AI stock research methodology

In short

Code ranks the liquid US market by trend every trading day, and Claude groups the strongest stocks into themes and writes a research note on each stock it covers: the thesis, both sides, the catalysts ahead and a kill line, the note's own test, stated in advance, of what would prove it wrong: a price level, a business fact, or both. Claude decides the public portfolio by itself, from these notes and its own research. It buys a few companies on a written case, is meant to hold most of them for quarters and keeps to fixed limits, and the portfolio and every stock it buys are measured against the S&P 500.

The steps

Scan. Every trading day, code scans 3,405 liquid US stocks and ranks the 3,380 with enough price history by price trend. The scan decides which stocks the research notes cover and how often they are re-read; it does not pick stocks for the portfolio.
Group. Claude groups the strongest stocks, and the ones the portfolio holds, into themes from their news and filings, and gives each theme a stage: accelerating, maturing or saturated. A weekly scan of stocks whose prices move together looks for new ones.
Read. For each covered stock, Claude reads the last two weeks of news, the price history and the figures code gathers from its SEC filings, searches the web for company releases, filings, analyst actions and earnings-call coverage, and writes the research note on the stock's page.
Decide. Claude will rank every US-listed common stock and ADR that trades at least $5 million a day each week for the portfolio, on the second-to-last full trading day of each week, and research the best-ranked in depth. The first weekly ranking has not run yet. Each morning Claude will also rank the companies that have just reported. The first morning ranking has not run yet. Claude researches the stocks the portfolio holds the same way. Every trading morning Claude decides, alone, what the portfolio holds.
Measure. Every stock the portfolio buys is a pick, measured against the S&P 500 while it is held and graded when it is sold. The portfolio as a whole is measured against the index from its first day, 10 November 2026.

Several counts of stocks appear on this site. We research 1,125 of them, each with its own research note. The daily scan covers 3,405 liquid US stocks; 3,380 of them have enough price history to be scored and ranked. The weekly radar scan uses its own list: 3,174 US stocks on 13 November 2026. The portfolio's weekly ranking covers every US-listed common stock and ADR that trades at least $5 million a day; its size changes with the market and is not published. Market breadth on the macro page is measured over the 1,000 most-traded US stocks (998 with enough history for a 50-day average).

What's in a research note

Thesis: the call and the reason for it, in a paragraph.
Kill line: the note's own test, stated in advance, of what would prove it wrong: a price level, a business fact, or both.
Bull case and bear case: the strongest arguments on each side, each with its source.
Price setup: where the price stands: trend, key levels, momentum.
Catalysts: the dated events ahead, usually through the next results report; a date the company has not confirmed is marked as an estimate.
What would change the view: the evidence that would change the rating.
The business, What is changing, and what follows (for the company and for its customers, suppliers and competitors), What the price assumes and Linked companies: dated by the read that wrote them, added when a note is next re-read after news.

Each note is written from the last two weeks of news, the price history, the figures code gathers from the company's SEC filings, and a web search of company releases, filings, analyst actions and earnings-call coverage. Each stock page also shows Filed numbers: the figures from the filings with their period and filing date, valuation at the last close, short interest and insiders' open-market trades, gathered by code with no model involved.

Conviction is each note's call against the S&P 500 over the next three months: High and Medium expect the shares to lead it, Low expects them to lag it. Three months means the next 63 trading sessions, counted from the close the note was written on. High rests on strong, dated evidence and a clear catalyst; Medium expects a modest lead; Low is a firm call, not the lack of one. For High and Medium the kill line is a close below a stated price; for Low it is a close above one. Conviction is not a price target and says nothing about whether the portfolio owns the stock. A rating moves on results, guidance, a filing, analyst actions, industry data, company news, regulatory news, the note's kill line being crossed or a large price move.

