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

Yanlong Wang; Jian Xu; Hongkang Zhang; Shao-Lun Huang; Danny Dongning Sun; Xiao-Ping Zhang

q-fin.TRFeb 16, 2026

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

Authors: Yanlong Wang, Jian Xu, Hongkang Zhang, Shao-Lun Huang, Danny Dongning Sun, Xiao-Ping Zhang

Organizations: Tsinghua Shenzhen International Graduate School, Tsinghua University · Pengcheng Laboratory · Shenzhen Finance Institute, Chinese University of Hong Kong

Abstract

Formulaic alpha factor mining is a critical yet challenging task in quantitative investment, characterized by a vast search space and the need for domain-informed, interpretable signals. However, finding novel signals becomes increasingly difficult as the library grows due to high redundancy. We propose FactorMiner, a lightweight and flexible self-evolving agent framework designed to navigate this complex landscape through continuous knowledge accumulation. FactorMiner combines a Modular Skill Architecture that encapsulates systematic financial evaluation into executable tools with a structured Experience Memory that distills historical mining trials into actionable insights (successful patterns and failure constraints). By instantiating the Ralph Loop paradigm -- retrieve, generate, evaluate, and distill -- FactorMiner iteratively uses memory priors to guide exploration, reducing redundant search while focusing on promising directions. Experiments on multiple datasets across different assets and Markets show that FactorMiner constructs a diverse library of high-quality factors with competitive performance, while maintaining low redundancy among factors as the library scales. Overall, FactorMiner provides a practical approach to scalable discovery of interpretable formulaic alpha factors under the "Correlation Red Sea" constraint.

Figures & tables

Figure 1. FactorMiner System Architecture. The Ralph Loop framework integrates three key components: (1) Experience Memory that stores successful patterns and forbidden regions from past mining sessions; (2) Agent Skill that encapsulates the multi-stage validation pipeline (IC screening, correlation checking, deduplication, and full validation); (3) Factor Library that grows dynamically while maintaining orthogonality constraints. The agent iteratively retrieves memory priors, generates candidates through the skill, and distills outcomes back into memory for improved future exploration.

Figure 2. Pairwise Spearman correlation heatmap of the released A-share factor library (110 admitted factors), computed from cross-sectionally standardized realized factor signals over the common time–asset panel. The average off-diagonal absolute correlation is Avg $|\rho|$ = 0.203.

Dataset	Method	Factor Library (Top-40)			Factor Combination (Top-40)				Factor Selection (Train’24/Test’25)
Dataset	Method	IC (%)	ICIR	Avg $\|\rho\|$	EW IC	EW ICIR	ICW IC	ICW ICIR	Las. IC	Las. ICIR	XGB. IC	XGB. ICIR
CSI500	RF ‡	2.68	0.25	0.13	6.98	0.41	12.02	0.90	13.81	1.15	8.99	0.90
	Alpha101 (Classic)	4.49	0.42	0.19	10.85	0.88	12.89	0.99	14.09	1.27	12.07	1.21
	Alpha101 (Adapted)	5.06	0.43	0.21	11.53	0.86	14.71	1.13	13.70	1.08	13.76	1.20
	GPLearn	6.04	0.43	0.44	10.30	0.62	13.38	1.00	12.44	1.17	9.86	0.95
	AlphaForge	4.48	0.38	0.36	7.12	0.60	11.13	0.93	10.49	0.87	11.30	1.25

Table 1. Out-of-Sample performance comparison with a stricter protocol. Top-40 factors are selected on CSI500 ( 2024 ) and evaluated on 2025 across datasets. For Alpha101, Classic is restricted to the same candidate set as Adapted. All reported IC and ICIR use the paper’s absolute-IC summary:

|\mathbb{E}[IC_{t}]|

and

|\mathbb{E}[IC_{t}]|/\mathrm{std}(IC_{t})

. Factor Combination uses the frozen Top-40 with EW and ICW (weights/signs determined on 2024). Factor Selection trains on 2024 and tests on 2025 using Lasso and XGBoost.

Metric	Period	CSI500	CSI1000	CSI300	Crypto
CSD (‰)	2024H1	5.58	5.00	3.50	2.74
	2024H2	5.05	10.56	6.44	2.64
	2025H1	3.87	4.73	3.09	2.39
	2025H2	4.26	4.41	3.64	2.48
	Avg	4.69	6.18	4.17	2.56
SYN	2024H1	-1.04	-0.80	-1.41	0.47

Table 2. Market heterogeneity measured by cross-sectional dispersion (CSD, ‰) and return synchronicity (SYN) over half-year windows.

Market	PatchTST		Chronos-2		LightGBM		FactorMiner
Market	IC (%)	ICIR	IC (%)	ICIR	IC (%)	ICIR	IC (%)	ICIR
CSI500	2.21	0.19	4.52	0.37	5.53	0.51	8.25	0.77
CSI1000	1.84	0.20	3.36	0.26	4.85	0.45	7.78	0.76
CSI300	1.80	0.15	3.01	0.19	4.66	0.34	7.46	0.38
Crypto	1.58	0.06	2.62	0.16	2.88	0.13	3.82	0.28

Table 3. Comparison with end-to-end forecasting baselines on the 2025 test period. IC is in %; higher is better.

Period	Lib. IC (%)	Lib. ICIR	XGB IC (%)	XGB ICIR
2025-10	5.52	0.49	13.42	1.16
2025-11	6.74	0.65	15.83	1.46
2025-12	6.83	0.68	16.16	1.55
2026-01	4.78	0.46	12.56	1.03
2026-02	6.23	0.64	15.23	1.38
2026-03	5.52	0.51	14.96	1.17

Table 4. Recent-period robustness of the full factor library on CSI500.

