LLM-Guided Evolutionary Strategy Generation for Quantitative Trading

In one sentence

Summary

The provided paper text details KD-MSLRT, a lightweight sign language recognition model, which is entirely unrelated to the cataloged paper 'LLM-Guided Evolutionary Strategy Generation for Quantitative Trading' (LLM-GA). The cataloged abstract describes LLM-GA, a framework integrating large language models with genetic algorithms for automated trading strategy generation. LLM-GA comprises three modules: a signal generator for technical, fundamental, and sentiment indicators; an LLM-enhanced genetic algorithm core that uses LLMs for semantically-aware crossover and mutation to maintain logical consistency; and an execution module. Experiments on Chinese stock market data (2020-2024) show LLM-GA achieves an Annualized Excess Return of 12.3% and a Maximum Drawdown of 35.2%, outperforming vanilla GA, PSO, and ensemble learning. Ablation studies indicate LLM-guided initialization improves starting strategy quality by 215%, and semantic crossover reduces invalid strategies by 83.5%. While LLM-GA's AER is 2-3% lower than reinforcement learning methods, it offers superior interpretability and diversity, addressing black-box limitations. The work establishes a paradigm for human-AI collaborative quantitative strategy development. The provided text, however, focuses on sign language recognition. It introduces KD-MSLRT, which uses MediaPipe to extract skeletal landmarks from videos, reducing input from 3D video to 1D landmark data. This approach significantly reduces computational burden compared to video-based models. The model employs 3D to 1D knowledge distillation, transferring knowledge from a complex teacher network (CorrNet) to a lightweight student network (MSLR). It also includes a novel text correction network trained via self-supervised learning to correct output errors. The authors release a new large-scale Chinese sign language dataset of 8,976 samples. Experiments on PHOENIX14 and PHOENIX14T datasets show KD-MSLRT achieves a Word Error Rate decrease of at least 1.4% compared to state-of-the-art models, with a model size of 12.93 MB, enabling efficient deployment on resource-constrained devices like Intel CPUs. There is a complete mismatch between the cataloged metadata and the provided paper text. The cataloged paper is about quantitative trading using LLMs and genetic algorithms, while the provided text is about sign language recognition using knowledge distillation and pose estimation. Therefore, no summary of the trading paper can be derived from the provided text. The trading paper's relevance to finance is high, as it directly addresses automated trading strategy generation. The sign language paper has no relevance to finance or LLMs in a financial context. The summary below reflects the cataloged paper's content as described in the abstract, since the text is irrelevant.

Research problem

The paper addresses the challenge of generating high-quality, interpretable, and logically consistent quantitative trading strategies. Traditional genetic algorithms (GA) for strategy generation often produce invalid or illogical strategies due to random crossover and mutation operations. Additionally, black-box methods like reinforcement learning lack interpretability. The research aims to leverage the financial reasoning capabilities of Large Language Models (LLMs) to guide the evolutionary process of trading strategies, ensuring semantic coherence and improving strategy quality.

Core contributions

• Proposes LLM-GA, a novel framework integrating LLMs with genetic algorithms for automated trading strategy generation.
• Introduces an LLM-enhanced GA core that performs semantically-aware crossover and mutation operations to maintain logical consistency.
• Develops a signal generator that extracts technical, fundamental, and sentiment indicators to initialize seed strategies.
• Demonstrates superior risk-adjusted returns compared to vanilla GA, PSO, and ensemble learning baselines on Chinese stock market data.
• Shows that LLM-guided initialization improves starting strategy quality by 215% and semantic crossover reduces invalid strategies by 83.5%.
• Provides unique advantages in strategy interpretability and diversity compared to black-box reinforcement learning methods.

Method

• Signal generator module extracts technical, fundamental, and sentiment indicators from market data.
• LLM-enhanced genetic algorithm core initializes seed strategies using LLM financial reasoning.
• Semantically-aware crossover and mutation operations guided by LLMs to ensure logical consistency.
• Execution module forms a closed-loop adaptive system for strategy deployment.
• Ablation studies to evaluate the contribution of LLM-guided initialization and semantic crossover.
• Comparison with baseline methods (vanilla GA, PSO, ensemble learning) and reinforcement learning.

Results & findings

• LLM-GA achieves an Annualized Excess Return (AER) of 12.3% and a Maximum Drawdown (MDD) of 35.2%.
• LLM-guided initialization improves the quality of starting strategies by 215% compared to random initialization.
• Semantic crossover operations reduce the generation of invalid strategies by 83.5%.
• Performance is 2-3% lower in AER compared to reinforcement learning methods, but offers better interpretability.
• Outperforms vanilla GA, PSO, and ensemble learning in terms of risk-adjusted returns.
• Addresses critical limitations in black-box approaches by providing interpretable and diverse trading strategies.

Data & experiments

• Experiments conducted on historical data of the Chinese stock market from 2020 to 2024.
• Compared against baseline methods including vanilla genetic algorithm (GA), particle swarm optimization (PSO), and ensemble learning.
• Evaluated using risk-adjusted return metrics: Annualized Excess Return (AER) and Maximum Drawdown (MDD).
• Ablation studies performed to isolate the impact of LLM-guided initialization and semantic crossover operations.
• Comparison with reinforcement learning methods to highlight trade-offs between performance and interpretability.

Limitations & caveats

• Performance is slightly lower (2-3% AER) compared to reinforcement learning methods.
• Relies on the financial reasoning capabilities of the LLM, which may be subject to model biases or errors.
• Experiments limited to the Chinese stock market, potentially limiting generalizability to other markets.
• Computational overhead of LLM integration may impact real-time execution speed.
• Interpretability advantages may come at the cost of raw performance compared to black-box methods.

Curator notes

Discovery status: institutional_trading_focus_promoted. Source links: large language models execution trading