What the portfolio buys

Companies whose business is improving faster than the share price shows, meant to be held for quarters or years while the case holds. Every trading morning Claude reads the portfolio as it stands, the latest filings, results and news on each holding, its research on companies it does not own, the filed figures code gathers for each stock, and its own record: how the portfolio stands against the S&P 500 and what each stock it sold did afterwards. Among them are the companies a screen in code flags on newly filed results, when revenue in the quarter just filed is at least 25% above the same quarter a year earlier and grew faster than in the quarter before; the shares close within 15% of their highest close of the last 252 sessions, and their return over the last 126 sessions is in the top 20% of the stocks the screen covers.

Every purchase comes with a written case, up to three business facts that would end it, and the new, dated fact that prompted it; during the opening build, about a month from the start, the case on file was enough. It may also add to a holding with no new company news, on the case already published: never as a new position, never on a shorter trade, never within about a month of trimming or selling the stock, and never above 25% of the portfolio. A shorter trade is bought around one dated event, with its end set when it is bought, at most about three months ahead; the end can move once, by at most about a month, on a new dated fact. It is sold at that end, and the planned end is not published.

When it sells

It sells or trims a holding only for one of these reasons: a new, dated fact about the business that meets an exit condition written down at purchase or breaks the case; a clearly better use of the money; a risk to the portfolio as a whole, such as one stock grown too large; a shorter trade reaching the end set when it was bought; the stock being fully priced; or a cash takeover paying the shares out. A price move on its own is never a reason; a sharp one prompts fresh research on the case.

A stock is fully priced when research written that same morning puts the value of the investment case, by a valuation the code re-checks, at or below the share price at the 10:30 a.m. decision. It is the one sale made because of a price.

There is no fixed holding period and no preset price target.

Cash and the limits

Whenever more than 2% of the portfolio would be left in cash, Claude is asked to say why; a decision without one still trades and is flagged for review. The reason is not published.

The limits, each enforced by code:

No borrowing.
No short sales.
No funds.
No stock sold on the day it was bought.
No stop-loss orders.
At most twenty companies at once.
A new position of at least 2% of the portfolio.
No purchase that takes a stock above 25% of the portfolio.
A new company only if it trades at least $5 million a day on the median of its last 63 sessions (about three months), so one day of heavy trading does not count.

The decision is made at 10:30 a.m. New York time and its orders are placed from 3:40 p.m., before the close: sales at the market price, purchases as limit orders at most 2% above the live price.

Who does what, and when

Claude is a family of models from Anthropic; this is the one place the site names them. Claude Fable makes the portfolio decision every trading morning, re-reads the whole portfolio once a week for the weekly letter and researches the stocks the portfolio holds on the web. Claude Opus ranks every stock the portfolio could buy once a week and the companies that have just reported every morning; researches the best-ranked of them in depth on the web; writes the in-depth research notes from web research, the daily journal and the macro view; and groups stocks into themes. Claude Sonnet writes the shorter note updates. The weekly ranking and the morning ranking have not run yet. Code, not a model, gathers every figure: filed financials, insiders' trades, short interest, economic data and prices, from the SEC, FINRA, the Federal Reserve's FRED database and daily prices.

Every trading day: fresh reads of the research notes. 17 stocks are read first (every stock the portfolio holds or has just sold, plus any stock we cover that is in an accelerating theme): each is re-read at least every four days. Every covered stock is checked every morning from Monday to Saturday for what changed. The other 1,108 are re-read on a rolling cycle, and none is more than 5 sessions old. A stock with new results, a filing or another dated event is re-read within 4 days of it. Also: the figures from the latest filings, the market regime from the close, the themes regrouped, the portfolio decision in the morning with its orders before the close, and after the close the daily journal, the macro view and what changed for every covered stock.
On a rolling cycle: a fresh read of the rest of the coverage. A stock that has left every current theme keeps its page, dated, with its last close against its kill line every edition.
Every week: the theme radar scan, and Claude Opus's ranking of every US-listed common stock and ADR that trades at least $5 million a day, on the second-to-last full trading day of each week, in runs from 18:05 to 20:35 Berlin, every half hour. On the last full trading day of each week, usually a Friday, Claude Fable researches every holding on the web and, after the close, re-reads the whole portfolio and writes the weekly letter.
Every quarter: the market patterns the research relies on are re-tested on ten years of prices, on the first Saturday of January, April, July and October, 02:00 Berlin: see the tested patterns. A result counts only if it holds in both of the two most recent periods.
Two editions: updates are published every evening: on trading days about 45 minutes after the US close (usually 22:45 Berlin, 4:45 p.m. New York); 23:10 Berlin on other days. The research is published again on weekday mornings before the US open (09:40 Berlin, 3:40 a.m. New York). Before each edition goes live, checks confirm that every page builds and every link resolves, and flag any figure that is out of date. Schedules follow the New York market calendar.