Rank	ID	Factor	Imp.	Category
1	006	VWAP Deviation	6.04	VWAP
2	061	Alpha101-12 Modern	4.06	Classic
3	023	Norm.-Momentum TsRank	3.59	Momentum
4	068	Skewness Regime PV-Div	3.55	Regime
5	028	Close-Low $\times$ Volume	3.03	Price range
6	070	Price-Pos. Vol. Interact.	2.27	Interaction

Table 5. XGBoost feature importance on the full FactorMiner library (top 20 factors).

Pattern	Search prior	Yield
Higher moments	Use Skew/Kurt as regime conditions for reversal logic.	High
PV corr. interaction	Combine price-volume correlation with amount or trend operators.	High
Robust efficiency	Smooth return/amount efficiency using median or robust statistics.	High
Smoothed efficiency rank	Apply EMA/WMA before cross-sectional ranking.	High
Trend regression	Switch between Slope and Resi using Rsquare reliability.	High
Extreme regimes	Use logical Or over price/volume extremes as state triggers.	High

Table 6. Recommended mining directions distilled from experience memory.

Figure 3. Ablation comparison between Have Memory and No Memory. High-quality candidates who pass the IC threshold. The bar chart reports the counts and the corresponding yield and rejection rates.

Figure 4. Operator and factor level computation time (ms, log scale) under Python, optimized CPU (C), and GPU backends; lower is better. The GPU backend yields consistent order-of-magnitude gains.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Representative operators
Arithmetic	Add, Sub, Mul, Div, Neg, Abs, Log, SignedPower, Power, Inv, Sqrt, Square, Exp, Tanh
Statistical	Mean, Std, Var, Skew, Kurt, Med, Sum, Product
Time-series	Delay, Delta, TsRank, TsMax, TsMin, TsArgMax, TsArgMin, TsDecay
Cross-sectional	CsRank, Scale
Smoothing	SMA, EMA, WMA
Regression	Slope, Rsquare, Resi

Appendix

Table 7. Operator categories in the factor mining skill (representative operators; 60+ in total).

Model	Avg $\|\rho\|$	Max $\|\rho\|$	Nearest factor	2nd	3rd
PatchTST	0.0365	0.18	002 EMA Deviation	108	006
Chronos-2	0.0399	0.23	062 Amount-stability adj. returns	002	081
LightGBM	0.0146	0.08	062 Amount-stability adj. returns	006	084

Appendix

Table 8. Correlation between end-to-end model signals and the 110-factor library.

Model	Chronos-2	LightGBM	PatchTST
Chronos-2	1.000	0.043	0.171
LightGBM	0.043	1.000	0.224
PatchTST	0.171	0.224	1.000

Appendix

Table 9. Pairwise correlations among end-to-end signals.

Direction	Correlated factors	$\|\rho\|$
Standardized Returns/Amount	006, 008, 009	$>$ 0.6
VWAP Deviation variants	006, 009, 012, 013, 016	$>$ 0.5
Mean Reversion / Price Deviation	001, 002	$>$ 0.5
Simple Delta Reversal	023, 024	$>$ 0.5
Close-Position Location	028, 044	0.87+
Volatility + Price Position Branch	046, 064	$>$ 0.6

Appendix

Table 10. Forbidden mining directions (high correlation risk).

Alpha Factor Discovery

See all 21 papers

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
From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery
May 14, 2026Lingzhe Zhang, Tong Jia, Yunpeng Zhai +5RL for Language Models Alpha Factor Discovery
Peking University, Beijing, China · Alibaba Group, China · Nanjing University, Nanjing, China +1
AlphaPilot: Multi-Turn Tool-Using LLM Agents for Adaptive Alpha Mining with Reinforcement Learning
Oct 11, 2026Peiqi Jiang, Fan Zhang, Xin Xu +2Alpha Factor Discovery Quantitative Finance
Department of Mathematics, Hong Kong University of Science and Technology

Mar 17, 2026J/K move · Enter open · S save

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

Qinhong Lin, Ruitao Feng, Yinglun Feng, Zhenxin Huang+6

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

Alpha Factor Discovery Quantitative Finance Automated Feature Engineering Algorithmic Trading Program Synthesis

Open paper AskPDF arXiv

We study alpha factor mining, the automated discovery of predictive signals from noisy, non-stationary market data-under a practical requirement that mined factors be directly executable and auditable, and that the discovery process remain computationally tractable at scale. Existing symbolic approaches are limited by bounded expressiveness, while neural forecasters often trade interpretability for performance and remain vulnerable to regime shifts and overfitting. We introduce FactorEngine (FE), a program-level factor discovery framework that casts factors as Turing-complete code and improves both effectiveness and efficiency via three separations: (i) logic revision vs. parameter optimization, (ii) LLM-guided directional search vs. Bayesian hyperparameter search, and (iii) LLM usage vs. local computation. FE further incorporates a knowledge-infused bootstrapping module that transforms unstructured financial reports into executable factor programs through a closed-loop multi-agent extraction-verification-code-generation pipeline, and an experience knowledge base that supports trajectory-aware refinement (including learning from failures). Across extensive backtests on real-world OHLCV data, FE produces factors with substantially stronger predictive stability and portfolio impact-for example, higher IC/ICIR (and Rank IC/ICIR) and improved AR/Sharpe, than baseline methods, achieving state-of-the-art predictive and portfolio performance.

Abstract

Alpha Factor Discovery

Explore similar work

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

From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM-Based Alpha Factor Discovery

AlphaPilot: Multi-Turn Tool-Using LLM Agents for Adaptive Alpha Mining with Reinforcement Learning