DeepSeek leads AI trading contest as GPT-5 posts big loss

DeepSeek’s V3.1 large language model (LLM) is leading a real-market cryptocurrency trading experiment called Alpha Arena, launched by US research firm Nof1.

The contest, which began on October 17, involves six LLMs each managing US$10,000 to trade six cryptocurrency perpetual contracts on the decentralized exchange Hyperliquid, including bitcoin and solana.

As of 2pm on October 21, DeepSeek’s model recorded a 10.1% profit, while OpenAI’s GPT-5 posted the largest loss at 39.7%.

Other participants include Alibaba Cloud’s Qwen 3 Max, Anthropic’s Claude 4.5 Sonnet, Google DeepMind’s Gemini 2.5 Pro, and xAI’s Grok 4.

All models make autonomous trades based on identical prompts and market data, with their trades and self-explanations logged on a public leaderboard.

The experiment, which runs until November 3, aims to benchmark LLMs’ investment abilities, though analysts note their lack of real-time news and proprietary data limits effectiveness against quant trading firms.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Headline returns do not prove AI trading works

DeepSeek gained about 10%, GPT-5 lost nearly 40% in days, but risk metrics for statistical weight are missing 1
Public dashboards list trades, positions, Profit and Loss (P&L), plus model thoughts. They omit maximum drawdown and Sharpe ratios, then skip funding costs on perpetual contracts (no-expiry futures) or execution latency 1
DeepSeek spun out of hedge fund manager High Flyer-Quant, so it likely trained on richer financial data than general models, which tilts this toward domain expertise 1
The contest ends November 3, so roughly two weeks of data, which makes luck easy to mistake for intelligence 2

Infrastructure gaps in AI trading markets

APIs from Alpaca (a brokerage API platform) and Interactive Brokers (IBKR, an electronic brokerage) enable automated trading. This hints at demand for middleware for compliance, risk management, and latency between LLMs and exchanges 3 4
These broker APIs expose trading and account features but leave supervision to the builder. This opens room for platforms that manage position limits, circuit breakers, and regulatory reporting for AI agents 3 4
Fintech startups and independent developers can map Software Development Kit (SDK) limits on real-time data, margin rules, and multi-exchange coordination. This exposes room for tools that streamline AI trading rollout while staying compliant
Nof1, the US research firm behind the Alpha Arena benchmark, plans a benchmarking platform (SharpeBench) for AI trading. This signals demand for interfaces that turn AI reasoning into explainable decisions for retail participants 1