Paper: arXiv 2508.02366
Authors: Adam Darmanin, Vince Vella
Abstract
Algorithmic trading requires short-term tactical decisions consistent with long-term financial objectives. Reinforcement Learning (RL) has been applied to such problems, but adoption is limited by myopic behaviour and opaque policies. Large Language Models (LLMs) offer complementary strategic reasoning and multi-modal signal interpretation when guided by well-structured prompts. This paper proposes a hybrid framework in which LLMs generate high-level trading strategies to guide RL agents. We evaluate (i) the economic rationale of LLM-generated strategies through expert review, and (ii) the performance of LLM-guided agents against unguided RL baselines using Sharpe Ratio (SR) and Maximum Drawdown (MDD). Empirical results indicate that LLM guidance improves both return and risk metrics relative to standard RL.
Complexity vs Empirical Score
- Math Complexity: 6.5/10
- Empirical Rigor: 6.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper employs advanced reinforcement learning algorithms, hierarchical prompting frameworks, and regret minimization heuristics for prompt tuning, indicating significant mathematical complexity; it also provides empirical backtesting using standardized financial metrics (Sharpe Ratio, Max Drawdown) on real data, with a structured benchmark environment and reproducible methodology.
Research Flowchart
flowchart TD A["Research Goal: Address myopic, opaque RL policies in quantitative trading using LLM guidance"] --> B["Methodology: Hybrid Framework LLM generates high-level strategy guiding RL agent"] B --> C["Data/Inputs: Multi-modal market signals, financial objectives, expert review criteria"] C --> D["Computational Process 1: LLM prompt analysis & strategy generation"] D --> E["Computational Process 2: RL agent learning guided by LLM strategy"] E --> F["Computational Process 3: Performance evaluation vs. baseline RL"] F --> G["Key Findings/Outcomes: Expert validation of economic rationale; Improved SR & MDD vs. unguided RL"]
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 a hybrid framework where large language models generate high‑level trading strategies to guide reinforcement‑learning agents.
Source sentence
This paper proposes a hybrid framework in which LLMs generate high-level trading strategies to guide RL agents.
Assumes 2
- Sharpe Ratio calculations assume a zero risk‑free rate. statistical
Source sentence
their SR assumes a zero risk-free rate.
- The hybrid design does not modify the underlying RL algorithm architecture. model
Source sentence
improving out‑of-sample performance without modifying the RL algorithm.
Data
2018 – 2020 · US large‑cap stocks · equities · daily
- Multi‑modal trading dataset market-prices unknown
Source sentence
To support strategy generation, the LLM agents consumed a multi-modal dataset spanning 2012–2020, aligning with the benchmark’s dataset dates in [3]. The dataset includes traditional Open, High, Low, Close, and Volume (OHLCV) price data...
- Interactive Brokers market data market-prices commercial
Source sentence
Market data was sourced from Interactive Brokers1 and iVolatility2, including OHLCV time series, SPX and NDX index returns, the VIX index, and Options implied volatility (IV).
- Fundamental data via SEC‑API and FRED fundamentals public
Source sentence
Fundamental data, comprising firm-level financial ratios and macroeconomic indicators (e.g., GDP, PMI, interest rates), were retrieved via SEC-API3 and the FRED API4.
- Technical analytics features alternative public
Source sentence
Analytics features were computed using TA-Lib5, applying rolling windows to extract indicators.
- News headlines text-news commercial
Source sentence
Alternative data consisting of news headlines were collected from Alpaca6 and processed into explanatory factors using few-shot LLM prompting...
Costs
Not modeled
Finds 2
Sharpe 1.70 (benchmark 0.64) · max drawdown 0.3%
- LLM guidance improves both return and risk metrics relative to standard RL. performance
Source sentence
Empirical results indicate that LLM guidance improves both return and risk metrics relative to standard RL.
- The hybrid LLM+RL agent out‑performed the RL‑only baseline in four out of six assets with statistically significant improvements (p < 0.05). performance
Source sentence
All resulting p-values were below 0.05, indicating statistically significant improvements.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 15 related papers). Coverage is partial, so this is neither replication nor refutation.
