From Deep Learning to Large Language Models: A Survey of Artificial Intelligence in Quantitative Investment

In one sentence

Summary

The paper addresses the need for a unified framework to understand the integration of artificial intelligence into quantitative investment, specifically alpha strategies. It positions the field as evolving through three stages: traditional human-crafted features and statistical models, the rise of deep learning for scalable pattern recognition, and the current emergence of large language models (LLMs) that enable autonomous agents to process unstructured data and support self-iterative workflows. The authors aim to bridge the gap between isolated technical studies and a holistic view of the quant pipeline, highlighting practical challenges and future directions for AI-driven alpha generation. The methodology involves a comprehensive literature review structured around the standard alpha investment pipeline: data processing, model prediction, portfolio optimization, and order execution. The authors categorize financial data into numerical, relational, alternative, and simulation types, analyzing how deep learning techniques like graph neural networks and transformers handle these modalities. For LLMs, the survey examines their dual roles as predictors for financial time series and as agents for tasks such as sentiment analysis, factor mining, and automated trading. The paper synthesizes existing works to compare the strengths and weaknesses of each technological stage, emphasizing the shift from supervised learning to agentic, reasoning-based systems. Key findings indicate that while deep learning has significantly improved predictive accuracy by capturing complex non-linear patterns, it often suffers from interpretability issues and overfitting risks. LLMs offer new capabilities in handling multimodal data and generating alpha signals through natural language reasoning, but their practical deployment is still nascent due to latency, hallucination risks, and the need for robust evaluation benchmarks. The survey concludes that the future of quant investment lies in hybrid systems that combine the precision of deep learning with the reasoning and automation capabilities of LLMs, though significant challenges in data quality, regulatory compliance, and real-world validation remain.

Research problem

The paper addresses the fragmentation of research on AI in quantitative investment, where studies often focus on isolated tasks or specific algorithms without a unified perspective. It seeks to provide a comprehensive survey of how deep learning and large language models contribute to the entire alpha strategy pipeline, from data processing to order execution. The problem is to map the evolution of these technologies and identify the practical challenges and future directions for integrating AI into real-world quantitative trading systems.

Core contributions

• Provides the first comprehensive survey connecting deep learning and LLM applications across the entire alpha strategy pipeline.
• Introduces an interdisciplinary framework that links technical AI advancements with practical quantitative investment challenges.
• Systematically compares traditional statistical models, deep learning methods, and LLM-based approaches in terms of capabilities and limitations.
• Identifies key research gaps and future directions for AI-driven alpha generation, particularly in the context of autonomous agents.

Method

• Literature survey covering papers from traditional statistical models to recent LLM applications.
• Categorization of financial data into numerical, relational, alternative, and simulation types.
• Analysis of deep learning models for feature extraction, prediction, and portfolio optimization.
• Review of LLM architectures and their roles as predictors and autonomous agents in trading.
• Comparison of technical approaches across three stages of quant investment evolution.
• Discussion of evaluation metrics and benchmarks for AI models in finance.

Results & findings

• Deep learning has enabled scalable modeling across the alpha pipeline, improving prediction accuracy for complex patterns.
• LLMs are emerging as powerful tools for processing unstructured data and enabling autonomous agent-based workflows.
• Traditional statistical models rely heavily on human-crafted features, limiting scalability and adaptability.
• LLMs show potential in generating alpha signals through natural language reasoning and multimodal data integration.
• Practical deployment of LLMs in trading is still in early stages due to latency, hallucination, and evaluation challenges.
• Hybrid systems combining deep learning precision with LLM reasoning represent the future direction of quant investment.

Data & experiments

• Comprehensive literature review of academic and industry works on AI in quantitative finance.
• Categorization of financial data modalities including numerical, relational, alternative, and simulation data.
• Analysis of deep learning architectures (e.g., GNNs, Transformers) applied to various stages of the alpha pipeline.
• Review of LLM applications in financial prediction, sentiment analysis, and agentic trading workflows.
• Discussion of benchmarks and evaluation metrics used in the field, noting the lack of standardized LLM benchmarks.

Limitations & caveats

• The survey is based on existing literature, which may not capture the latest industry developments.
• It focuses primarily on alpha strategies, potentially overlooking other quantitative investment approaches.
• The discussion of LLMs is based on early-stage research, with limited evidence of large-scale deployment.
• The paper does not provide empirical results or new benchmarks, relying on a synthesis of existing work.
• Practical challenges such as regulatory compliance and data privacy are mentioned but not deeply analyzed.

Curator notes

Bridges deep learning, alpha strategy, and LLM-based investing.