{"id":16175,"date":"2025-11-14T09:28:24","date_gmt":"2025-11-14T08:28:24","guid":{"rendered":"https:\/\/haimagazine.com\/uncategorized\/experimental-trading-with-ai-which-llm-earns-the-most\/"},"modified":"2025-11-17T15:54:08","modified_gmt":"2025-11-17T14:54:08","slug":"experimental-trading-with-ai-which-llm-earns-the-most","status":"publish","type":"post","link":"https:\/\/haimagazine.com\/en\/redakcja-poleca\/experimental-trading-with-ai-which-llm-earns-the-most\/","title":{"rendered":"\ud83d\udd12 Experimental trading with AI: which LLM earns the most?"},"content":{"rendered":"
As artificial intelligence evolves and large language models (LLM) become widespread, there are new attempts to harness it in the financial world \u2014 particularly in areas where, at least in theory, you can quickly make big money. Of course, I’m talking about trading or, if you will, speculating with financial instruments. Let’s dive into the Alpha Arena experiment, where six LLMs were each given $10,000 to trade on the cryptocurrency market. Can AI can replace humans in trading? Check it out!<\/p>
The origin and purpose of the experiment<\/h4>
The Alpha Arena Experiment was kicked off by the research lab Nof1.ai and stands as the world’s first global benchmark designed to measure the investment capabilities of leading LLMs in real, dynamic market conditions. The main goal of the project was to determine if the general intelligence of LLMs is adequate for generating a market edge (known as “alpha”) and effectively managing risk in an environment that, by definition, is “chaotic, antagonistic, non-stationary, and unpredictable.” Unlike static knowledge tests, Alpha Arena put these models under market pressure, testing their decision-making abilities with real capital and volatility.<\/p>
Main principles:<\/p>
6 models:
Qwen3 Max (Alibaba),<\/li><\/ul>
DeepSeek Chat V3.1,<\/li><\/ul>
GPT-5 (OpenAI),<\/li><\/ul>
Gemini 2.5 Pro (Google\/DeepMind),<\/li><\/ul>
Grok 4 (xAI),<\/li><\/ul>
Claude Sonnet 4.5 (Anthropic),<\/li><\/ul><\/li>\n\n
Duration: October 18 to November 3, 2025<\/li>\n\n
Starting capital: $10,000<\/li>\n\n
Instruments:
Bitcoin ($BTC),<\/li><\/ul>
Ethereum ($ETH),<\/li><\/ul>
Solana ($SOL),<\/li><\/ul>
Binance Coin ($BNB),<\/li><\/ul>
Doge ($DOGE)<\/li><\/ul>
Ripple ($XRP),<\/li><\/ul><\/li>\n\n
Exchange: Hyperliquid<\/li><\/ul>
Experiment architecture and system assumptions (The Harness)<\/h4>
The key to testing LLMs in a transactional environment is to create what’s called a Harness \u2014 an architectural system that transforms the language model into an agent capable of executing an investment strategy. In Alpha Arena, the models operated in a short, repetitive decision-making loop that refreshed about every 3 minutes, according to initial assumptions.<\/p>
The models were given only raw numerical data, intentionally limiting their access to broader contexts like news, global economy or market sentiments, which could have been analyzed by humans. This data included:<\/p>
Technical indicators:<\/strong> Current prices and technical analysis indicators such as exponential moving averages (EMA), moving average convergence divergence (MACD) and relative strength index (RSI). These indicators were provided over different time intervals, e.g., 10 minutes and 4 hours.<\/li>\n\n
Account status:<\/strong> Current financial situation including available cash, open positions, current profit\/loss (P&L), the Sharpe Ratio (risk-adjusted return) and transaction fees at the Hyperliquid exchange.<\/li><\/ol>
It’s also worth noting that the models were required to adopt a specific approach to making transactions. When opening a position, the model was to have an exit plan ready, which included a price at which to take profits (known as Take Profit [TP]) and a price at which to cut further losses (known as Stop Loss [SL]). Beyond the exit strategy, a justification based on a Chain of Thought (CoT) was needed, along with confidence in the position, which is a measure of the model’s subjective certainty about the decision, expressed as a percentage.<\/p>
![\"\"](\"https:\/\/haimagazine.com\/wp-content\/uploads\/2025\/11\/wykres.png\")
![\"\"](\"https:\/\/haimagazine.com\/wp-content\/uploads\/2025\/11\/tabela1-1024x572.png\")
![\"\"](\"https:\/\/haimagazine.com\/wp-content\/uploads\/2025\/11\/Rysunek-1.png\")
![\"\"](\"https:\/\/haimagazine.com\/wp-content\/uploads\/2025\/11\/tabela-1024x187.png\")