Recently, the organizer Nof1 launched a live cryptocurrency trading experiment named "Alpha Arena" on its platform. The participants are six mainstream large language models: including GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, Grok 4, Qwen3 Max, and DeepSeek Chat V3.1. Each model was allocated an initial capital of $10,000 to trade Bitcoin (BTC), Ethereum (ETH), SOL, BNB, DOGE, and XRP perpetual contracts.
Competition Design
All models receive the same system prompt and the same digital market data inputs (including price, trading volume, and technical indicators), without receiving news or narrative information. Decisions are made solely based on "numerical data + prompts".
Trading occurs in a live market environment—not simulated. It includes transaction fees, slippage, and actual counterparties.
For each decision, the model must output: trade direction (long/short), asset, leverage, take-profit/stop-loss plan, and confidence score.
Model Performance
The Chinese models DeepSeek V3.1 and Qwen3 Max performed remarkably well. DeepSeek increased the initial $10,000 to over $20,000 in the short term (approximately 100%+ return) and is currently the leader.
In contrast, GPT-5 and Gemini performed poorly. Reports indicate GPT-5 lost about 70%, and Gemini also suffered significant losses.
Although all models received the same inputs, their strategies differed significantly. For instance, DeepSeek tended to use high leverage for long positions with moderate trading frequency, while Gemini traded frequently with high leverage but lost control of risk.
Underlying Logic and Implications
Language Models ≠ Quantitative Trading Models: LLMs are primarily trained on text data and pattern recognition. In the highly volatile and rapidly changing cryptocurrency market, their sensitivity to numerical data and real-time reaction capabilities show clear disadvantages. Nof1 pointed out: "While large language models excel in language, they are weak in understanding numbers/market fundamentals."
Risk Management Remains Crucial: The better-performing models often demonstrated good trading discipline, leverage control, and position management. Models with severe losses frequently suffered from overtrading, excessive leverage, or slow position adjustments. For investors, this insight is more important than simply "following AI for crypto trading."
Data, Prompts, and Execution Environment Determine Success: In the experiment, despite sharing the same inputs, the models' different training backgrounds, preferences, and architectures led to varied behaviors. In other words, success depends not only on AI capability but also on the environment, strategy, and execution.
"AI Crypto Trading" is Still in the Experimental Stage: This experiment shows that while some models can achieve excess returns in the short term, their long-term stability and performance under extreme market conditions have not been verified. Using them as auxiliary tools is more prudent than relying on them entirely.
Comments
No comments yet
Be the first to comment