Paper: arXiv 2507.08584
Authors: Dimitrios Emmanoulopoulos, Ollie Olby, Justin Lyon, Namid R. Stillman
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.
Complexity vs Empirical Score
- Math Complexity: 7.5/10
- Empirical Rigor: 6.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper involves advanced stochastic calculus (SDEs) and iterative model discovery, indicating high mathematical density. It reports backtesting with real equities and a synthetic market simulator (Simudyne), demonstrating substantial empirical implementation.
Research Flowchart
flowchart TD A["Research Goal: Develop an agentic system that uses LLMs<br>to estimate market risk and inform trading decisions"] --> B["Methodology: Iterative SDE Discovery"] B --> C["Data/Inputs: Historical equity price time series<br>News events (for synthetic market simulator)"] C --> D["Computational Process: Agentic Framework<br>1. LLM proposes SDE form<br>2. System fits parameters & evaluates<br>3. Iterates until model is validated"] D --> E["Outputs: Estimated Risk Metrics<br>from discovered SDE models"] E --> F["Final Action: Model-informed trading strategy<br>executes daily decisions"] F --> G["Key Findings/Outcomes:<br>- Outperforms standard LLM agents<br>- Improves Sharpe ratios across equities<br>- Combines LLMs with model discovery for better risk estimation"]
Evidence Card
What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card
Proposes
The paper proposes an agentic system that uses large language models to iteratively discover stochastic differential equations for financial time series, generating risk metrics to inform daily trading decisions.
Source sentence
We address this gap by developing an agentic system that uses LLMs to iteratively discover stochastic differential equations for financial time series.
Assumes 6
- The Brownian motion increment dW_t is sampled from a normal distribution with mean zero and standard deviation proportional to the time step. model
Source sentence
the stochastic term includes composition with Brownian motion, ◦dWt, which we assume to be sampled from the normal distribution with mean zero and standard deviation proportional to the time step, dWt∼N(0,dt).
- The price process S_t is a continuous R^n‑valued process representing asset price dynamics. model
Source sentence
we assume that St is a continuous Rn-valued process representing the evolution of price dynamics of an asset.
- Only a single independent variable (price) is modeled; volatility and other factors are not included. model
Source sentence
we restrict the space of SDEs to a single independent variable, considering only the evolution of the price, St, and not, for example stochastic models of both volatility and price paths.
- Transaction costs are a fixed proportion κ of the trade value, with κ = 8.8 bps. execution
Source sentence
for an assumed transaction charge, κ, given in basis points (bps). NYSE stocks are estimated to have an average charge of 8.8 bps of the total value of the trade, such that κ = 8.8 × 10−4.
Show 2 more
- Calibration loss is measured as mean absolute error between synthetic and historical price paths. statistical
Source sentence
we calculate the loss, L, as the MAE between synthetic and historical price paths, L(θ)=MAE(St;θ,S∗t;θ)= ΣNsims i |St,i,θ−S∗t,i,θ| / N sims
- All LLMs are pre‑trained on a large portion of the internet up to their cut‑off date, which may introduce bias. other
Source sentence
Assuming that all LLMs have been trained on a significant proportion of the internet up to their cut‑off date, we cannot preclude the possibility of a biasing factor from their pre‑training data affecting their trading performance.
Data
2024-09 – 2025-04 · US equities (e.g., AAPL) · equities · daily
- Simudyne Horizon synthetic market and news generator synthetic unknown
Source sentence
generate an alternative price path and synthetic news cycle for the S&P 500 covering the 6 month trading period from 1st January 2023 to 1st July 2023.
- Historical price data S∗t market-prices unknown
Source sentence
given some historical price data, S∗t , and model-specific parameters, θ.
Costs
Modeled
Transaction cost TCt = κ Pt St with κ = 8.8×10⁻⁴ (8.8 bps) per trade.
Source sentence
for an assumed transaction charge, κ, given in basis points (bps). NYSE stocks are estimated to have an average charge of 8.8 bps of the total value of the trade, such that κ = 8.8 × 10−4.
Finds 5
Sharpe 4.56 · max drawdown 7.0%
- Model‑informed trading strategies outperform standard LLM‑based agents, yielding higher Sharpe ratios across multiple equities. performance
Source sentence
We find that model‑informed trading strategies outperform standard LLM‑based agents, improving Sharpe ratios across multiple equities.
- Adding the model‑discovery loop increases Sharpe ratio by an average of 22% compared with using news‑only context. performance
Source sentence
We see an average 22% increase in SR when comparing LLMs using contextN to contextN M.
- Most LLMs outperform a simple buy‑and‑hold strategy when using news‑only signals, and performance further improves with model discovery. performance
Source sentence
We find that most LLMs are able to outperform the simple buy-and-hold strategy by using news stories only, contextN. The trading results are further improved when combined with a model discovery loop.
