Paper: arXiv 2611.25732
Authors: Hui Cai, Kangcheng Deng, Tianyi Ma, Qixuan Yang, Jiacheng Lu, Rui Sun, Beidi Luan, Jing Li, Zuo Bai
Abstract
Autonomous factor research requires generating economic hypotheses as well as implementing and testing them. We present Auto-Alpha, a framework that connects report-guided initialization to self-directed research through a test-driven implementation loop. An idea agent proposes economic mechanisms and signal constructions; a research agent develops them with backtest feedback. Semantic audit, quantitative review, and human admission select factors for a persistent library, while an Experience Library retains research insights. Starting from 120 research reports, Auto-Alpha builds an initial library of 32 factors. It then generates new ideas and refines existing candidates, producing 420 implementations and retaining 62 additional factors. We evaluate factor quality and portfolio performance using a fixed library snapshot of 105 factors. The median 1-day Rank IC is 0.0301, and a weekly Top-300 portfolio returns 16.81% annualized after stated costs over 2020-2026, versus 1.71% for the CSI 1000 price index. These retrospective results document an operating research process and its artifacts; controlled evaluation is needed to isolate the contribution of its components.
Complexity vs Empirical Score
- Math Complexity: 5.0/10
- Empirical Rigor: 6.0/10 (capped from 7.0: no out-of-sample test; see penalty flags)
- Quadrant: Holy Grail — high math complexity, high empirical rigor (effective empirical rigor capped from 7.0)
Why this score: The paper presents a novel framework for autonomous factor research, combining LLM agents with a test-driven loop. It demonstrates strong empirical results with a substantial factor library and portfolio performance, though it acknowledges the need for controlled evaluation of component contributions. The methodology is well-described, but full reproducibility would benefit from explicit code/data availability.
Research Flowchart
flowchart TD
A["Research Goal: Autonomous Factor Research"] --> B{"Report-Guided Initialization"};
B -- 120 Research Reports --> C["Idea Agent: Propose Economic Hypotheses & Signals"];
C --> D{"Research Agent: Develop & Test Signals via Backtest Feedback"};
D -- Test-Driven Implementation Loop --> E["Factor Selection: Semantic Audit, Quantitative Review, Human Admission"];
E -- Persistent Factor Library & Experience Library --> F["Key Findings: 105 Factors, Median 1-day Rank IC 0.0301, 16.81% Annualized Return"];
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 Auto-Alpha, a framework that connects report‑guided initialization to self‑directed factor research through a test‑driven implementation loop.
Source sentence
We present Auto-Alpha, a framework that connects report-guided initialization to self-directed research through a test-driven implementation loop.
Assumes 6
- Trades are executed using VWAP and the account is marked at the close. execution
Source sentence
Trades use VWAP and the account is marked at the close.
- The execution model does not estimate market impact or impose a volume‑participation cap. execution
Source sentence
This execution model does not estimate market impact or impose a volume-participation cap.
- Constant transaction fees of 1 bp for buys and 6 bp for sells (plus minimum RMB 5 and 10 bp slippage per side) are applied throughout the period. model
Source sentence
Constant fees throughout the period are 1 basis point for buys and 6 for sells, with a RMB 5 minimum per stock and direction; proportional slippage adds 10 basis points on each side.
- The eligible universe excludes ST stocks and stocks listed for fewer than 60 days. data
Source sentence
The eligible Shanghai and Shenzhen A-share universe excludes ST stocks and stocks listed for fewer than 60 days; actual fills respect the engine’s suspension, price-limit, and approximate lot-size rules.
Show 2 more
- No additional turnover or stock‑replacement limit is imposed. execution
Source sentence
No additional turnover or stock-replacement limit is imposed.
- Weekly rebalancing uses the first trading day of each week to generate a Top‑300 target executed the next day. execution
Source sentence
Each week’s first trading day produces a new Top-300 target for execution on the next trading day.
Data
2020-01 – 2026-11 · Shanghai and Shenzhen A-share universe · equities · weekly
- Shanghai and Shenzhen A‑share price data market-prices unknown
Source sentence
The account runs from the first trading day of 2020 through 2026-11-03. The eligible Shanghai and Shenzhen A-share universe excludes ST stocks and stocks listed for fewer than 60 days; actual fills respect the engine’s suspension, price-limit, and approximate lot-size rules.
