Paper: arXiv 2512.23515
Authors: Zuoyou Jiang, Li Zhao, Rui Sun, Ruohan Sun, Zhongjian Li, Jing Li, Daxin Jiang, Zuo Bai, Cheng Hua
Abstract
Signal decay and regime shifts pose recurring challenges for data-driven investment strategies in non-stationary markets. Conventional time-series and machine learning approaches, which rely primarily on historical correlations, often struggle to generalize when the economic environment changes. While large language models (LLMs) offer strong capabilities for processing unstructured information, their potential to support quantitative factor screening through explicit economic reasoning remains underexplored. Existing factor-based methods typically reduce alphas to numerical time series, overlooking the semantic rationale that determines when a factor is economically relevant. We propose Alpha-R1, an 8B-parameter reasoning model trained via reinforcement learning for context-aware alpha screening. Alpha-R1 reasons over factor logic and real-time news to evaluate alpha relevance under changing market conditions, selectively activating or deactivating factors based on contextual consistency. Empirical results across multiple asset pools show that Alpha-R1 consistently outperforms benchmark strategies and exhibits improved robustness to alpha decay. The full implementation and resources are available at https://github.com/FinStep-AI/Alpha-R1.
Complexity vs Empirical Score
- Math Complexity: 6.0/10
- Empirical Rigor: 8.5/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper integrates reinforcement learning and advanced LLM reasoning with a substantial mathematical foundation, including statistical concepts for non-stationary markets and RL optimization. Empirically, it presents extensive backtesting across multiple asset pools, provides a public GitHub repository with implementation, and reports robustness to alpha decay, indicating high practical applicability.
Research Flowchart
flowchart TD A["Research Goal<br>Address Signal Decay & Regime Shifts"] --> B["Data Inputs<br>Market Data, Unstructured News, Factor Series"] B --> C["Methodology<br>Alpha-R1 8B Model with RL Training"] C --> D["Process<br>Reasoning: Factor Logic + News Context"] D --> E["Outcome<br>Context-Aware Alpha Screening<br>(Activate/Deactivate Factors)"] E --> F["Key Findings<br>Outperforms Benchmarks, Robust to Decay"]
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 Alpha-R1, an 8B-parameter reasoning model trained via reinforcement learning for context‑aware alpha screening.
Source sentence
We propose Alpha-R1, an 8B-parameter reasoning model trained via reinforcement learning for context-aware alpha screening.
Assumes 6
- Execution assumes a transaction fee of 0.1% per trade to account for commissions and slippage. execution
Source sentence
A transaction fee of 0.1% (10 bps) is applied to both buy and sell orders to account for commissions and slippage.
- Slot rotation assumes a holding period H (e.g., 5 days) and rebalances only one slot per day, reducing turnover to 1/H. execution
Source sentence
independent sub‑portfolios (referred to as slots), where
- VWAP execution assumes price is computed from the first 30 minutes of the trading session as a conservative estimate of slippage. execution
Source sentence
is computed using transaction data from the first 30 minutes of the trading session (09:31–10:00).
- The factor backtesting assumes that the historical performance vector Pi (returns, volatility, decay) provides the ground‑truth for factor effectiveness. data
Source sentence
which includes key metrics such as returns, volatility, and decay characteristics.
Show 2 more
- The dynamic factor zoo assumes the Alpha101 library provides 82 computationally feasible factors after filtering. data
Source sentence
We construct a dynamic factor zoo by filtering the Alpha101 library [18], retaining 82 computationally feasible factors.
- Training assumes the CSI 300 constituent stocks represent a high‑turnover environment for the training phase. market
Source sentence
Using constituent stocks from the CSI 300, we simulate a high‑turnover environment.
Data
2025-01 – 2025-06 · CSI 300 stocks · equities · daily
- Alpha101 library other unknown
Source sentence
We construct a dynamic factor zoo by filtering the Alpha101 library [18], retaining 82 computationally feasible factors.
- CSI 300 constituents market-prices unknown
Source sentence
Using constituent stocks from the CSI 300, we simulate a high‑turnover environment.
- CSI 1000 universe market-prices unknown
Source sentence
The CSI 1000 experiment (Figure 2b) evaluates the model’s ability to generalize under a zero‑shot setting.
Costs
Modeled
A flat transaction fee of 0.1% (10 bps) per side is applied to both buy and sell orders.
Source sentence
A transaction fee of 0.1% (10 bps) is applied to both buy and sell orders to account for commissions and slippage.
Finds 5
Sharpe 4.03 · annual return 78.2% · total return 42.5% · max drawdown 6.8%
- Alpha‑R1 achieves a Sharpe ratio of 1.62 and limits maximum drawdown to 6.76% on the CSI 300, outperforming traditional baselines. performance
Source sentence
Alpha‑R1 attains a Sharpe ratio of 1.62 while limiting maximum drawdown to 6.76%.
