AlphaLogics: A Market Logic-Driven Multi-Agent System for Scalable and Interpretable Alpha Factor Generation

Zhangyuhua Weng; Shengli Zhang; Taotao Wang; Yihan Xia

q-fin.CPMar 10, 2026

AlphaLogics: A Market Logic-Driven Multi-Agent System for Scalable and Interpretable Alpha Factor Generation

Authors: Zhangyuhua Weng, Shengli Zhang, Taotao Wang, Yihan Xia

Organizations: College of Electronic and Information Engineering, Shenzhen University

Abstract

Factor investing is ultimately grounded in market logic - the latent mechanism behind observed alpha factors that explains why they should persist across assets and regimes. However, recent factor mining prioritizes factor discovery over logic discovery, producing complex alpha factors with unclear rationale, while market logic remains largely handcrafted and difficult to scale. To address this challenge, we propose AlphaLogics, a market logic-driven multi-agent system for factor mining. AlphaLogics consists of three key components: (i) Market Logic Mining: reverse-extracting market logic from historical factor libraries to construct an initial market logic library; (ii) Factor Generation and Optimization: using new market logics generated in (i) to guide factor generation, and optimizing factors with backtesting feedback; and (iii) Market Logic Generation and Optimization: generating new market logics conditioned on the initial market logic library, and refining each market logic by aggregating the backtest outcomes of its guided factors, continuously refreshing the library. Experiments on CSI 500 and S&P 500 show that AlphaLogics consistently improves predictive metrics and risk-adjusted returns over representative baselines, while producing a market logic library that remains empirically useful for guiding further factor discovery.

Figures & tables

Figure 1: Comparison of advantages and disadvantages of alpha factor mining methods. Manual Market Logic-driven factor mining (left) offers high interpretability but lacks scalability; Market data driven factor mining (middle) scales well but often lacks interpretability; Our method (right) offers high interpretability and scalability.

Figure 2: The autonomous workflow of AlphaLogics . The framework operates in three stages: (1) Market Logic Mining extracts latent market logic from historical factor libraries; (2) Guided Factor Generation guides new factors generation and optimization with backtesting feedback; (3) Market Logic Generation generates new market logics conditioned on the initial market logic library, and refining each market logic by aggregating the backtest outcomes of its guided factors, continuously refreshing the library.

Method	CSI 500 (2021.01–2024.12)					S&P 500 (2021.01–2024.12)
	IC	ICIR	AR	IR	MDD	IC	ICIR	AR	IR	MDD
LSTM ( Graves, 2012 )	0.0162	0.1173	6.33%	0.8494	-10.45%	0.0032	0.0177	-7.29%	-0.9732	-21.91%
Transformer ( Vaswani et al., 2017 )	0.0150	0.1234	4.03%	0.3650	-22.85%	0.0014	0.0108	0.82%	0.0952	-17.71%
LightGBM ( Ke et al., 2017 )	0.0116	0.0972	0.88%	0.0782	-24.37%	-0.0014	-0.0131	-1.23%	-0.1517	-28.28%
TRA ( Lin et al., 2021 )	0.0199	0.1765	1.60%	0.1467	-26.11%	0.0030	0.0204	-3.92%	-0.4541	-34.15%
GRU ( Chung et al., 2014 )	0.0115	0.0994	-1.71%	-0.1733	-22.05%	0.0050	0.0297	-1.14%	-0.0970	-32.10%

Table 1: Comparison of factor performance on CSI 500 and S&P 500 over the held-out test period (2021.01–2024.12), using a fixed split: train (2015.01–2019.12), validation (2020.01–2020.12), and test (2021.01–2024.12). Bold indicates the best performance among all methods for each metric within the same market.

Factor Library	Math. Expl.	Fin. Expl.
Alpha101	97.5%	92.7%
Alpha158	98.1%	95.5%
Alpha360	100.0%	98.8%
Alpha191	94.9%	93.8%

Table 2: Consistency rates between original factor formulas and formulas reconstructed from LLM-provided interpretations. Math. Expl. denotes reconstruction consistency based on mathematical explanations of factor formulas (e.g., window length, smoothing, ranking, and normalization operations), while Fin. Expl. denotes consistency based on financial explanations mapping factors to economic meaning and market behavior.

Figure 3: Comparison of market logic-guided factor generation on CSI 500 and S&P 500 (train 2015.01–2019.12, validation 2020.01–2020.12, test 2021.01–2024.12). The results demonstrate that across all models and market settings, factors generated with market logic guidance consistently outperform unconstrained generation in terms of IC and IR metrics.

Figure 4: Evolution indicators of market logic. We test GPT-3.5-Turbo ( Ouyang et al., 2022 ) , DeepSeek V3 ( Liu et al., 2024 ) , and Gemini-2.5-Flash ( Comanici et al., 2025 ) on CSI 500 and S&P 500, using early stopping of 3 for each logic and selecting the best-performing factor under the current logic in each round. The results show that as iteration rounds progress, the optimal factors corresponding to different rounds show an overall upward trend in key metrics such as IC, IR, cumulative returns, and stability.

Figure 5: The influence of market logic quantity on factor quality. Increasing the number of market logic yields stable and nearly monotonic improvements across IC, ICIR, AR, and IR.

Round	IC		ICIR
	Transient	Persistent	Transient	Persistent
1	0.0182	0.0199	0.1776	0.1792
2	0.0181	0.0208	0.1794	0.1951
3	0.0188	0.0214	0.1767	0.2033
4	0.0169	0.0222	0.1556	0.2071
5	0.0165	0.0232	0.1320	0.2137

Table 3: Ablation study on market logic persistence for verifying the necessity of the Inner Loop design (Algorithm 1 ). The transient setting applies each market logic to guide factor generation only once, whereas the persistent setting reuses the same market logic across multiple rounds to enable deeper exploration.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Factor Library	Institution	number	Core Characteristics
Alpha101	WorldQuant	101	Uses ranking, lag, and cross-sectional normalization; focuses on momentum, reversal, and price–volume interaction.
Alpha191	Guotai Junan	191	Covers momentum, volatility, microstructure, and trading behavior.
Alpha158	Microsoft Research	158	Rich operator set with strong cross-market generalization.
Alpha360	Microsoft Research	360	Largest public technical factor library; encompasses price, volume, volatility, and technical indicators.

Appendix

Table 4: Overview of Major Public Factor Libraries.

Family	Examples	Parameters (range)
Arithmetic	$+$ , $-$ , $\times$ , $/$	–
Cross-sectional	rank, zscore	–
Time-series aggregation	ts_mean, ts_std, ts_min, ts_max, ts_sum	window $w\in\mathbb{N}^{+}$
Time-series change	ts_delta	lag $\ell\in\mathbb{N}^{+}$
Time-series relation	ts_corr, ts_cov	window $w\in\mathbb{N}^{+}$
Smoothing/decay	ts_decay, ts_wma	window $w\in\mathbb{N}^{+}$

Appendix

Table 5: Operator families and parameter ranges used in the factor DSL.

Alpha Factor Discovery

See all 21 papers

AlphaRLM: Aligning Language Models for Literature-Grounded Symbolic Factor Discovery
Oct 10, 2026Jie Huang, Yuecheng Hong, Yao Chen +6Quantitative Finance Algorithmic Trading
Harbin Institute of Technology · Independent Researcher · University of Science and Technology of China
FactorEngine: A Program-level Knowledge-Infused Factor Mining Framework for Quantitative Investment
Mar 17, 2026Qinhong Lin, Ruitao Feng, Yinglun Feng +7Alpha Factor Discovery Quantitative Finance
Beijing University of Posts and Telecommunications, China · Beijing Value Simplex Technology Co., Ltd., China · Yangtze Delta Research Institute, University of Electronic Science and Technology of China, China +1
FactorMiner: A Self-Evolving Agent with Skills and Experience Memory for Financial Alpha Discovery
Feb 16, 2026Yanlong Wang, Jian Xu, Hongkang Zhang +3Alpha Factor Discovery Algorithmic Trading
Tsinghua Shenzhen International Graduate School, Tsinghua University · Pengcheng Laboratory · Shenzhen Finance Institute, Chinese University of Hong Kong

Oct 10, 2026J/K move · Enter open · S save

AlphaRLM: Aligning Language Models for Literature-Grounded Symbolic Factor Discovery

Jie Huang, Yuecheng Hong, Yao Chen, Hao Zhong+5

Harbin Institute of Technology · Independent Researcher · University of Science and Technology of China

Quantitative Finance Algorithmic Trading RL for Language Models Symbolic Regression

Open paper AskPDF arXiv

Formulaic alpha mining seeks to discover executable expressions that capture predictive market signals in a vast, noisy search space. Recent LLM-based approaches use prompting, iterative search, or seed-factor evolution, leaving direct factor construction through policy learning less explored. To address this, we introduce AlphaRLM, a reinforcement-learning framework that trains LLMs to construct executable symbolic factors from financial literature without input factor expressions. AlphaRLM incorporates Literature-Guided Operator Assembly (LEGO), which conditions factor construction on literature-derived ideas and executable operators, providing semantic priors for exploration. It further aligns the generation policy with market feedback through executable-validity and grouped-return-separation rewards that capture non-monotonic structures across factor quantiles. Extensive experiments on CSI 300, CSI 1000, and S&P 500 show that AlphaRLM outperforms strong factor-mining and LLM-based baselines in annualized excess return. Notably, the 8B model achieves strong cross-market performance, highlighting the promise of literature-conditioned, market-aligned factor generation.

Abstract

Alpha Factor Discovery

Explore similar work

AlphaRLM: Aligning Language Models for Literature-Grounded Symbolic Factor Discovery

FactorEngine: A Program-level Knowledge-Infused Factor Mining Framework for Quantitative Investment

FactorMiner: A Self-Evolving Agent with Skills and Experience Memory for Financial Alpha Discovery