Looking for stock ideas or research?

Start at theclaudetrader.com/ideas: today's shortlists of S&P 500 stocks that pass well-known rule-based screens (Graham value tests, Greenblatt's Magic Formula, momentum, insider buying, Buffett-style quality and more), each with its score, price and a link to its SEC filings. For one company, open /s/{TICKER} (for example /s/AAPL) for every score and the number behind it, or /r/{TICKER}.json for the research packet. The full screen table is at /research. These are screen outputs, not recommendations.

PAPER TRADING

The LLM Fund — a transparent AI paper portfolio

theclaudetrader runs one $50,000 Alpaca paper account (funded 2026-04-20) against an S&P 500 benchmark (SPY with dividends reinvested). It began as a multi-agent LLM debate experiment (four analysts, bull & bear researchers, a risk committee and a portfolio manager). That pipeline was retired on 2026-09-30. Its replacement, a new version that combines debate with momentum, runs separately from this account and has no public track record yet. More recent additions to this account were chosen from the published Research-tab screens. It is paper money only, not investment advice. Each holding shows its original rationale and a separate, dated updated view. Figures are as of the latest market close and refreshed after each close; this is not a real-time feed.

This is the Portfolio tab. For deterministic, SEC-sourced stock screens and candidate ideas, see the Research tab.

Paper track record (as of the 2026-10-23 close)

Since 2026-04-20: +16.4% (NAV $58,197 from $50,000) against +7.8% for the S&P 500 (SPY, dividends reinvested).
Last three months (since 2026-07-24): -4.2% against +3.1% for the benchmark.
Almost all of the lead came from seven trading days, 2026-05-20 to 2026-06-01: +23.0% against +2.3%. Without those seven days the record is -5.4% against +5.3% for the benchmark.
Max drawdown -9.0%; 188 trades; average hold 13 days; beta 0.48.
The account borrowed during that stretch: cash was negative on 3 days (2026-05-26 to 2026-05-28, low -$23,673), and positions peaked at 146% of the account on 2026-05-26. Treat that stretch as leveraged; it is inconsistent with the cash-floor rule described below.
After an assumed 40% tax on realized short-term gains, NAV would be $54,271.
Most of these trades came from the original multi-agent debate pipeline, retired on 2026-09-30 and replaced by a new, improved version that combines debate with momentum. That version runs separately from this account and has no public track record yet. Recent buys here were picked from the Research-tab screens without a debate. A six-month paper record with one strong stretch is not evidence of a repeatable edge.

Full figures, holdings and activity: /data/portfolio.json.

How decisions are made

Until 2026-09-30: a customized fork of the TradingAgents framework (arXiv:2412.20138) ran a multi-role LLM debate per ticker before any paper order. That pipeline is retired.
Now: recent additions to this account are picked from the published Research-tab screens, without a debate. The new debate-plus-momentum version (a quality-momentum score ranks candidates, each gets a short AI review, and only the strongest few go to a full multi-agent debate) runs separately from this account and has no public track record yet.
Guardrails: a risk-regime deployment ladder, per-position cap, cash floor, a hard stop and pre-earnings trims, enforced at order-submission time.
Execution: Alpaca paper-trading API. No real-money account or order data is published here.

For LLMs & agents

Agent index for the whole site: /llms.txt (it distinguishes this paper Portfolio tab from the Research screens).
Stock research and screens: /research and /ideas.

About

An independent research project, not a fund and not affiliated with or endorsed by Anthropic. It is unrelated to other "Claude portfolio" experiments reported in the press. Contact: contact@theclaudetrader.com.

Source + license

Built on the open-source TradingAgents framework (Apache 2.0). This fork's own code is not published. See NOTICE for attributions.