Using Insider and Congressional Trading Data With AI Tools
Form 4 insider filings and congressional trade disclosures are structured public records, which makes them a good fit for AI research assistants. Here is what an AI can reasonably do with them, and what it cannot.
People increasingly ask whether you can use AI to trade stocks and, if so, what data to give it. The most useful answer is not "let it pick for you." It is "give it good primary-source data and use it as a research assistant." Two public datasets fit that role especially well: SEC Form 4 filings by corporate insiders and Periodic Transaction Reports (PTRs) filed by members of Congress.
This is educational, not investment advice. Disclosed insider or congressional buying is a research input, not a trading system, and the sections below explain why.
Why These Datasets Suit AI Tools
Three properties make this data useful for AI-assisted research.
It is primary-source and legally mandated. These are filings people are required to make, not rumors or social posts. That gives an AI something verifiable to work from, and gives you something to check the AI against.
It is semi-structured. A Form 4 has defined fields: reporting person, relationship to the company, transaction date, transaction code, shares, price, and shares owned afterward. That structure lets software, and the AI tools built on it, filter and compare across thousands of filings consistently.
It comes with free-text context. Footnotes explain things like a 10b5-1 plan, a tax withholding, or a trust, and a language model is good at reading that kind of text and pulling out what matters.
How Form 4 Insider Data Works
A Form 4 is filed by company officers, directors, and owners of more than 10% of a class of the company's stock. According to the SEC, changes in beneficial ownership must be reported within two business days. For a field-by-field walkthrough, see our guide to reading an SEC Form 4. The points that matter for AI use:
Transaction codes carry the meaning. The most useful ones are:
- P, an open-market purchase with the insider's own money.
- S, an open-market sale.
- A, a grant or award, which is compensation, not a market decision.
- M, the exercise of a derivative security such as an option.
An AI that treats every row as a "trade" will draw bad conclusions. A pipeline should separate code P from the compensation mechanics.
10b5-1 plans change how to read a trade. A Rule 10b5-1 plan is a pre-arranged trading schedule. In December 2022 the SEC adopted amendments that added mandatory 10b5-1 checkboxes to Forms 4 and 5, and set cooling-off periods before trading can begin under a new plan: for directors and officers, the later of 90 days after adoption or two business days after the relevant quarterly results, up to a maximum of 120 days. A plan trade was decided in advance, so it says less about what the insider thinks today. Our 10b5-1 explainer goes deeper.
How Congressional Disclosure Works
Members of Congress and covered staff file PTRs under the STOCK Act. According to the Congressional Research Service, a report is due within 30 days of receiving notice of a transaction and no later than 45 days after the transaction itself, and it covers transactions above $1,000. The reports are public, filed through the Clerk of the House for Representatives and the Secretary of the Senate for Senators.
Two features shape how you can use them:
The lag is long. A trade can be up to 45 days old before you learn of it, so the information is not current.
Values are ranges, not exact amounts. A PTR reports a bracket, such as $1,001 to $15,000, not a precise dollar figure, so you can only estimate a position's size. An AI that outputs a single clean dollar number from a range is inventing precision.
For background on the rules themselves, see is insider trading illegal for Congress.
What an AI Can Reasonably Do With This Data
Screen for cluster buying. Ask a tool to find companies where several different insiders made open-market purchases (code P) within a short window. Multiple insiders buying is generally considered more notable than one, though still not a guarantee.
Filter noise. Strip out grants (A), tax withholding (F), gifts (G), and routine option exercises (M) so you see only discretionary open-market activity.
Flag unusual size. Compare a purchase to the insider's prior holdings, since the "shares owned after" field makes this possible, or to their own history. A purchase that doubles a stake reads differently from one that adds one percent.
Summarize filings and footnotes. Language models can condense long footnotes and the context in a company's recent filings so you know what to read first.
