Paper: arXiv 2507.17211
Authors: Haochen Luo, Yuan Zhang, Chen Liu
Abstract
Sparse portfolio optimization is a fundamental yet challenging problem in quantitative finance, since traditional approaches heavily relying on historical return statistics and static objectives can hardly adapt to dynamic market regimes. To address this issue, we propose Evolutionary Factor Search (EFS), a novel framework that leverages large language models (LLMs) to automate the generation and evolution of alpha factors for sparse portfolio construction. By reformulating the asset selection problem as a top-m ranking task guided by LLM-generated factors, EFS incorporates an evolutionary feedback loop to iteratively refine the factor pool based on performance. Extensive experiments on five Fama-French benchmark datasets and three real-market datasets (US50, HSI45 and CSI300) demonstrate that EFS significantly outperforms both statistical-based and optimization-based baselines, especially in larger asset universes and volatile conditions. Comprehensive ablation studies validate the importance of prompt composition, factor diversity, and LLM backend choice. Our results highlight the promise of language-guided evolution as a robust and interpretable paradigm for portfolio optimization under structural constraints.
Complexity vs Empirical Score
- Math Complexity: 6.5/10
- Empirical Rigor: 6.0/10 (capped from 7.0: transaction costs not modelled; see penalty flags)
- Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)
Why this score: The paper introduces a novel evolutionary framework combining LLM generation with rigorous backtesting on multiple real-market datasets, though the mathematical depth is moderate (combining ranking theory with evolutionary algorithms) rather than exceptionally advanced.
Research Flowchart
flowchart TD
A["Research Goal: Enhance Sparse Portfolio Optimization with LLM-Generated Factors"] --> B["Data Collection & Setup"]
B --> C["Core Methodology: Evolutionary Factor Search EFS"]
subgraph C ["EFS Framework"]
C1["LLM Prompt Engineering<br>Initial Factor Generation"]
C2["Factor Evaluation<br>Top-m Ranking Performance"]
C3["Evolutionary Loop<br>Selection & Mutation"]
end
B --> D["Baselines & Benchmarks<br>Fama-French + Market Datasets"]
C --> E["Computational Process<br>Iterative Refinement"]
E --> F["Key Findings & Outcomes"]
subgraph F ["Results"]
F1["Outperforms Statistical & Optimization Baselines"]
F2["Scales Well with Large Asset Universes"]
F3["Robust in Volatile Market Conditions"]
F4["Validated via Ablation Studies"]
end
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 Evolutionary Factor Search (EFS), a novel LLM‑guided framework that autonomously generates, evolves, and selects alpha factors for sparse portfolio optimization under ℓ0 constraints.
Source sentence
To address the limitations in current alpha mining algorithms and the challenges of sparse portfolio optimization, we propose Evolutionary Factor Search (EFS). EFS is a novel framework that leverages
Assumes 4
- Portfolio selection is limited to a sparse set of assets (ℓ0 constraint). model
Source sentence
(LLMs) to autonomously generate, evolve, and select alpha factors under ℓ0 constraints.
- Only closing price and returns are used as input variables for factor construction. data
Source sentence
Considering implementation constraints and LLM context limits, we use only two variables—closing price and returns—for factor construction.
- Factor scores are normalized to the range [-1, 1] and the top‑5 scoring assets are selected at each step. execution
Source sentence
All factor scores are normalized to the range [-1, 1], and we select top-5 scoring assets at each step.
- Selected assets are equally weighted throughout the backtest. execution
Source sentence
To isolate factor quality, we apply equal weighting to selected assets throughout.
Data
US equities (US50, HSI45, CSI300) · equities · daily
- FF25 benchmark public
Source sentence
We first use five standard benchmark datasets from Kenneth R. French Data Library [Fama and French, 2023]: FF25, FF32, FF49, FF100 and FF100MEOP.
- FF32 benchmark public
Source sentence
We first use five standard benchmark datasets from Kenneth R. French Data Library [Fama and French, 2023]: FF25, FF32, FF49, FF100 and FF100MEOP.
- FF49 benchmark public
Source sentence
We first use five standard benchmark datasets from Kenneth R. French Data Library [Fama and French, 2023]: FF25, FF32, FF49, FF100 and FF100MEOP.
- FF100 benchmark public
Source sentence
We first use five standard benchmark datasets from Kenneth R. French Data Library [Fama and French, 2023]: FF25, FF32, FF49, FF100 and FF100MEOP.
- US50 market-prices public
Source sentence
US market (US50), comprises the top 50 U.S. large‑cap stocks by market capitalization, selected from the S&P 500 and NASDAQ indices, with data spanning from 2019 to 2024.
- HSI45 market-prices public
Source sentence
Hong Kong market (HSI45), includes the top 45 companies from the Hang Seng Technology Index, based on market capitalization, from 2022 to 2025.
Costs
Not modeled
Finds 2
- EFS achieves the best performance across all five Fama‑French benchmark datasets under the standard sparse setting (m = 10), with only minor differences between GPT‑4.1 and DeepSeek back‑ends. performance
Source sentence
Under the standard sparse setting (m = 10), our proposed EFS framework achieves the best performance across all asset pools, with only minor differences between the GPT-4.1 and DeepSeek backends.
