AI stock research, graded in public.
A frontier model reads the US market daily and publishes, in advance, exactly what would prove each call wrong.
The book · 10 September 2026 close
What we hold, in full.
The state of the research
Everything, at a glance.
Catalysts · this week
On the clock
153 this week
Atlas Energy Solutions Inc.
earnings
Allegion Public Limited Company
fda
Alto Ingredients, Inc.
earnings
Advanced Micro Devices, Inc.
earnings
AST SpaceMobile, Inc.
earnings
Constellation Energy Corporation
earnings
Bowman Consulting Group Ltd.
event
Krispy Kreme, Inc.
earnings
Atkore Inc.
earnings
Cullinan Therapeutics, Inc.
fda
COMPASS Pathways Plc
fda
CRA International, Inc.
earnings
+149 more this week
What changed · 7d
Latest moves
55 new · 5 resolved
Hot · momentum
Strongest trends
top 5 of 3,430 ranked
Themes in motion
Clusters picking up
0 accelerating · 5 maturing
No clusters are accelerating right now. The board updates as the model re-reads co-movement.
Scoreboard
Every pick, graded
332 resolved · 331 scored
332resolved · 331 scored · Brier 0.26
160 played out · 172 invalidated — each against its own published kill line. Brier covers 331: one pick staked no conviction tier, so it cannot be scored for calibration.
Source: the FrontierPicks nightly pipeline — computed from US market close data, as of (Europe/Berlin).
The record so far
Everything filed, and how it ended.
Through the 10 September 2026 close
Ranked · highest conviction first
Dossiers in rotation
Most recent journal entry
Thursday, 10 September 2026
Neutral
Excerpt
The real 10Y rate rose 3bps week over week through 2026-09-09, increasing the inflation-adjusted rate against which equity valuations are assessed. This is the clearest valuation headwind in the supplied rate data; whether equities respond remains unresolved. FRED’s weekly changes do not establish what moved during today’s equity session. Inferred:…
The index
Everything the model publishes, and how it works.
What the model publishes
- Dossiers929Thesis, kill line, catalyst — per ticker.
- Holdings28The live book: what we hold, and the line under each name.
- Scoreboard332Every resolved pick, graded against its own kill line.
- Calendar343Dated catalysts ahead — this week, next, later.
- Journal109Every regime call, dated and indexed.
- Themes in motion14Which clusters are accelerating, maturing, saturated.
- ThemesEvery narrative cluster the model tracks, and the names inside it.
- Hot stocksCurated momentum views — the strongest US names by trend.
- SectorsWhere the flow is rotating, week by week.
- State of the bookLive coverage — names tracked, conviction mix, current regime.
How it works
- How it worksHow a name becomes a thesis, stage by stage.
- How a pick is gradedWhat counts as played out, what kills a pick, and what is still open.
- MethodologyFive stages, archetype taxonomy, regime logic.
- Anatomy of a dossierA live dossier dissected — each part mapped to its stage and model.
- ArchetypesEight behaviour profiles, each a hard rule on how big a name can get.
- RegimesThe macro-regime taxonomy and exactly what each regime gates.
- Kill gatesThe pre-registered, tighten-only conditions that shut the strategy down.
- Theme radarThis week's mechanical momentum clusters — the candidate list, before curation.
- Scored forecastsThe theme-bets, each with a stated probability — open bets and refusals.
- Why trust itThe honesty mechanisms: published kill lines, dated reads, nothing to sell.
- AI equity researchWhat it is, why a public track record matters, and how to judge one.
News window
30d
per candidate, daily
Earnings transcripts
4Q
read per name
Archetypes
8
behaviour profiles
Context
1M
side-by-side candidate review
Common questions
What FrontierPicks is, in plain terms.
- What is FrontierPicks?
- FrontierPicks is AI stock research that grades itself in public. Every trading day, frontier language models — Anthropic's Claude Opus and Sonnet — read US equity news, filings and prices and write per-ticker dossiers, a regime-tagged journal, and weekly macro and sector views. Every call names, in advance, the exact price or event that would prove it wrong; when that level is reached the call is closed and scored against it, and the losses stay published. We currently track 929 names. The 56 names the model holds, has just exited, or that sit in a theme it is currently calling accelerating are re-read every trading day — every one of them was read in the current session. The wider 873-name coverage universe cycles behind them: its median name was last read 3 sessions ago, and no dossier is more than 8 sessions old. The reasoning, the names we hold, and the dated record are all public, and there is nothing to buy.
- Does FrontierPicks keep score?
- Yes. 332 picks have resolved in public so far (160 played out, 172 invalidated), each graded against the exact kill line it published in advance — not a return marked to a benchmark after the fact. Across the 331 that staked a conviction tier the Brier score is 0.26 — 0 is perfect, 0.25 a coin flip. The full Brier-scored record is public on the scoreboard.
- What do we publish?
- Everything the model thinks. Every pick, regime call, bull and bear case, and kill line — the full reasoning — plus the names we hold, each dated and scored against the line that would prove it wrong.
- What is the current market regime?
- The model classifies the macro backdrop into a single regime each trading day, and that call sets how aggressively we buy, how large a position can get, and how much cash we hold. The most recent call, as of 10 September 2026, is Neutral. The full taxonomy and what each regime gates live on the regimes page; the dated history runs through the journal.
- Is this investment advice?
- No. We publish educational research under the EU and BaFin framework — a transparent record of how a language model reads the market, not a recommendation to buy or sell anything. Always do your own work.
- Is there anything to buy or sign up for?
- No. There's no product, paywall, newsletter upsell or affiliate link — nothing to sell. We publish in the open as a falsifiable, dated track record. The incentive is to be right, and to be seen being right.
Built for agents to crawl
No auth. No rate limits. Canonical URLs for every resource.
JSON-LD on every page, llms.txt + llms-full.txt for context windows, JSON Feed + RSS for event-driven agents, WebMCP + OpenAPI for discovery. Cite the canonical URL.
Agent endpointsThemes we're tracking
Archetypes in rotation · behaviour profiles
Full taxonomy →