Alibaba's AI Model Dominates Inaugural Investment Competition

November 4, 2025 - The first-ever AI large model investment competition, Alpha Arena, concluded with Alibaba's Tongyi Qianwen Qwen3-Max emerging as the clear winner. Hosted by U.S. research institution Nof1, the event demonstrated artificial intelligence's growing capabilities in financial markets.

The Competition Framework

Six leading large language models (LLMs) participated in the month-long challenge, each allocated $10,000 in virtual capital to trade Bitcoin and Ethereum perpetual contracts on the Hyperliquid platform. The unique constraints:

• Data Limitations: Models could only access numerical market data (prices, volumes, technical indicators)
• No External Context: News events or qualitative information were excluded
• Performance Metrics: Judged on both raw returns (22.32% for Qwen3-Max) and risk-adjusted Sharpe ratio

Divergent Trading Strategies Emerge

Researchers observed significant behavioral differences among participants:

• Active vs Passive: Some models executed frequent trades while others maintained longer positions
• Risk Appetite: Varied approaches to leverage and position sizing emerged organically
• Data Sensitivity: Simple prompt adjustments (like changing data presentation order) dramatically impacted some models' performance

The competition revealed unexpected nuances in how AI systems interpret financial data. "We saw certain models misread descending price sequences as bullish signals until we reformatted the input," noted a Nof1 researcher.

Limitations and Future Directions

The study acknowledged several constraints:

• Small sample size of six models
• Short evaluation period (30 days)
• Limited asset classes (only BTC/ETH derivatives)

The research team plans expanded trials in Season 2 with:

1. Additional control variables
2. More asset classes
3. Extended evaluation windows
4. Enhanced statistical measurement tools
5. Deeper behavioral analysis protocols

Implications for Financial AI Development

The Alpha Arena represents a paradigm shift from static benchmark testing to dynamic market environments. Key findings suggest:

• Untuned LLMs can generate alpha without financial-specific training
• Data presentation formats critically impact decision quality
• Risk management remains an evolving challenge for autonomous systems Nof1's initiative bridges academic research and practical fintech applications, potentially accelerating adoption of AI trading systems.

Key Points:

• 🏆 Alibaba's Qwen3-Max achieved 22.32% returns under constrained conditions
• 🔍 Models demonstrated distinct trading personalities without explicit programming
• ⚠️ Data formatting unexpectedly emerged as a critical performance factor
• 📈 Competition validates AI's potential in quantitative finance applications
• 🔬 Future iterations will incorporate more variables and assets