI stocks are trading based on speculation (particularly around “Project Stargate”) rather than concrete financial performance, creating vulnerability to sentiment shifts

  63. [72]

    DeepSeek problem

    Short-T erm Negative Catalysts: The “DeepSeek problem” specifically identified for NVIDIA represents an immediate concern that could continue to pressure the stock in the near term. While NVIDIA maintains a strong position in the AI market with potential long-term govern- ment...

    2025

  64. [73]

    • The Maximum Drawdown (MDD) of 52.70% is substantial and indicates historical vulnerability to major corrections

    Mixed Risk Profile : • The risk metrics show some concerning elements: The EVTVaR (20.79) and EVTCVaR (25.50) are significantly higher than the standard VaR (9.90) and CVaR (16.08), indi- cating that extreme negative movements could be worse than standard risk measures suggest...

  65. [74]

    • The positive Drift Score suggests some upward momentum in the price action

    Neutral T echnical Indicators: • The RSI at 52.21 is in neutral territory, neither overbought nor oversold. • The positive Drift Score suggests some upward momentum in the price action. • Slightly negative residuals (-0.48) indicate recent underperformance relative to the mode...

  66. [75]

    Magnificent 7,

    Balanced F undamental Picture: • Strong positives include NVIDIA’s dominant AI market position, inclusion in the “Magnificent 7,” and potential benefits from the incoming Trump administration’s “Project Stargate.” • Significant concerns include valuation worries, competition f...

  67. [76]

    AI bubble

    Risk vs. Reward Assessment : • While the positive drift suggests upside potential, the elevated tail risk metrics (EVTVaR, EVT CVaR) indicate substantial downside risk, particularly given the “AI bubble” concerns mentioned in the analyst thoughts. • The neutral RSI and current...

    2022

Pith tools

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