- Sonnet 3.5v2, Sonnet 3.7, and 4o‑mini achieve profitable trading decisions in the synthetic backtest, while R1 and o1‑mini perform worse. performance
Source sentence
Sonnet 3.5v2, and Sonnet 3.7 and 4o-mini resulted in profitable trading decisions across the synthetic backtest.
- No LLM outperforms the buy‑and‑hold baseline for NVIDIA due to its large 47% return over the period. negative
Source sentence
for NVIDIA, no LLM is able to outperform the buy-and-hold baseline. This is likely due to the drastic increase in value of the NVIDIA stock over this period (roughly 47% returns).
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 44 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive
Code
No repository by the authors found in the paper.
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (62)
- The Illusion of the Illusion of Thinking (2025)
- tau-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains (2024) — link
- The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity (2025) — link
- From llm reasoning to autonomous ai agents: A comprehensive review (2025) — link
- Quantitative Risk Management. (2011)
- Nonparametric statistical methods (2013)
- Optimal transport: old and new (2008)
- Forecasting: principles and practice (2018)
and 54 more
- Revisiting stylized facts for modern stock markets (2023)
- International Financial Markets Through 150 Years: Evaluating Stylized Facts — → 2504.08611
- Analysis of financial time series (2005)
- Financial modelling with jump processes (2003)
- Automated statistical model discovery with language models (2024) — link
- The ai scientist: Towards fully automated open-ended scientific discovery (2024) — link
- Towards an AI co-scientist (2025) — link
- MacGyver: Are Large Language Models Creative Problem Solvers? (2023) — link
- Human Creativity in the Age of LLMs: Randomized Experiments on Divergent and Convergent Thinking (2024) — link
- A Survey of Frontiers in LLM Reasoning: Inference Scaling, Learning to Reason, and Agentic Systems (2025) — link
- Agentic Large Language Models, a survey (2025) — link
- From Deep Learning to LLMs: A survey of AI in Quantitative Investment — → 2503.21422
- Large language models in finance: A survey — → 2311.10723
- A survey of large language models for financial applications: Progress, prospects and challenges — → 2406.11903
- Revolutionizing finance with llms: An overview of applications and insights (2024) — link
- Can Large Language Models Trade? Testing Financial Theories with LLM Agents in Market Simulations — → 2504.10789
- Large language model agent in financial trading: A survey — → 2408.06361
- FinMem: A performance-enhanced LLM trading agent with layered memory and character design — → 2311.13743
- A modern view on Merton’s jump-diffusion model (2012)
- Estimation in the cox-ingersoll-ross model (1997)
- The constant elasticity of variance model and its implications for option pricing (1980)
- Stochastic processes with applications to finance (2002)
- Stochastic processes for insurance and finance (2009)
- Weisfeiler-lehman graph kernels. (2011)
- DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines (2024)
- Demonstrate-Search-Predict: Composing Retrieval and Language Models for Knowledge-Intensive NLP (2022) — link
- Neuro-Symbolic Traders: Assessing the Wisdom of AI Crowds in Markets — → 2410.14587
- Quantitative risk management: concepts, techniques and tools-revised edition (2015)
- An introduction to statistical modeling of extreme values (2001)
- Alpha-gpt: Human-ai interactive alpha mining for quantitative investment — → 2308.00016
- Quantifying Qualitative Insights: Leveraging LLMs to Market Predict — → 2411.08404
- Tradexpert: Revolutionizing trading with mixture of expert llms — → 2411.00782
- Tradinggpt: Multi-agent system with layered memory and distinct characters for enhanced financial trading performance — → 2309.03736
- Automate strategy finding with llm in quant investment — → 2409.06289
- TradingAgents: Multi-Agents LLM Financial Trading Framework — → 2412.20138
- Understanding uncertainty shocks and the role of black swans (2024)
- A Useful Method for Model-Building (1962)
- On neural differential equations (2022) — link
- Ledex: Training LLMs to better self-debug and explain code (2024)
- Text summarization using large language models: a comparative study of mpt-7b-instruct, falcon-7b-instruct, and openai chat-gpt models (2023) — link
- Can Large Language Models Be an Alternative to Human Evaluations? (2023)
- OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code (2024) — link
- Openai o1 system card (2024) — link
- Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning (2025) — link
- A comparative study of the MACD-base trading strategies: evidence from the US stock market (2022) — link
- LLM Knowledge Cutoff Dates (2024) — link
- Measuring and Modeling Execution Cost and Risk (2006)
- Open-endedness is essential for artificial superhuman intelligence (2024) — link
- Efficient Capital Markets: A Review of Theory and Empirical Work (1970) — link
- Application of natural language processing in financial risk detection — → 2406.09765
- A survey of large language models for financial applications: Progress, prospects and challenges — → 2406.11903
- Stochastic modeling and fair valuation of drawdown insurance (2013)
- Optimization of conditional value-at risk (2000)
- Longrope: Extending llm context window beyond 2 million tokens (2024) — link
16 of 62 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.