2 papers 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 (95)
- IEEE tran Homepage (2007) — link
- IEEE tran Webpage on CTAN (2007) — link
- The IEEE Website (2007) — link
- \BibTeX ing (1988) — link
- Designing \BibTeX Styles (1988) — link
- Tame the BeaST — The B to X of \BibTeX (2005) — link
- \BibTeX Tips and FAQ (2007) — link
- The \TeX FAQ (2007) — link
and 87 more
- A \BibTeX Guide via Examples (2004) — link
- \TeX User Group Bibliography Archive (2006) — link
- The natbib.sty package (2006) — link
- The url.sty Package (2005) — link
- The hyperref.sty Package (2006) — link
- The breakurl.sty Package (2006) — link
- The Babel Package (2005) — link
- The multibib.sty package (2004) — link
- The biblatex package (2007) — link
- A Novel Ultrathin Elevated Channel Low-temperature Poly- Si TFT (1999)
- Butt-jointed DBR Laser With 15 nm Tunability Grown in Three MOVPE Steps (1995)
- Pull-in Time Dynamics as a Measure of Absolute Pressure (1997)
- Introduction to Magnetic Materials (1972)
- Midgap Traps Related to Compensation Processes in CdTe Alloys (1997)
- Analysis and Comparison of PM Synchronous Motor and Induction Motor Type Magnetic Bearings (1995)
- Internet Tomography (2002)
- On the Minimum Distance of Parallel and Serially Concatenated Codes (submitted for publication)
- Oversampling Delta-Sigma Data Converters Theory, Design and Simulation (1992)
- Laser Assisted Microtechnology (1998)
- The Analysis of Directional Time Series: Applications to Wind Speed and Direction (1989)
- A Course in Number Theory (1988)
- Theory Principles and Design of Magnetic Circuits (1964)
- Optical Reflectometry for Component Characterization (1998)
- Properties of the Tailbiting BCJR Decoder (2000)
- Theory for Transmission of Vector Quantization Data (1995)
- Design of an Improved Pixel for a Polysilicon Active-Matrix Organic LED Display (1998)
- FLEXChip Signal Processor ( MC68175/D ) (1996)
- FLEXChip Signal Processor (1996)
- Radial Basis Function Decision Feedback Equaliser Assisted Burst-by-burst Adaptive Modulation (1999)
- Measurement and Modeling of the Temporal Dependence in Packet Loss (1999)
- High Resolution Fiber Distributed Measurements With Coherent OFDR (2000)
- 160 Gbit/s Single-channel Transmission Over 300 km Nonzero-dispersion Fiber With Semiconductor Based Transmitter and Demultiplexer (1999)
- A Novel Approach to Automatic Protection Switching Using Trees (1997)
- High-Resolution Micromachined Interferometric Accelerometer (1992)
- Performance of TCP Congestion Control with Rate Feedback: TCP/ABR and Rate Adaptive TCP/IP (1999)
- Delay Characterization and Performance Control of Wide-area Networks (2000) — link
- Congestion Avoidance in Computer Networks with a Connectionless Network Layer (1987)
- A Stochastic Model of TCP R eno Congestion Avoidance and Control (1999)
- A Tutorial Review of Elements of Weak Signal Detection in Non- G aussian EMI Environments (1986)
- TCP over ATM : ABR or UBR
- Modified TCP Congestion Avoidance Algorithm (1990) — link
- Cache Array Routing Protocol v1.1 (1998) — link
- Optimal Partition of QoS Requirements on Unicast Paths and Multicast Trees (1998) — link
- High-Speed Digital-to- RF Converter (1997)
- Quadrature Modulation Circuit (1992)
- Estimateur d’un D 'e faut de Fonctionnement d’un Modulateur en Quadrature et 'E tage de Modulation l’Utilisant
- IEEE_M_PCOM # “, Special Issue on Wireless ATM " (1996)
- Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specification (1997)
- Fiber Channel Physical Interface ( FC-PI ) (1999)
- MATE : MPLS Adaptive Traffic Engineering (1999)
- Enhanced Proportional Rate Control Algorithm PRCA (1994)
- Advanced QoS Services for the Intelligent Internet (1997)
- PDCA12-70 Data Sheet
- A Proposal to Add Explicit Congestion Notification ( ECN ) to IP (1999)
- SaberDesigner Reference Manual (1998)
- Ucb/lbnl/vint Network Simulator—ns (Version 2) — link
- M essung von S t "o rfeldern an A nlagen und L eitungen der T elekommunikation im F requenzbereich 9 kHz bis 3 GHz
- Telemetry Channel Coding (1999) — link
- Telemetry Channel Coding (1999) — link
- Telemetry Channel Coding (1999) — link
- Telemetry Channel Coding (1999) — link
- Deep Reinforcement Learning: A Brief Survey (2017) — link
- Deep Reinforcement Learning with Double Q-Learning (2016) — link
- Reinforcement Learning: An Introduction (2018) — link
- An Application of Deep Reinforcement Learning to Algorithmic Trading (2021) — link
- Deep Reinforcement Learning Applied to a Sparse-Reward Trading Environment with Intraday Data (2024) — link
- FinRL: Deep Reinforcement Learning Framework to Automate Trading in Quantitative Finance (2021) — link
- Reinforcement Learning for Quantitative Trading (2023) — link
- Beyond Econometrics: A Roadmap Towards Financial Machine Learning (2019) — link
- The 10 Reasons Most Machine Learning Funds Fail (2018) — link
- FinMem: A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design — → 2311.13743
- A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist — → 2402.18485
- Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models — → 2304.07619
- The Memorization Problem: Can We Trust LLMs’ Economic Forecasts? — → 2504.14765
- Leveraging LLMs for Financial News Analysis and Macroeconomic Indicator Nowcasting (2024) — link
- LLMFactor : Extracting Profitable Factors through Prompts for Explainable Stock Movement Prediction (2024) — link
- QuantAgent: Seeking Holy Grail in Trading by Self-Improving Large Language Model — → 2402.03755
- Demystifying Prompts in Language Models via Perplexity Estimation (2023) — link
- The Prompt Report: A Systematic Survey of Prompting Techniques (2024) — link
- Towards Reasoning in Large Language Models: A Survey (2022) — link
- Can Large Language Models Faithfully Express Their Intrinsic Uncertainty in Words? (2024) — link
- Entropy Heat-Mapping: Localizing GPT -Based OCR Errors with Sliding-Window Shannon Analysis (2025) — link
- Artificial Intelligence in Asset Management (2020)
- A Five-Factor Asset Pricing Model (2015) — link
- Examining the Predictive Power of Moving Averages in the Stock Market (2022) — link
- High-frequency trading: Better than its reputation? (2011) — link
- Explainable AI for Interpretable Credit Scoring (2020) — link
5 of 95 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.