- On the out‑of‑domain CSI 1000 universe, Alpha‑R1 obtains a cumulative return of 42.49% and a Sharpe ratio of 4.03, demonstrating strong zero‑shot generalization. performance
Source sentence
Alpha‑R1 maintains strong performance on the out‑of‑domain CSI 1000 universe, achieving a cumulative return of 42.49% and a Sharpe ratio of 4.03.
- Baseline reinforcement‑learning agents (A2C, PPO) suffer negative or marginal returns and low Sharpe ratios in both in‑domain and out‑of‑domain tests. negative
Source sentence
A2C achieves a negative cumulative return of -5.52% with a Sharpe ratio of -0.85, while PPO shows only marginal positive returns (0.89%) with a low Sharpe ratio of 0.11.
- Generic reasoning LLMs (Claude 3.7 Sonnet, DeepSeek‑R1) exhibit substantial drawdowns (~15–16%) and negative Sharpe ratios, indicating lack of domain‑specific grounding. negative
Source sentence
They fail to adequately account for risk constraints, leading to substantial drawdowns (approximately 15–16%) and negative Sharpe ratios.
- Ablation studies confirm that semantic gating and the integration of news, price, and semantic descriptions are key contributors to Alpha‑R1’s superior performance. performance
Source sentence
tangible advantages over traditional heuristic gating strategies.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 30 related papers). Coverage is partial, so this is neither replication nor refutation.
1 paper in the archive
Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.
Cited References (90)
- Reinforcement-Learning Portfolio Allocation with Dynamic Embedding of Market Information — → 2501.17992
- 101 formulaic alphas (2016) — link
- DeepSeekMath : Pushing the Limits of Mathematical Reasoning in Open Language Models (2024) — link
- Common risk factors in the returns on stocks and bonds (1993) — link
- Capital Asset Prices: A Theory of Market Equilibrium under Conditions of Risk (1964) — link
- On Persistence in Mutual Fund Performance (1997) — link
- A Five-Factor Asset Pricing Model (2015) — link
- Taming the Factor Zoo: A Test of New Factors (2020) — link
and 82 more
- Can ChatGPT Forecast Stock Price Movements? Return Predictability and Large Language Models — → 2304.07619
- Dissecting Characteristics Nonparametrically (2020) — link
- Quantum Finance and Fuzzy Reinforcement Learning-Based Multi-Agent Trading System (2024) — link
- Alpha-GPT 2.0: Human-in-the-Loop AI for Quantitative Investment — → 2402.09746
- R&D-Agent-Quant : A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization — → 2505.15155
- Automate Strategy Finding with LLM in Quant Investment — → 2409.06289
- 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
- AlphaEval : A Comprehensive and Efficient Evaluation Framework for Formula Alpha Mining — → 2508.13174
- Fine-Tuning Large Language Models for Stock Return Prediction Using Newsflow — → 2407.18103
- From Deep Learning to LLMs : A Survey of AI in Quantitative Investment — → 2503.21422
- Alpha- GPT : Human- AI Interactive Alpha Mining for Quantitative Investment — → 2308.00016
- Regression Shrinkage and Selection via the Lasso (1996) — link
- Factor Investment or Feature Selection Analysis? (2025) — link
- Empirical Asset Pricing via Machine Learning (2020) — link
- Factor Models, Machine Learning, and Asset Pricing (2022) — link
- A Survey of Reinforcement Learning from Human Feedback (2025) — link
- Training language models to follow instructions with human feedback (2022) — link
- Proximal Policy Optimization Algorithms (2017) — link
- DeepSeek - R1 incentivizes reasoning in LLMs through reinforcement learning (2025) — link
- BloombergGPT : A Large Language Model for Finance (2023) — link
- Fin GPT : Democratizing Internet-scale Data for Financial Large Language Models (2023) — link
- Instruct-FinGPT : Financial Sentiment Analysis by Instruction Tuning of General-Purpose Large Language Models — → 2306.12659
- Lo RA : Low-Rank Adaptation of Large Language Models (2022) — link
- PIXIU : A Comprehensive Benchmark, Instruction Dataset and Large Language Model for Finance (2023) — link
- BBT-Fin : Comprehensive Construction of Chinese Financial Domain Pre-trained Language Model, Corpus and Benchmark (2023) — link
- XuanYuan 2.0 : A Large Chinese Financial Chat Model with Hundreds of Billions Parameters (2023) — link
- Trading-R1 : Financial Trading with LLM Reasoning via Reinforcement Learning — → 2509.11420
- Fin-R1 : A Large Language Model for Financial Reasoning through Reinforcement Learning (2025) — link
- FinO1 : On the Transferability of Reasoning-Enhanced LLMs and Reinforcement Learning to Finance (2025) — link
