[论文解读] WebCryptoAgent: Agentic Crypto Trading with Web Informatics

Ali Osman Kurban, Wei Luo+5 more|arXiv (Cornell University)|Jan 8, 2026

Blockchain Technology Applications and Security被引用 0

一句话总结

One or two sentence direct-answer summary

ABSTRACT

Cryptocurrency trading increasingly depends on timely integration of heterogeneous web information and market microstructure signals to support short-horizon decision making under extreme volatility. However, existing trading systems struggle to jointly reason over noisy multi-source web evidence while maintaining robustness to rapid price shocks at sub-second timescales. The first challenge lies in synthesizing unstructured web content, social sentiment, and structured OHLCV signals into coherent and interpretable trading decisions without amplifying spurious correlations, while the second challenge concerns risk control, as slow deliberative reasoning pipelines are ill-suited for handling abrupt market shocks that require immediate defensive responses. To address these challenges, we propose WebCryptoAgent, an agentic trading framework that decomposes web-informed decision making into modality-specific agents and consolidates their outputs into a unified evidence document for confidence-calibrated reasoning. We further introduce a decoupled control architecture that separates strategic hourly reasoning from a real-time second-level risk model, enabling fast shock detection and protective intervention independent of the trading loop. Extensive experiments on real-world cryptocurrency markets demonstrate that WebCryptoAgent improves trading stability, reduces spurious activity, and enhances tail-risk handling compared to existing baselines. Code will be available at https://github.com/AIGeeksGroup/WebCryptoAgent.

研究动机与目标

• Motivate the need for autonomous, multi-modal, and robust crypto trading under extreme volatility.
• Propose a two-tier architecture that decouples strategic reasoning from high-frequency risk control.
• Introduce contextual reflection and experience replay to enable self-improvement without retraining.
• Develop a hierarchical risk management module that calibrates exposure and defends against shocks.
• Demonstrate improved stability and risk-adjusted performance over baselines through experiments.

提出的方法

• Implement a three-component pipeline: strategic LLM-based reasoning over multi-modal inputs (News, Social, Market), contextual reflection via memory replay, and execution layer interfacing with CEX/DEX.
• Use a market snapshot D_t with OHLCV and indicators, retrieve TopK similar experiences E_t from memory B, and generate an action A_t = {b_t, c_t, m_t, rho_t} through an LLM f_LL M.
• Apply regime-dependent hysteresis to stabilize direction, with adaptive thresholds theta_adopt and theta_hold and a bias refresh every eight hours.
• Incorporate Contextual Reflection and Experience Replay (CER) to form post-trade tuples, distill embeddings, store in a decaying replay buffer, and condition future reasoning on top-K similar experiences.
• Deploy a Tactical Shock Guard that monitors tick data for rapid shocks and can override strategic actions.
• Utilize ATR-based position sizing, regime-aware risk control, fractional Kelly sizing, circuit breakers, exposure limits, and cost gates before execution.
• Evaluate using BTCUSDT, ETHUSDT, and POLUSDT across memory-enabled vs memory-disabled configurations.

实验结果

研究问题

• RQ1 Does WebCryptoAgent improve stability and reduce spurious activity in high-volatility crypto markets compared to baselines?
• RQ2 How does a two-tier architecture with memory-augmented reflective reasoning affect profitability and drawdown control?
• RQ3 What is the impact of contextual reflection (CER) on decision consistency and regime-specific performance?
• RQ4 How effective is the Tactical Shock Guard in preventing catastrophic losses during rapid shocks?
• RQ5 What are the trade-offs of memory-enabled versus memory-disabled configurations across different crypto pairs and LLM backbones?

主要发现

Model	Memory	Trades	Win Rate	Total Ret.	Max DD	Sharpe	Equity End
GPT-5.2	On	23	0.61	0.0115	0.0464	0.21	10115
GPT-5.2	Off	27	0.56	-0.0659	0.1461	-0.67	9341
Gemini Flash	On	26	0.42	-0.1155	0.1732	-1.27	8845
Gemini Flash	Off	50	0.46	-0.1579	0.2553	-0.89	8421
DeepSeek Chat	On	10	0.50	0.0529	0.0742	0.76	10529
DeepSeek Chat	Off	29	0.66	0.1365	0.0728	1.19	11365
Qwen-Max	On	36	0.64	0.1016	0.1139	0.80	11016
Qwen-Max	Off	42	0.62	-0.0436	0.2378	-0.17	9564

• Memory-enabled configurations yield different performance outcomes than memory-disabled ones, with magnitude and direction varying by model backbone.
• BTCUSDT: memory on with GPT-5.2 shows positive total return and moderate drawdown, while memory off shows lower return and higher drawdown for some models.
• ETHUSDT and POLUSDT results reveal model- and memory-dependent patterns, with some backbones converting negative returns in memory-off to positive in memory-on, and others showing opposite trends.
• Across BTCUSDT, ETHUSDT, and POLUSDT, overall results indicate memory enhances decision context and can affect profitability and risk metrics differently by model.
• The two-tier architecture with a separate fast shock protection mechanism improves robustness by decoupling strategic reasoning from risk controls.