The List / Trading, Investment, and Portfolio Management / paper
Alpha-GPT: Human-AI Interactive Alpha Mining for Quantitative Investment
Abstract
One of the most important tasks in quantitative investment research is mining new alphas (effective trading signals or factors). Traditional alpha mining methods, either hand-crafted factor synthesizing or algorithmic factor mining (e.g., search with genetic programming), have inherent limitations, especially in implementing the ideas of quants. In this work, we propose a new alpha mining paradigm by introducing human-AI interaction, and a novel prompt engineering algorithmic framework to implement this paradigm by leveraging the power of large language models. Moreover, we develop Alpha-GPT, a new interactive alpha mining system framework that provides a heuristic way to ``understand''the ideas of quant researchers and outputs creative, insightful, and effective alphas. We demonstrate the effectiveness and advantage of Alpha-GPT via a number of alpha mining experiments.
In one sentence
Summary
The paper addresses the inefficiency and labor-intensity of traditional alpha mining, which relies on manual factor synthesis or compute-heavy algorithmic search. It proposes a third paradigm centered on human-AI interaction, where Large Language Models (LLMs) act as mediators to interpret quant researchers' natural language intuitions and translate them into executable symbolic alpha expressions. This approach aims to bridge the gap between abstract trading concepts and precise mathematical formulations, reducing the cognitive load on researchers and accelerating the discovery of effective trading signals. The proposed system, Alpha-GPT, employs an agentic workflow comprising ideation, implementation, and review stages. In the ideation phase, a 'Trading Idea Polisher' agent uses Retrieval-Augmented Generation (RAG) over financial literature and data field databases to formalize user inputs. The 'Quant Developer' agent then generates seed alpha expressions, which are refined using genetic programming in the implementation stage. The 'Analyst' agent evaluates these alphas via backtesting engines, providing natural language feedback to guide iterative refinement. The system supports both interactive mode, where humans guide the process, and autonomous mode, which uses hierarchical RAG to explore large-scale quantitative databases without overwhelming the LLM's context window. Empirical evaluations demonstrate Alpha-GPT's effectiveness in improving research efficiency and alpha quality. In consistency tests, Alpha-GPT-generated factors outperformed those from junior human researchers, achieving an 86.6% win rate in pairwise comparisons. Search enhancement experiments showed significant improvements in Information Coefficient (IC) over 20 iterations, with out-of-sample IC stabilizing to indicate good generalization. Notably, in the WorldQuant International Quant Championship 2024, Alpha-GPT ranked among the top 10 worldwide out of 41,000 teams, generating 81 qualified alphas with competitive in-sample and out-of-sample scores. The system also successfully captured complex technical patterns like Bollinger band breakouts and moving average divergences, validating its ability to translate abstract ideas into concrete, profitable signals.
Research problem
Traditional alpha mining in quantitative investment suffers from two main limitations: manual factor synthesis is labor-intensive and dependent on individual expertise, while algorithmic search (e.g., genetic programming) is computationally expensive and difficult to guide effectively. There is a need for a more efficient paradigm that leverages human intuition while automating the translation of abstract trading ideas into precise, testable formulaic alphas.
Core contributions
- Proposes a new 'interactive alpha mining' paradigm that integrates human-AI collaboration to enhance the efficiency and effectiveness of factor discovery.
- Introduces AlphaBot, an algorithmic framework that uses domain knowledge compilation and decomposition to enable LLMs to mediate between human intent and algorithmic search.
- Develops Alpha-GPT, a complete system architecture featuring an agentic workflow (ideation, implementation, review) and hierarchical RAG for autonomous exploration of large databases.
- Demonstrates superior performance in the WorldQuant International Quant Championship 2024, ranking in the top 10 globally among 41,000 teams.
- Validates the system's ability to translate natural language trading ideas into high-quality formulaic alphas that outperform junior human researchers in consistency tests.
Method
- Agentic workflow with three stages: Ideation (Trading Idea Polisher), Implementation (Quant Developer), and Review (Analyst).
- Retrieval-Augmented Generation (RAG) using Faiss vector database for financial literature and alpha metadata.
- Hierarchical RAG in autonomous mode: High-level categories -> Second-level categories -> Specific data fields.
- Genetic programming for alpha search enhancement, evolving seed alphas over multiple iterations.
- Backtesting engine calculating IC, Sharpe ratio, and MDD on historical intraday data.
- LLM (Llama3 70B) for natural language understanding, prompt engineering, and alpha expression generation.
- Alpha computation acceleration using streaming algorithms, vectorized computation, and GPU/SIMD instructions.
Results & findings
- Alpha-GPT ranked in the top 10 worldwide in the WorldQuant IQC 2024, generating 81 qualified alphas.
- In pairwise comparisons, Alpha-GPT factors won 86.6% of the time against junior human researchers (score 8.16 vs 6.81).
- Search enhancement improved IC from 0.58% (seed) to 2.23% (after interaction and enhancement).
- Out-of-sample IC stabilized after initial rise, suggesting good generalization and reduced overfitting.
- Alpha-GPT achieved Top 5%-10% performance in a high-frequency trading competition compared to human leaderboards.
- The system successfully generated alphas capturing complex patterns like 'Three White Soldiers' and 'Golden Cross'.
- Autonomous mode with hierarchical RAG effectively navigated databases with >10^4 fields without context overflow.
Data & experiments
- Intraday volume-price data (OHLCV, VWAP) for Chinese and US stocks.
- Evaluation in the WorldQuant International Quant Championship 2024 with over 5,000 data fields and 100+ operators.
- Comparison of Alpha-GPT-generated alphas against those from five junior human quant researchers.
- Backtesting experiments measuring Information Coefficient (IC), Sharpe ratio, and Maximum Drawdown (MDD).
- Search enhancement curves tracking in-sample and out-of-sample IC over 20 iterations of genetic programming.
- Qualitative analysis of alpha consistency with specific technical patterns (e.g., Bollinger bands, moving averages).
Limitations & caveats
- The system relies on the quality of the LLM and the external knowledge library for grounding.
- Autonomous mode may still face challenges in interpreting highly nuanced or novel market conditions not present in training data.
- The effectiveness of genetic programming enhancement depends on the initial seed alphas generated by the LLM.
- Potential for overfitting if the search enhancement process is not carefully monitored, though out-of-sample IC suggests mitigation.
- The system's performance in live trading may differ from backtested results due to market impact and latency.
- Dependence on specific data fields and operators may limit generalizability to other markets or asset classes.
Curator notes
Discovery status: round4_promoted. Source links: MarketSenseAI 2.0: Enhancing Stock Analysis Through LLM Agents; Multimodal Financial Foundation Models (MFFMs): Progress, Prospects, and Challenges; Trade in Minutes! Rationality-Driven Agentic System for Quantitative Financial Trading