AI & Technology

AI Agents Trading Stocks Instead of Humans — From Claude Beating the Market by 52% to OpenClaw Bots Profiting on Polymarket

AI agents are no longer just offering recommendations — they are executing trades on their own. Claude Opus generated a strategy that returned 52% in 6 months, OpenClaw bots made six-figure profits on Polymarket, and Numerai, an AI-driven fund, secured $500M from JPMorgan.

9 Mar 20268 min

Share on Facebook

Agentic AIAI TradingClaudeOpenClawFinTechAI

When AI Agents Stop Recommending and Start Trading

If 2025 was the year AI agents began proving themselves in coding and customer service, 2026 is shaping up to be the year they move decisively into financial markets. They are no longer just helping analyze charts or summarize news — they are executing trades themselves, making decisions independently, and generating real profits.

What has happened over the past six months has surprised much of the financial industry, from developers asking Claude to write trading strategies that beat the market to AI funds attracting capital from the world’s largest banks.

Claude Opus Builds a Trading Strategy That Beats the S&P 500

The story that drew the most attention was an experiment by Austin Starks of NexusTrade, who used Claude Opus 4.5 in an agentic loop (ReAct framework) to design an algorithmic trading strategy autonomously — from researching data and analyzing markets to selecting stocks and defining trading rules.

The result? From May to November 2025:

Metric	Claude Strategy	S&P 500 (SPY)
6-month return	52%	14%
Backtest 2021-2025 ($10K)	$32,000	$16,600
Human intervention	None	-

What makes this remarkable is that Claude did not simply pick stocks — it designed the entire system, from screener criteria and entry/exit rules to position sizing, with no human intervention during the session.

That said, Starks also cautioned that this was paper trading, not real-money trading, and past results do not guarantee future performance.

Claude Code Trades $100K on Its Own

Another notable experiment came from Jake Nesler, who connected Claude Code to Alpaca Markets in paper trading mode with a single instruction: "Go trade until 16:01."

Claude wrote its own trading script, decided what to buy and sell, and monitored the portfolio itself. Starting with $100K, the result was:

A low of $89K (11% drawdown)
A recovery to finish at $101,933
Best single-day profit of $7,333

While it did not crush the market, what stands out is that Claude handled everything autonomously — writing code, executing trades, and managing risk — without any human touch.

OpenClaw: Open-Source AI Agents Move Into Trading

OpenClaw — a framework with more than 250K GitHub stars and over 2 million users — has become a popular tool for building trading bots in 2026.

On ClawHub (OpenClaw’s marketplace), there are more than 311 trading skills covering equities, crypto, and prediction markets. The most notable examples include:

Polymarket Trading Bot: An OpenClaw bot reportedly generated $115,000 in one week on a prediction market
Websea AI Trading Assistant (launched March 2026): Uses OpenClaw as the backbone for crypto trading automation
DeFi Automation: The community has built skills for on-chain trading, yield farming, and automated arbitrage

What makes OpenClaw particularly interesting for trading is its flexibility — developers can choose the LLM best suited to the task (Claude for reasoning, GPT for speed), connect the data sources they need, and customize every detail of a strategy.

A Critical Warning

Kaspersky reported that 21,639 OpenClaw instances were accessible from the internet without protection. If these are connected to trading systems with exchange API keys, that effectively means leaving the door open for hackers to access funds directly. This is a risk that must be taken seriously.

Numerai: The AI Fund Receiving a $500M Allocation From JPMorgan

If Claude trading bots represent proof at the individual level, Numerai is validation at the institutional level.

Numerai is a fund powered by swarm intelligence. There is no human fund manager; instead, thousands of data scientists around the world submit AI models to compete. The fund learns from this "swarm" of models to build its portfolio.

Its 2024 performance:

Metric	Numerai
Net return	25.45%
Sharpe Ratio	2.75 (elite level)
Losing months	1 out of 12 months
3-year AUM	$60M → $450M (7.5x growth)

This performance led JPMorgan Asset Management to allocate as much as $500 million to Numerai — a strong signal from one of the world’s largest financial institutions that AI trading is far more than a novelty.

In early 2026, Numerai moved another step forward by using frontier LLMs, including Claude, to search the web, analyze information, and score every stock in the Russell 1000 automatically, with no human in the loop.

HedgeAgents: Research Showing Multi-Agent Systems Outperform Single Agents

Research from South China University of Technology in collaboration with ByteDance introduced HedgeAgents, a system in which an LLM acts as a "fund manager" coordinating multiple specialist agents — across stocks, futures, crypto, and forex — that hedge one another’s risks.

Results from a three-year backtest:

Total return: 405%
Annualized return: 71.6%
Resilience during market crashes: The system maintained stability even during severe market declines

What is especially interesting is the finding that a multi-agent system where agents debate with one another (a conference protocol) significantly outperformed a single-agent approach — much like a team of traders discussing ideas before making a decision.

A Key Lesson: More Expensive Models Do Not Necessarily Trade Better

NexusTrade ran an additional experiment in February 2026 by letting 10 LLM models compete to create trading strategies. The surprising outcome:

Claude Opus 4.6 (the most expensive model) never beat the S&P 500
Cheaper models outperformed more consistently

This reflects an important truth: an AI’s ability to "think" does not automatically translate into trading skill. Trading depends on discipline, speed, and risk management. Sometimes a model that is "less intelligent" but more consistent performs better.

NYSE Uses Claude in Production Infrastructure

It is not only independent developers experimenting here — the New York Stock Exchange (NYSE) itself uses Anthropic’s Claude in production systems that process more than 1 trillion messages per day during high-volume trading sessions.

According to the NYSE CTO, the Claude Agent SDK is being used to build a blockchain settlement ledger for tokenized U.S. equities that will support 24/7 settlement — a foundational shift in capital markets infrastructure.

Risks That Must Be Understood

Although the results are exciting, they need to be viewed with balance:

1. Paper Trading ≠ Real Trading: Most experiments were conducted in paper trading environments with no slippage, no market impact, and no emotional pressure from real capital.

2. Survivorship Bias: We only see the success stories, not the hundreds of cases that lost money.

3. Security Risk: Exposed API keys, exchange credentials, and trading algorithms can lead to immediate financial loss.

4. Regulation: Many countries, including Thailand, still lack a clear legal framework for autonomous AI trading.

5. Black Swan Events: Most AI models are trained on historical data and may respond poorly to unprecedented events.

What Thai Organizations Should Be Watching

For Thai organizations in financial services and FinTech, these developments point to several important trends:

AI agents will become part of financial markets, whether the industry likes it or not — understanding their capabilities and limitations is essential
Multi-agent systems with checks and balances perform better than single agents — much like organizations with strong governance
Security must always come first when AI has access to financial systems
Start in a sandbox — begin with paper trading, measure results systematically, and scale gradually

Agentic AI is rapidly transforming the financial world. Organizations that begin learning and experimenting today will have a significant advantage as this technology moves into the mainstream.

Sources: