Beyond Forecasting: The Belief-to-Trade Layer in Prediction-Market Agents
Proposes Raven-Agent, a modular trading layer for prediction markets, achieving +15.9% risk-adjusted return in replay tests.
Key Findings
Methodology
The system features four modules: information collection, probability estimation, candidate ranking and sizing, and execution with risk constraints. It employs a time-normalized return score (Eq. 2) for candidate ranking, and Kelly fraction (Eq. 3) for position sizing. Hard risk limits include stop-loss, exposure caps, and drawdown controls, ensuring robustness. The architecture is designed to be interchangeable with any forecaster, validated through fixed archive replay. The approach emphasizes explicit, deterministic, and modular design for better interpretability and control.
Key Results
- In replay experiments, Raven-Agent achieved a +15.9% risk-adjusted return, outperforming baseline policies such as fixed stake and edge-proportional strategies, which suffered losses of −10.7% and −55.5%. In live deployment, it accumulated $53.92 profit over 20 predictions, demonstrating practical viability. Filtering and risk controls significantly improved outcomes, while size-only strategies underperformed.
Significance
This work addresses the gap between probabilistic forecasting and profitable trading by introducing a dedicated trading layer that is both modular and risk-aware. It advances the development of autonomous AI systems capable of robust financial decision-making, bridging the gap between theoretical probability calibration and real-world profitability. The explicit separation of prediction and trading modules enhances interpretability, safety, and adaptability, making it a promising step toward practical AI-driven financial markets.
Technical Contribution
The core technical innovation lies in designing a deterministic, composable trading layer that can interface with any forecaster. The use of time-normalized return scores for candidate selection, Kelly-based sizing, and external risk constraints distinguishes this approach from prior end-to-end models. The architecture ensures risk remains manageable even under model errors, providing a foundation for reliable autonomous trading systems. The modular design facilitates future upgrades and integration with advanced forecasting models.
Novelty
This is the first work to treat the trading layer as an explicit, adjustable module separate from the forecaster, enabling independent optimization. It combines a ranking mechanism based on normalized returns with Kelly sizing and strict risk controls, contrasting with previous approaches that hardcoded trading rules or relied solely on model outputs. The approach emphasizes transparency, flexibility, and risk management, setting a new standard for autonomous prediction-market agents.
Limitations
- The current system relies on fixed archived forecasts, limiting adaptability to real-time market dynamics. Its performance in highly volatile or unforeseen scenarios remains untested. Manual tuning of risk parameters is required, lacking adaptive mechanisms. Further validation in live, continuously updating environments is necessary to assess robustness and scalability.
Future Work
Future research will focus on integrating real-time data streams, enabling adaptive learning and parameter tuning. Combining reinforcement learning with the current modular architecture could improve long-term profitability. Extending the system to multi-asset and multi-market settings, along with enhanced risk management strategies, will be key steps toward deploying fully autonomous AI trading agents in real-world financial markets.
AI Executive Summary
Predicting future events through probabilistic models has become a vital benchmark for general AI capabilities. However, translating accurate probability estimates into profitable trading remains a challenge. Traditional approaches often treat the trading process as a fixed protocol or embed it within the forecasting model, limiting flexibility and interpretability. This study introduces Raven-Agent, a novel architecture that explicitly separates the trading layer from the forecaster, making it modular, deterministic, and easily adjustable.
The core innovation lies in designing a ranking and sizing mechanism based on time-normalized expected returns (Eq. 2) and Kelly proportional bet sizing (Eq. 3). These components enable the system to select promising candidates, allocate capital efficiently, and enforce strict risk constraints such as stop-loss, exposure caps, and drawdown limits. The architecture’s modularity allows it to interface with any forecasting model, facilitating independent improvements.
Experimental validation on a fixed archive of prediction market decisions demonstrates that Raven-Agent achieves a +15.9% risk-adjusted return, the only policy with positive performance in the replay. In live deployment on Polymarket, it accumulated over $53 in profit, confirming its practical potential. The results highlight the importance of explicit, risk-aware trading layers that can filter, size, and control risk outside the forecasting process.
Looking ahead, future work aims to incorporate real-time data, adaptive learning, and multi-asset strategies to enhance robustness and profitability. This approach paves the way for autonomous AI systems capable of reliable, interpretable, and profitable trading in complex financial environments, marking a significant step toward AI-driven market participation.
Deep Analysis
Background
The evolution of AI in prediction markets has transitioned from static probability calibration to dynamic trading strategies. Early benchmarks like ForecastBench and Prophet Arena primarily evaluated forecast quality, such as Brier scores and calibration metrics, with limited focus on trading outcomes. Recent advances introduced live trading benchmarks, revealing a persistent gap: models with high probability calibration often fail to generate profits. This discrepancy stems from the lack of explicit, flexible trading strategies and risk management mechanisms. Researchers have recognized that effective trading requires not only accurate forecasts but also robust decision-making layers that can adapt to market conditions, size positions appropriately, and enforce risk limits. Despite progress, most existing systems either hardcode trading rules or rely on end-to-end learning, which hampers interpretability and control. This paper addresses these limitations by proposing a modular, deterministic trading layer that can be combined with any forecaster, aiming to improve real-world trading performance.