- TradingGPT : Multi-Agent System with Layered Memory and Distinct Characters for Enhanced Financial Trading Performance — → 2309.03736
- FinMem : A Performance-Enhanced LLM Trading Agent with Layered Memory and Character Design — → 2311.13743
- TradingAgents : Multi-Agents LLM Financial Trading Framework — → 2412.20138
- MetaGPT : Meta Programming for A Multi-Agent Collaborative Framework (2024) — link
- A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist — → 2402.18485
- Learning to Generate Explainable Stock Predictions using Self-Reflective Large Language Models — → 2402.03659
- LLMF actor: Extracting Profitable Factors through Prompts for Explainable Stock Movement Prediction (2024) — link
- Direct Preference Optimization: Your Language Model is Secretly a Reward Model (2023) — link
- SimPO : Simple Preference Optimization with a Reference-Free Reward (2024) — link
- Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback (2023) — link
- R eward B ench: Evaluating Reward Models for Language Modeling (2025) — link
- Measuring Massive Multitask Language Understanding (2021) — link
- Towards Mitigating LLM Hallucination via Self Reflection (2023) — link
- Beyond the Black Box: Interpretability of LLMs in Finance — → 2505.24650
- FinRL : A Deep Reinforcement Learning Library for Automated Stock Trading in Quantitative Finance (2020) — link
- Continuous control with deep reinforcement learning (2016) — link
- Addressing Function Approximation Error in Actor-Critic Methods (2018)
- Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor (2018)
- Stable-Baselines3 : Reliable Reinforcement Learning Implementations (2021)
- MAPS : Multi-agent Reinforcement Learning-based Portfolio Management System (2020) — link
- Reinforcement-Learning Based Portfolio Management with Augmented Asset Movement Prediction States (2020) — link
- Chain-of-Alpha : Unleashing the Power of Large Language Models for Alpha Mining in Quantitative Trading (2025) — link
- Returns to Buying Winners and Selling Losers: Implications for Stock Market Efficiency (1993)
- Evidence of Predictable Behavior of Security Returns (1990)
- The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting and Non-Normality (2014)
- The Probability of Backtest Overfitting (2017)
- The Sharpe Ratio Efficient Frontier (2012)
- The Stationary Bootstrap (1994)
- A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix (1987)
- The Probabilistic Relevance Framework: BM25 and Beyond (2009)
- AlphaAgentEvo : Evolution-Oriented Alpha Mining via Self-Evolving Agentic Reinforcement Learning (2026) — link
- Navigating the Alpha Jungle: An LLM -Powered MCTS Framework for Formulaic Alpha Factor Mining (2026) — link
- QuantaAlpha : An Evolutionary Framework for LLM -Driven Alpha Mining — → 2602.07085
- FactorMiner : A Self-Evolving Agent with Skills and Experience Memory for Financial Alpha Discovery — → 2602.14670
- AlphaCrafter : Harnessing Multi-Agent Workflows for Cross-Sectional Quantitative Trading (2026) — link
- From Feedback Loops to Policy Updates: Reinforcement Fine-Tuning for LLM -Based Alpha Factor Discovery (2026) — link
- Can LLM -based Financial Investing Strategies Outperform the Market in Long Run? — → 2505.07078
- Time Travel is Cheating: Going Live with DeepFund for Real-Time Fund Investment Benchmarking (2025) — link
- QuantEvolve : Automating Quantitative Strategy Discovery through Multi-Agent Evolutionary Framework (2025) — link
- AlphaAgents : Large Language Model based Multi-Agents for Equity Portfolio Constructions — → 2508.11152
- AlphaQuanter : An End-to-End Tool-Augmented Agentic Reinforcement Learning Framework for Stock Trading (2026) — link
- Trade- R1 : Bridging Verifiable Rewards to Stochastic Environments via Process-Level Reasoning Verification — → 2601.03948
- Reasoning through Verifiable Forecast Actions: Consistency-Grounded RL for Financial LLMs (2026) — link
- FinWorld : An All-in-One Open-Source Platform for End-to-End Financial AI Research and Deployment (2026) — link
- Understanding R1 -Zero-Like Training: A Critical Perspective (2025) — link
- DAPO : An Open-Source LLM Reinforcement Learning System at Scale (2025) — link
- Group Sequence Policy Optimization (2025) — link
- The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models (2025) — link
- Forest through the Trees: Building Cross-Sections of Stock Returns (2025) — link
- Growing the Efficient Frontier on Panel Trees (2025) — link
- APT or `` AIPT ‘’? The Surprising Dominance of Large Factor Models (2024) — link
- Structural Deep Learning in Conditional Asset Pricing (2022) — link
23 of 90 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.