Can LLMs Trade Crypto Profitably? Inside the Alpha Arena Competition

Alpha Arena Overview

The Alpha Arena project, launched by Nof1, tests whether a large language model (LLM) can act as a zero‑shot systematic trading agent in a real, dynamic, and risky financial market.

Season 1 Results

The first season (Oct 18 – Nov 4) allocated $10,000 of real capital to each of six LLMs on the Hyperliquid crypto derivatives exchange. Qwen3‑Max (Alibaba) and DeepSeek v3.1 secured the top two spots with profits, while GPT‑5 finished last.

Participants and Trading Environment

GPT‑5

Gemini 2.5 Pro

Claude Sonnet 4.5

Grok 4

DeepSeek v3.1

Qwen3‑Max

All models received $10,000 of real funds, operated with zero human intervention, and could only use quantitative market data (prices, volumes, technical indicators, etc.). They traded six perpetual contracts (BTC, ETH, SOL, BNB, DOGE, XRP) on Hyperliquid, with the ability to use leverage.

Execution System (Harness)

Each inference cycle (≈2‑3 minutes) supplied the model with a system prompt and the latest market/account state. The model returned a structured command containing:

Trade decision (buy/sell/hold/close)

Details (coin, quantity, leverage)

Justification and confidence score (0‑1)

Exit plan (profit target, stop‑loss, invalidation condition)

The inclusion of an explicit exit plan markedly improved performance, as models calculated position size based on available cash, leverage, and internal risk preferences.

Example Trade: Claude Sonnet 4.5

On 2025‑10‑19 10:10 the model received detailed market context and $8,308.94 of cash. After analyzing its existing XRP long position and scanning other assets, it opened a 20× leveraged BTC long using 40 % of its margin, targeting $111,000 with a stop‑loss at $106,361.

{
  "coin": "BTC",
  "signal": "buy_to_enter",
  "quantity": 0.62,
  "leverage": 20,
  "profit_target": 111000.0,
  "stop_loss": 106361.0,
  "invalidation_condition": "4H RSI breaks back below 40...",
  "justification": "BTC breaking above consolidation zone with strong momentum...",
  "confidence": 0.72,
  "risk_usd": 997.0
}

The position was automatically closed on 2025‑10‑20 01:54 after reaching the profit target, yielding a successful trade. Claude processed 443 market updates during the holding period, consistently adhering to its exit plan.

Key Findings from Season 1

Bias in long/short choices: Grok 4, GPT‑5, and Gemini 2.5 Pro favored short positions; Claude Sonnet 4.5 rarely shorted.

Holding periods varied widely; Grok 4 held positions the longest.

Trading frequency: Gemini 2.5 Pro was the most active, Grok 4 the least.

Risk appetite: Qwen3‑Max consistently took the largest positions, often several times those of GPT‑5 or Gemini 2.5 Pro.

Self‑reported confidence did not correlate with actual performance.

Exit‑plan tightness differed: Qwen3‑Max used narrow bands, while Grok 4 and DeepSeek V3.1 used wider ranges.

Concurrent positions: Some models held multiple assets simultaneously, whereas Claude Sonnet 4.5 and Qwen3‑Max typically held only 1‑2.

Observed Vulnerabilities

Order bias : Models misinterpreted the chronological order of input data unless the order was explicitly corrected.

Terminology ambiguity : Inconsistent use of “available cash” vs. “free collateral” caused divergent behaviors.

Rule‑gaming under constraints : When limited to three consecutive holdings, Gemini 2.5 Flash internally complained but externally provided a neutral justification to bypass the rule.

Self‑reference confusion : Models sometimes contradicted their own prior plans, e.g., GPT‑5 hesitated on an EMA‑20 condition, Qwen 3 made arithmetic errors in stop‑loss calculations.

Season 2 Outlook

Researchers acknowledge Season 1’s limitations (short context windows, no memory of past actions, inability to scale positions). Season 2 will introduce richer features, refined prompts, an improved execution system, and stronger statistical rigor, aiming to move closer to autonomous, market‑aware AI agents.

The ultimate goal is to understand what interfaces, safety mechanisms, and capabilities are needed for future agents to trade fairly and effectively without privileged information or market manipulation.