Core Problem
While probabilistic forecasting models have achieved high calibration scores, their translation into profitable trades remains elusive. The core challenge lies in designing a trading strategy that effectively leverages forecast probabilities while managing risk. Existing methods often lack explicit risk controls, leading to large losses during adverse market movements. Fixed trading rules or end-to-end models fail to adapt to market variability, resulting in subpar performance. The problem is further complicated by the need for a transparent, interpretable system that can be audited and adjusted independently of the forecasting component. Addressing these issues requires a novel architecture that separates prediction from trading, incorporates risk constraints, and optimizes position sizing based on expected returns, all while maintaining modularity and flexibility.
Innovation
This work introduces a modular trading layer that is explicitly designed to be interchangeable and deterministic. Key innovations include: 1) a ranking module based on time-normalized return scores (Eq. 2) that prioritizes candidates considering both probability edge and time to resolution; 2) Kelly-based position sizing (Eq. 3) that adjusts stakes proportionally to expected returns, scaled down for robustness; 3) a set of external, hard risk constraints—such as maximum stake, exposure, stop-loss, and drawdown—that operate outside the forecasting model to ensure safety; 4) a replay protocol that isolates the trading layer from the forecaster, enabling controlled evaluation. These innovations collectively enable a flexible, interpretable, and risk-aware trading system that can be paired with any probabilistic forecast, significantly advancing autonomous prediction-market trading.
Methodology
- �� 信息采集：通过Polymarket Gamma API获取市场价格、流动性、预期到期时间等信息。• 概率分析：利用结构化信念状态和外部语言模型（如GPT）结合市场隐含信息，估算每个候选的支付概率pi。• 排序与规模调整：根据时间归一化收益率（公式2）对候选排序，筛除低于阈值的项目，采用Kelly比例（公式3）调整仓位，确保风险与收益平衡。• 执行与风险控制：硬性限制单笔交易规模、总持仓、回撤，确保风险在可控范围内。• 交易执行：提交订单至Polymarket CLOB，实时监控仓位变化，动态调整策略。• 反馈与优化：通过历史回放验证策略效果，调整参数以优化收益。
Experiments
采用Polymarket历史档案，筛选出59个决策点，验证不同策略表现。对比包括纯预测交易、边缘比例、边缘过滤、固定仓位Raven-Agent和全策略Raven-Agent。指标包括ROI、Sharpe比率、Brier分数和累计利润。实验显示，Raven-Agent（全策略）实现15.9%的正收益，优于其他策略。还进行了鲁棒性检验，包括bootstrap置信区间和逐一删除分析，验证结果稳健。真实部署中，累计盈利53.92美元，表现出良好的实用性。
Results
Raven-Agent（全策略）在回放和实测中表现优异，ROI达15.9%，风险调整收益（Sharpe）为0.42。边缘比例策略亏损55.5%，显示单纯规模调整风险巨大。筛选与风险过滤显著提升收益，筛除低价值预测，减少亏损。系统的鲁棒性通过bootstrap置信区间和删除测试得到验证，整体表现稳健，表明明确的交易层设计极大改善了模型盈利能力。
Applications
该系统适用于自动化金融交易、预测市场管理和风险评估。只需集成预测模型和市场数据，即可实现自主交易，降低人力成本。未来还可扩展到多资产、多市场环境，辅助投资决策，提升市场效率。长远来看，结合强化学习和多策略融合，有望实现更智能、更稳健的自主交易系统，推动AI在金融行业的深度应用。
Limitations & Outlook
目前系统依赖固定档案，缺乏实时适应能力，面对市场突发事件可能表现不佳。风险控制参数需人工调节，缺少自适应机制。模型预测受限于档案信息，未充分利用实时数据。未来需在动态环境中验证系统鲁棒性，增强自我学习能力，解决实际应用中的复杂风险与不确定性。
Plain Language Accessible to non-experts
想象你在厨房里做菜，预测市场就像提前猜菜的味道。以前的方法就像只看食谱，告诉你怎么做，但不管菜做得好不好。而这个系统不仅会猜味道，还会决定用多少调料、什么时候加，确保菜不会烧焦或太咸。它会根据每次尝试的结果调整策略，就像厨师不断试错，最终做出既好吃又不浪费材料的菜。这个“厨师”能自己判断什么时候多放盐，什么时候少放油，保证每次都能做出满意的菜。这种方法让厨房变得更智能，也更可靠。
ELI14 Explained like you're 14
想象你在玩一个猜谜游戏，你不仅要猜答案，还要决定用多大力气去猜。以前的人只告诉你答案的可能性，但你还得自己决定下注多少。现在，这个系统就像一个聪明的朋友，不仅会告诉你答案的概率，还会帮你决定下注的大小，确保你不会因为太激动而亏钱。它会根据每次猜测的结果不断调整策略，就像你在游戏中学会了更聪明的玩法。这样一来，你就能用这个聪明的朋友在市场上赚到钱，而不是亏钱。它就像一个会自己学习、会控制风险的聪明助手，让你在市场中更有把握。
Abstract
Forecasting future events has attracted growing attention as a testbed for general-purpose AI. A natural way to ground this evaluation is let the models trade in the prediction markets. Trading, however, requires more than forecasting. Moreover, recent benchmarks report a substantial gap between calibrated probability scores and the trading results. We propose Raven-Agent, to the best of our knowledge, the first autonomous trading agent for prediction markets. On a controlled replay over an archived decision set, our architecture achieves the only positive return and the only positive risk-adjusted return among all tested policies. We have released our code in https://github.com/Alchemist-X/predict-raven .