- CSI 1000 price index benchmark unknown
Source sentence
CSI 1000 (000852.SH) is the reference benchmark, with no index hedge or constraint restricting selection to its constituents. The net portfolio returns 16.81% annualized, compared with 1.71% for the CSI 1000 price index.
Costs
Modeled
Constant fees of 1 bp for buys and 6 bp for sells (minimum RMB 5 per stock) plus proportional slippage of 10 bp on each side.
Source sentence
Constant fees throughout the period are 1 basis point for buys and 6 for sells, with a RMB 5 minimum per stock and direction; proportional slippage adds 10 basis points on each side.
Finds 4
Sharpe 0.76 · annual return 16.8% · max drawdown 32.2%
- The median 1‑day Rank IC of the 105‑factor library is 0.0301. predictive-accuracy figure from elsewhere in the paper
Source sentence
The median 1-day Rank IC is 0.0301, and a weekly Top-300 portfolio returns 16.81% annualized after stated costs over 2020-2026, versus 1.71% for the CSI 1000 price index. Its median 1-day Rank IC is 0.0301.
- A weekly Top‑300 portfolio constructed from the library returns 16.81% annualized after stated costs over 2020‑2026, versus 1.71% for the CSI 1000 benchmark. performance
Source sentence
The median 1-day Rank IC is 0.0301, and a weekly Top-300 portfolio returns 16.81% annualized after stated costs over 2020-2026, versus 1.71% for the CSI 1000 price index. The net portfolio returns 16.81% annualized, compared with 1.71% for the CSI 1000 price index.
- The portfolio’s Sharpe ratio is 0.763 and its maximum drawdown is 32.16%. performance
Source sentence
Its Sharpe ratio is 0.763, with a maximum drawdown of 32.16%.
- Annualized one‑way turnover of the portfolio is 19.15 times account value. performance
Source sentence
Annualized one-way turnover is 19.15 times account value.
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.
Code
No repository by the authors found in the paper.
Extracted 2026-11-20 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (24)
- From Search to Research: Exploring Search Scaling in Autonomous Quantitative Factor Mining (2026) — link
- X-Alpha : An AI Quant Researcher with Memory-Driven Self-Evolution (2026) — link
- Cognitive Alpha Mining via LLM -Driven Code-Based Evolution (2026) — link
- R&D-Agent-Quant : A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization — → 2505.15155
- AlphaAgent : LLM -Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay (2025) — link
- AutoAlpha : an Efficient Hierarchical Evolutionary Algorithm for Mining Alpha Factors in Quantitative Investment (2020) — link
- AlphaForge : A Framework to Mine and Dynamically Combine Formulaic Alpha Factors — → 2406.18394
- Alpha-GPT 2.0 : Human-in-the-Loop AI for Quantitative Investment — → 2402.09746
and 16 more
- FactorMiner : A Self-Evolving Agent with Skills and Experience Memory for Financial Alpha Discovery — → 2602.14670
- Reflexion : Language Agents with Verbal Reinforcement Learning (2023) — link
- The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery (2024) — link
- Large Language Monkeys: Scaling Inference Compute with Repeated Sampling (2024) — link
- Scaling Test-time Compute for LLM Agents (2025) — link
- ResearchGym : Evaluating Language Model Agents on Real-World AI Research (2026) — link
- PostTrainBench : Can LLM Agents Automate LLM Post-Training? (2026) — link
- PaperBench : Evaluating AI ’s Ability to Replicate AI Research (2025) — link
- OpenFinGym : A Verifiable Multi-Task Gym Environment for Evaluating Quant Agents (2026) — link
- Generating Synergistic Formulaic Alpha Collections via Reinforcement Learning — → 2306.12964
- ReAct : Synergizing Reasoning and Acting in Language Models (2023) — link
- Scaling LLM Test-Time Compute Optimally Can be More Effective than Scaling Parameters for Reasoning (2025) — link
- AlphaBench : Benchmarking Large Language Models in Formulaic Alpha Factor Mining (2026) — link
- AlphaQT-Bench : Diagnosing the Gap between Financial Code Generation and Quantitative Reasoning in LLMs (2026) — link
- RE-Bench : Evaluating Frontier AI R&D Capabilities of Language Model Agents against Human Experts (2025) — link
- The Probability of Backtest Overfitting (2017) — link
5 of 24 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2601 linked papers.