- On the US50 dataset, EFS‑GPT (CW = 39.67, SR = 0.154) and EFS‑DeepSeek (CW = 32.99, SR = 0.149) deliver a 7× improvement over the 1/N baseline. performance
Source sentence
On US50, EFS-GPT (CW = 39.67, SR = 0.154) and EFS-DeepSeek (CW = 32.99, SR = 0.149) deliver 7× improvement over the 1/N baseline.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 22 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper 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 (58)
- Short-term sparse portfolio optimization based on alternating direction method of multipliers (2018)
- A survey of large language models for financial applications: Progress, prospects and challenges — → 2406.11903
- Revolutionizing Finance with LLMs: An Overview of Applications and Insights (2024) — link
- Navigating the Alpha Jungle: An LLM-Powered MCTS Framework for Formulaic Alpha Factor Mining (2026) — link
- A lpha F in: Benchmarking Financial Analysis with Retrieval-Augmented Stock-Chain Framework (2024) — link
- Constructing optimal sparse portfolios using regularization methods (2015)
- Qlib: An AI-oriented Quantitative Investment Platform (2020) — link
- The Markowitz contribution to portfolio theory (1979)
and 50 more
- Asset management: A systematic approach to factor investing (2014)
- Incorporating global industrial classification standard into portfolio allocation: A simple factor-based large covariance matrix estimator with high-frequency data (2016)
- A generalized approach to sparse and stable portfolio optimization problem (2018)
- Sparse and stable Markowitz portfolios (2009)
- Sparse index tracking with $K$-sparsity or $\epsilon$-deviation constraint via $\ell_0$-norm minimization (2022)
- Sparse portfolio selection via the sorted $\ell_1$-Norm (2020)
- A brief review of portfolio optimization techniques (2023)
- Autonomous Sparse Mean- CV a R Portfolio Optimization (2024) — link
- AutoAlpha: an Efficient Hierarchical Evolutionary Algorithm for Mining Alpha Factors in Quantitative Investment (2020) — link
- Generating synergistic formulaic alpha collections via reinforcement learning — → 2306.12964
- Mathematical discoveries from program search with large language models (2024) — link
- Evolution of Heuristics: Towards Efficient Automatic Algorithm Design Using Large Language Model (2024) — link
- Toward automated algorithm design: A survey and practical guide to meta-black-box-optimization (2025)
- ReEvo: Large Language Models as Hyper-Heuristics with Reflective Evolution (2024)
- Multi-Objective Evolution of Heuristic Using Large Language Model (2025) — link
- Discovering Preference Optimization Algorithms with and for Large Language Models (2024)
- Perspective reformulations of mixed integer nonlinear programs with indicator variables (2010)
- F in T ral: A Family of GPT -4 Level Multimodal Financial Large Language Models (2024) — link
- Temporal data meets LLM–explainable financial time series forecasting — → 2306.11025
- Harnessing LLM s for Temporal Data - A Study on Explainable Financial Time Series Forecasting (2023) — link
- FinCon : A Synthesized LLM Multi-Agent System with Conceptual Verbal Reinforcement for Enhanced Financial Decision Making (2024)
- Active portfolio management (2000)
- A Globally Optimal Portfolio for m-Sparse Sharpe Ratio Maximization (2024) — link
- AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay (2025) — link
- AlphaForge: A Framework to Mine and Dynamically Combine Formulaic Alpha Factors — → 2406.18394
- Production of US RM-RF , SMB , and HML in the Fama-French Data Library (2023)
- Alpha- GPT : Human-AI Interactive Alpha Mining for Quantitative Investment — → 2308.00016
- Alpha-GPT 2.0: Human-in-the-Loop AI for Quantitative Investment — → 2402.09746
- LLMF actor: Extracting Profitable Factors through Prompts for Explainable Stock Movement Prediction (2024) — link
- GPT -Signal: Generative AI for Semi-automated Feature Engineering in the Alpha Research Process (2024) — link
- Can Large Language Models Mine Interpretable Financial Factors More Effectively? A Neural-Symbolic Factor Mining Agent Model (2024) — link
- 101 Formulaic Alphas (2016) — link
- Alpha191 (2024) — link
- GPT-4 Technical Report (2024) — link
- DeepSeek-V3 Technical Report (2025) — link
- Algorithm Evolution Using Large Language Model (2023) — link
- LightGBM : A Highly Efficient Gradient Boosting Decision Tree (2017)
- XGBoost : A Scalable Tree Boosting System (2016) — link
- Distribution of eigenvalues for some sets of random matrices (1967)
- A well-conditioned estimator for large-dimensional covariance matrices (2004)
- Noise dressing of financial correlation matrices (1999)
- Cleaning large correlation matrices: tools from random matrix theory (2017)
- The New Quant: A Survey of Large Language Models in Financial Prediction and Trading — → 2510.05533
- Large Language Model Agents in Finance: A Survey (2025)
- QuantaAlpha: An Evolutionary Framework for LLM-Driven Alpha Mining — → 2602.07085
- Cognitive Alpha Mining via LLM-Driven Code-Based Evolution (2025) — link
- Trading-R1: Financial Trading with LLM Reasoning via Reinforcement Learning — → 2509.11420
- AlphaAgents: Large Language Model based Multi-Agents for Equity Portfolio Constructions — → 2508.11152
- HedgeAgents: A Balanced-aware Multi-agent Financial Trading System — → 2502.13165
- Alphabench: Benchmarking large language models in formulaic alpha factor mining (2026)
11 of 58 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.