The New Quant: A Survey of Large Language Models in Stock Return Prediction and Investment Decision-Making

In one sentence

Summary

The paper addresses the emerging paradigm of the "new quant," where large language models transform unstructured financial information into evidence-grounded signals and executable investment decisions. It positions this shift as a move from traditional feature-centric text mining to end-to-end decision systems that read heterogeneous disclosures, generate auditable hypotheses, and interact with external tools. The survey consolidates insights from domain-specific surveys and more than fifty primary studies published between 2023 and 2025, focusing specifically on equity return prediction and trading with portfolio construction. It aims to provide a structured understanding of how upstream natural language processing components feed tradable signals, addressing the gap between academic research and production-ready quantitative investing. The authors propose a comprehensive task-centered taxonomy that categorizes LLM capabilities into sentiment and opinion extraction, information extraction and knowledge graphs, numerical question answering, summarization, multimodal analysis, and agentic workflows. The methodology involves synthesizing empirical evidence on predictability, reviewing design patterns that improve faithfulness such as retrieval-augmented generation and tool-verified numerics, and analyzing how these signals integrate into portfolio construction under exposure, turnover, and capacity controls. The survey assesses existing benchmarks and datasets, highlighting the need for time-safe evaluation protocols that report costs, latency, and capacity, while also examining challenges like temporal leakage, hallucination, and deployment economics. Key findings indicate that language-derived views on news, filings, and earnings calls can predict returns, but evaluation practices often lack trading standards regarding leakage control and market microstructure realism. The paper highlights that retrieval-first prompting and tool-verified numerics significantly improve faithfulness and reduce hallucination risks. It outlines specific use cases for agentic systems that coordinate tools for research, backtesting, and execution, emphasizing the separation of signal generation from portfolio allocation. Limitations discussed include temporal leakage in pretrained models, data coverage gaps, and the need for interpretability and governance. The survey concludes with recommendations for standardizing evaluation, building auditable pipelines, and advancing multilingual research to ensure robust, risk-controlled performance in practice.

Research problem

The paper addresses the challenge of systematically organizing and evaluating the rapidly growing body of research on large language models in quantitative finance, specifically for equity return prediction and trading. It seeks to bridge the gap between academic NLP capabilities and production-grade quantitative investing by defining a clear taxonomy of tasks and evaluation standards. The problem includes ensuring that LLM-derived signals are economically meaningful, time-safe, and integrated into risk-controlled portfolio construction processes.

Core contributions

• Proposes a task-centered taxonomy mapping LLM capabilities to financial prediction and trading workflows.
• Synthesizes insights from over fifty primary studies on equity return prediction and trading systems.
• Reviews empirical evidence for predictability and highlights design patterns like retrieval-augmented generation for faithfulness.
• Assesses benchmarks and datasets, outlining desiderata for time-safe and economically meaningful evaluation.
• Analyzes production challenges including temporal leakage, hallucination, and deployment economics.
• Provides recommendations for standardizing evaluation and building auditable pipelines for LLM-driven investing.

Method

• Systematic review of primary studies and domain surveys on LLMs in finance.
• Development of a task-centered taxonomy categorizing NLP capabilities for trading.
• Analysis of empirical evidence on return predictability using LLM-derived signals.
• Evaluation of retrieval-augmented and tool-verified methods for improving faithfulness.
• Assessment of agentic frameworks for research, backtesting, and execution coordination.
• Critique of existing benchmarks for time-safety, cost reporting, and market realism.
• Synthesis of governance and interpretability requirements for production deployment.

Results & findings

• LLMs can predict equity returns from unstructured data, but evaluation often lacks trading realism.
• Retrieval-augmented generation and tool-verified numerics improve signal faithfulness and reduce hallucination.
• Agentic systems with memory and tool use show promise for coordinating research and execution.
• Temporal leakage and look-ahead bias remain significant threats to validity in LLM-based prediction.
• Portfolio construction must separate signal generation from allocation to manage exposure and turnover.
• Standardized evaluation requiring cost, latency, and capacity reporting is critical for production deployment.
• Multilingual and cross-market research is needed to generalize LLM performance beyond English-centric datasets.

Data & experiments

• Synthesis of more than fifty primary studies and domain surveys from 2023 to 2025.
• Review of benchmarks including FinQA, FinanceBench, and TradingGPT for prediction and trading.
• Analysis of datasets for sentiment, information extraction, and numerical reasoning in finance.
• Evaluation of agentic systems and multi-agent frameworks for research and execution.
• Discussion of evaluation protocols emphasizing time-safe splits, cost reporting, and capacity constraints.

Limitations & caveats

• Evaluation practices often lack trading standards regarding leakage control and market realism.
• Temporal leakage and look-ahead bias remain significant threats to validity in LLM-based prediction.
• Data coverage and structure issues can limit the generalizability of LLM signals.
• Deployment economics, including cost and latency, are critical but often underreported.
• Interpretability and governance requirements are unevenly addressed in current research.
• Multilingual and cross-market research is limited, restricting generalizability beyond English-centric datasets.

Curator notes

Good second-level survey after general financial LLM surveys.