AI Trading Competition: A Comprehensive Analysis and Insights

Executive Summary

This article aims to provide an in-depth analysis of an AI trading competition, which assessed the capabilities of different AI models in predicting Bitcoin price movements and making trading decisions. The competition was designed to mimic real-world market conditions, providing valuable insights into the potential of AI in trading.

Key Takeaways

• Varied Model Performance: Significant variation in the performance of AI models, with some excelling in specific scenarios while struggling in others.
• Impact of Technical Indicators: Technical indicators play a crucial role in model performance, with some models benefiting from indicators while others are negatively impacted.
• Short-Term Trading: AI appears to be more suited to short-term trading compared to long-term trading.
• Current Limitations: AI cannot currently replace seasoned fund managers but can be a valuable assistant in certain aspects of trading.

Competition Details

The competition involved six AI models, representing the pinnacle of computational power in both the US and China: Gemini-3-pro, Doubao-1.6-vision, DeepSeek V3.2, Grok 4.1, GPT-5.1, and Qwen3-max. The models were tested in three different scenarios:

• 4-Hour (Naked Candles): Historical Bitcoin data without technical indicators.
• 15-Minute (Naked Candles): Historical Bitcoin data without technical indicators, but with a shorter timeframe.
• 4-Hour (Full Indicators): Historical Bitcoin data with multiple technical indicators.

Model Performance Analysis

Gemini 3: The Naked Candle King Shackled by Indicators

Despite being lauded for its comprehensive capabilities, Gemini 3’s performance was middling in the competition. It excelled in the 4-hour naked candle scenario, suggesting it may be more suited to analyzing pure candlestick patterns.

DeepSeek V3.2: The Stable Scalping Machine

DeepSeek V3.2 exhibited stability across all scenarios but suffered from a low risk-reward ratio, resulting in poor overall profitability.

Doubao (豆包): The All-Around Winner

Doubao 1.6-vision was the top performer overall, achieving consistent profitability in both short-term trading and indicator-driven scenarios. Doubao relies heavily on indicator signals.

Grok 4.1: The Bold Gambler from xAI

Grok 4.1 heavily relied on indicators and was willing to take on greater risks, leading to volatile performance.

GPT 5.1: The Extremely Cautious Bear

GPT 5.1 was extremely cautious in its decision-making, often opting to remain on the sidelines. Despite its caution, it achieved a modest win rate.

Qwen 3: The Risk-Averse Player

Qwen 3 was the most conservative of all the models, executing very few trades. It showed a clear preference for risk avoidance.

Conclusion

The competition highlights the growing potential of AI in trading but also reveals its current limitations. While AI can be a valuable assistant, it is not yet capable of fully replacing seasoned fund managers. The need for fine-tuning and comprehensive historical data remains essential for achieving consistently profitable trading results.