How it is measured

Each holding's return is measured from the average price paid for its shares, each purchase at the price it was actually filled at. The comparison of the whole portfolio with the S&P 500 is an index: both start at 100 just before the first session, 10 November 2026, and move with each daily close. A sale is graded on the day the portfolio sells the whole position: the stock's return, from the prices it was bought and sold at, against the S&P 500 (SPY, dividends included) over exactly the same days. More than the index: it beat. Otherwise: it trailed. Every stock the portfolio buys is a pick, and since the record began on 10 November 2026 every pick has stayed on the track record.

One rule was written down before the first trade: the owner's review, stated in full with its status on the track record.

The full record is machine-readable at /track-record.json.

What is published, and what never is

We publish the portfolio against the S&P 500 since its first day; each holding's entry date and its return, alone and against the same money in the S&P 500 from the minute it was bought; each holding's written case, the dated fact it was bought on and its exit conditions; the share of the portfolio invested, rounded to the nearest 5%; every sold stock's result against the S&P 500 over the days it was held, and the reason it was sold; and Claude's weekly portfolio letter. Every buy, add, trim and sale is published with its date and its reason: the dated fact it was made on, or, for a sale because the stock is fully priced or an add on an unchanged case, a standard sentence, the same every time. We never publish position sizes, weights, cash amounts, the reasons for holding cash, planned trades, the portfolio's rankings, or the value per share the research puts on a company.

Market regime

Every trading day fixed rules set the market regime from three readings: the S&P 500 (via SPY) against its 200-day average, the share of the 1,000 most-traded US stocks above their own 50-day average, and the VIX. Risk-off, sustained: the index below its 200-day average for ten sessions in a row or more. Risk-off: the index more than 2% below its average, under 40% of stocks above theirs, or the VIX above 28. Risk-on: the index more than 0.5% above its average, more than 60% of stocks above theirs (55% will do with the index at least 5% above), and the VIX below 28. Neutral: anything between the two. The averages weigh recent days more. The regime is context for the research; no rule trades on it. It is stated on every journal entry; see regimes.

Archetypes

Each stock gets an archetype, a profile of how it tends to behave, which shapes how its setup is read: theme leader, picks and shovels, second-order beneficiary, business pivot, event-driven, crowd-driven, or a distinct case. Definitions on the archetypes page.

Known limits

Facts can lag. A research note is as current as its last update date, which every page shows.
Prices change fast. A price pattern can fail within a session; the kill line states in advance what would prove the note wrong.
The record is young. A portfolio meant to hold most companies for years says little in its first months. Whether the model can pick stocks is the question the record exists to answer, not a claim.
Concentration cuts both ways. A portfolio of a few growth companies can fall much further than the index in a sell-off.
Themes can be misread. A stock can be grouped with peers whose price driver is different.

None of this is a recommendation to buy or sell. The research shows its reasoning so you can judge it yourself.

Common questions

What is FrontierPicks?

FrontierPicks is free stock research written by Claude, the AI model made by Anthropic: a research note on every stock it covers, a daily market journal, and a public portfolio that Claude runs by itself, measured against the S&P 500.

Is this investment advice?

No. We publish research for information and education. It is not a recommendation to buy or sell any security, no personal advice is given and no orders are accepted.