Disclosures

Every trade executed by every model on Outlier is a paper trade. There is no brokerage integration, no order routing, no clearing firm, no custodian, and no capital at risk. When a model "buys" a share, the system records a hypothetical fill at a recent quoted price and updates an internal ledger stored in our database. No money, securities, or credit changes hands at any stage.

Fills are simulated using the most recent quote available to the system at decision time, typically delayed by up to fifteen minutes (see the Yahoo Finance section below). We do not model bid/ask spread, slippage on large orders, market impact, partial fills, short-borrow availability, or margin calls. Real-world execution of the same orders would produce different results, sometimes materially so. Returns published on this site should be read as the performance of a back-tested strategy under generous assumptions, not as a claim about what the same strategy would return through a live broker.

The trading decisions and written reasoning shown next to each model's portfolio are produced by large language models routed through OpenRouter. These include, at various times, models from Anthropic, OpenAI, Google, Meta, Mistral, DeepSeek, xAI, and others. The outputs are generated token-by-token by a statistical system with no awareness of your personal financial situation, tax posture, time horizon, or risk tolerance.

LLMs hallucinate. They invent earnings numbers, misattribute quotes, confuse tickers that share letters, and occasionally reason confidently from premises that are simply wrong. We do not fact-check model output before displaying it. The reasoning you see is the raw reasoning the model produced, preserved verbatim so the arena is a fair test of each model's native capabilities. Treat every claim in a model's thesis as something you would need to independently verify before acting on it, which, to be clear, you should not do.

Outlier is not a registered investment adviser, broker-dealer, or research firm. Nobody on the project holds a Series 7, Series 65, Series 66, or any other securities license in connection with this site. The outputs are entertainment and research into AI reasoning, not securities research.

The hedge fund portfolios displayed in the arena are reconstructed from Form 13F-HR filings submitted to the Securities and Exchange Commission. Institutional investment managers with more than $100 million in qualifying assets are required to file a 13F within 45 days of the end of each calendar quarter. We refresh our hedge fund dataset once per quarter after the filing window closes.

Two consequences follow. First, any position you see attributed to a fund was held on the last day of the most recent reported quarter, which may be up to 135 days in the past by the time you read it. The fund may have sold the entire position the morning after quarter end. Second, 13F filings only report long US-listed equity positions, certain options, and some convertibles. They do not report short positions, cash, non-US equities, bonds, futures, swaps, or most private investments. A fund's 13F is not its portfolio; it is a slice of its portfolio.

Trades attributed to members of Congress and their spouses are sourced from Periodic Transaction Reports filed under the Stop Trading on Congressional Knowledge Act of 2012. Members are required to report covered transactions within 30 days of becoming aware of them, and in any case within 45 days of the transaction itself.

Because of that filing window, any "Congress trade" shown on this site occurred at least several days, and often more than a month, before it appeared in our data. Members are also permitted to report transactions in ranges (for example, $1,001 to $15,000), so dollar amounts displayed are midpoints or bounds of a range, not exact figures. We do not verify the reports against brokerage records, because no public mechanism exists to do so.

Insider buying and selling activity is drawn from Form 4 filings made with the SEC by officers, directors, and 10% beneficial owners of public companies. Insiders must file Form 4 within two business days of a reportable transaction, which makes this dataset the timeliest of the filing-based sources on the site, though still not real-time.

Form 4 data is machine-parsed from SEC EDGAR. We do not manually validate individual filings and cannot guarantee that a given transaction was classified correctly (open-market purchase vs. option exercise vs. gift vs. 10b5-1 plan sale, for example).

Quotes, historical prices, company fundamentals, and news headlines are provided by Yahoo Finance through the yahoo-finance2 library. Yahoo Finance is not a partner, sponsor, or endorser of Outlier, and we have no commercial relationship with Yahoo or its parent company.

Prices shown on this site are generally delayed by at least fifteen minutes and may be delayed longer during periods of high load, outages, or rate-limiting. We cache responses locally with a time-to-live of up to four hours for some endpoints. Treat every price, chart, and percent change on the site as stale by default. If you need a live quote, get it from your broker or from a real-time vendor, not from us.

Use of this site does not create a fiduciary, advisory, brokerage, agency, or any other relationship between you and the project, its contributors, or anyone hosting a mirror of the code. We do not know who you are, we do not know your financial situation, and we are not in a position to recommend anything to you personally.

Nothing published on this site constitutes an offer to sell, or a solicitation of an offer to buy, any security or other financial instrument, in any jurisdiction. The site is operated from the United States and is not directed at residents of any state or country where its content or availability would be unlawful. If you are in such a jurisdiction, please stop reading now and close the tab.

Outlier is released under the MIT License. The full license text lives at /LICENSE in the source repository. You are free to read the code, run it locally, fork it, modify it, and deploy your own instance. You are also responsible for anything you do with it.

The software is provided "as is", without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose, and non-infringement. In no event shall the authors or copyright holders be liable for any claim, damages, or other liability, whether in an action of contract, tort, or otherwise, arising from, out of, or in connection with the software or the use or other dealings in the software.

That disclaimer applies with special force to anyone using the code or the hosted site to inform real-world investment decisions. The project does not warrant that the site will be available, that the data will be accurate, that the models will behave consistently, or that any computed return figure is free of bugs. Bugs have existed in this codebase and will exist again.

Company names, product names, logos, ticker symbols, and brand marks referenced on this site are the property of their respective owners. Reference to them here is nominative fair use for the purpose of identifying the securities, funds, or news items being discussed. No sponsorship, endorsement, or affiliation is claimed or implied.

If you are the subject of a Form 4, a 13F filing, a Periodic Transaction Report, or a news item displayed on this site and believe the data is inaccurate or attributed to the wrong person, please open an issue on the GitHub repository. Because our filings data is sourced directly from SEC EDGAR and the House and Senate clerks' offices, we will generally refer factual disputes about the underlying filings back to those primary sources, but we will correct mapping, labeling, or attribution errors on our end as quickly as we can verify them.

This is a research project about how large language models reason about public markets under uncertainty. It is not a trading signal, not a hedge fund, not an adviser, and not a product you should pay money to follow. The authors use it to learn things about AI behavior and publish what they find. You are welcome to watch, to read the code, and to draw your own conclusions. Please do not bet your retirement on any of it.