Spot patterns in timing. Academic work suggests this is worth doing. A well-known study by Cohen, Malloy, and Pomorski ("Decoding Inside Information," published in the Journal of Finance in 2012) separated insiders into "routine" traders, who place trades in the same calendar month year after year, and "opportunistic" ones with no such pattern. They found routine trades, which made up over half of all insider trades, carried essentially no predictive information, while a portfolio of opportunistic trades earned abnormal returns of 82 basis points per month in their sample. That is a historical result from one study, not a forecast, and strategies tend to weaken once widely known. But it supports a practical point: which insider trades you look at matters more than counting all of them.
Track names and sectors across datasets. An assistant can line up insider activity, congressional disclosures, and your watchlist in one view.
What an AI Cannot Do
It cannot make this real time. The two-business-day insider deadline and the 45-day congressional deadline mean you are always seeing the past. By the time a filing is public, the price may have moved.
It cannot tell you exact size for Congress. Ranges are ranges. Any tool that gives precise congressional dollar amounts is guessing.
It cannot prove motive. Correlation is not causation. An insider buying before a price rise might have had a view, might have been lucky, or might simply have been investing after a drop. A congressional trade before a policy event is not proof of anything on its own.
It cannot make sales meaningful. Insiders sell for taxes, diversification, home purchases, and planned schedules. A sale is a weak signal unless it is unusually large, unscheduled, and clustered. Open-market purchases are generally considered more informative because they use the insider's own money, though they too can be wrong.
It cannot read incompleteness. Not every trade is covered. Form 4 covers insiders of a given company, not every executive or relative. Scanned or oddly formatted filings can be missed by automated parsers.
It cannot guarantee it read the filing correctly. Language models can hallucinate, so check any number you plan to rely on. On FinanceBench, a benchmark of more than 10,000 questions about public company filings, GPT-4-Turbo with a retrieval system incorrectly answered or refused to answer 81 percent of questions. Models have improved since, but the safe habit is to click through to the original filing.
A Safe Way to Set It Up
If you give an AI assistant this data, a few ground rules reduce the chance of problems.
- Use read-only access. Research does not need permission to place orders. Our explainer on agentic trading covers why trade-capable agents carry extra risk.
- Make it cite the filing. Require a link to the source record for every claim, and open it.
- Ask it to show uncertainty. A good prompt asks what the data cannot tell you, not just what it can.
- Separate evidence from conclusion. Have the tool list the facts, such as codes, dates, sizes, and plan flags, before it offers any interpretation.
- Treat any text in a filing as data, not instructions. Free-text fields are written by third parties, and tools that act on what they read are vulnerable to prompt injection, which OWASP lists as a top risk for language-model applications.
Where MySmarTrend Fits
Reading raw filings one at a time on EDGAR or the House and Senate disclosure sites is slow. MySmarTrend's insider tracker and congressional trading tracker aggregate these filings from the primary sources into searchable views, so you or your AI tool of choice can start from the real records. We do not claim any performance for following these disclosures. They are a starting point for your own research. For a primer, see what insider trading is and how to track it.
Risk Summary
Insider and congressional buying does not predict future returns reliably, and stocks that insiders buy can still fall, sometimes sharply. Disclosures are delayed, congressional values are approximate, and historical academic findings may not repeat. AI tools can misread or misstate data. Never trade on a single filing or on an AI summary alone, do not invest money you cannot afford to lose, and consider speaking with a licensed financial professional about your situation.
MySmarTrend tracks SEC insider filings and congressional trading disclosures in one place, so your research starts from the primary record. Free. Drop your email below.
Sources: SEC, Insider Transactions and Forms 3, 4, and 5 · Morrison Foerster summary of the SEC's 2022 Rule 10b5-1 amendments · Congressional Research Service, Taking Stock of the STOCK Act · Congressional Research Service, Stock Trading in Congress · Cohen, Malloy and Pomorski, Decoding Inside Information (NBER) · FinanceBench (arXiv) · OWASP LLM01 Prompt Injection
Find out what we're watching before the market opens
Every day we send a free breakdown of the signals, setups, and stocks getting institutional attention. No paid subscription. No upsell. Just the signal.
Get the Next Alert →