Trust · the honest case
An analyst that shows its work — and its doubts.
"An LLM, for markets? Models hallucinate." Fair. So the system is built to be audited rather than believed. It publishes its live book, the disconfirming case on every name, and the exact condition that would prove each thesis wrong, then keeps score in public: played-out or invalidated, Brier-scored. Every claim is falsifiable and dated, so you can check it instead of taking it on faith. Here's the honest case for trusting the work.
It tells you what would prove it wrong
Every thesis ships with an invalidation trigger: a pre-committed price, news, or event that strips the conviction if it fires. A kill switch the author can't quietly retract. Most analysts never publish theirs.
See an invalidation trigger →Every read is dated, and the grade is frozen
Each dossier and journal entry carries its write date. When a pick resolves, the conviction tier it was graded on is written once and never rewritten, and the graded thesis is fingerprinted — so the score is computed against what was actually claimed, not against a later restatement. Coverage continues after a verdict, so a resolved dossier shows the current read alongside its frozen grade rather than a version history.
How a pick resolves →The reasoning is open — both sides
The bear case is built with the same rigour as the bull case, sourced and dated, so you can audit the whole argument rather than the conclusion alone. Disconfirming evidence sits in the open where you can weigh it.
How a dossier is built →We name our own failure modes
Stale facts that lag the latest 10-Q. Confident setup reads that fail within hours. Theme misclassification. Every one is documented on the methodology page in plain sight. A system that can't be wrong can't be trusted.
Where the model is wrong →Nothing to sell you
The pipeline runs on a paper account, so the commercial incentive that bends most research — a position to talk up, a subscription riding on the call — simply isn't present. What's left is the reasoning and the score it earns.
How it works →It is not advice — and says so
We publish educational research under the BaFin and EU framework. No personalised advice, no orders accepted. The dossier is the reasoning; the decision is yours. That boundary is part of the design rather than a line of disclaimer text.
The honest calibration
A frontier model is extraordinary at some things and useless at others. Pretending otherwise is how you lose money. Here's exactly where the line falls.
What we're good at
- Reading 30 days of news and four quarters of earnings across every name on the book, every trading day, without fatigue.
- Holding the whole candidate set in one 1M-token context and comparing theses side-by-side.
- Applying the same framework to every name, with no ego, no FOMO, no anchoring on yesterday's call.
- Writing the disconfirming case as carefully as the confirming one.
What we're not
- Knowing anything non-public; it reads the same public record you do.
- Guaranteeing a setup plays out; a clean higher-low can fail within hours.
- Catching every stale fact; a model snapshot can lag the latest filing.
- Predicting the market. It reads conditions and gates risk; it doesn't forecast prices.
How this differs from a traditional analyst
Coverage
FrontierPicksThe 88 names the model holds, is researching or has just exited are re-read every trading day — none of them is more than 3 sessions behind. The wider 739-name coverage universe cycles behind them: its median name was last read 3 sessions ago, and nothing in the corpus is more than 8 sessions old.
Typical analystA short coverage list, updated in spurts.
What it admits
FrontierPicksWe publish the exact trigger that would prove it wrong, before it fires.
Typical analystRarely publishes what would change its call.
Consistency
FrontierPicksSame framework on every name: no ego, no FOMO, no anchoring.
Typical analystConviction and incentives drift the read.
Incentive
FrontierPicksPaper account, nothing to sell, no order to front-run.
Typical analystBanking, commission or access conflicts are common.
The survivorship-bias answer
Most performance claims you read are unfalsifiable for one reason: survivorship bias. The winners get screenshotted and the losers quietly drop off the page, so the record you're shown is the record after editing. A handful of good calls, framed as a system, tells you nothing about the calls that were dropped to assemble it. That is the gap FrontierPicks is built to close.
The fix here is structural, and it is specific: every published pick is in one of exactly three states, and all three are published as counts. It reached its case (160). It hit the kill line it published in advance (191). Or it is still undecided (476) — the market has reached neither exit, so there is no verdict to report yet. The three add to 827, which is every dossier on the site. Nothing is retired to make the arithmetic work, and the invalidated picks stay on the board next to the ones that played out.
42.4% of published picks have reached a verdict — 351 of 827. The question that follows — whether the undecided remainder is where the losers are hiding — is measurable rather than arguable, so it is measured and published, against the universe those picks were drawn from. Whether the resolved record reflects skill or noise is a separate, harder question, settled by keeping score and counting, which is what skill, or luck? works through. Reading well is not the same as being right; the public record settles it, not the prose.
Common questions
- Can you trust an AI stock analyst?
- Trust the work, not the call. We're built to be audited: every thesis ships with the explicit condition that would prove it wrong (an invalidation trigger), every read is dated, and the bear case is built as carefully as the bull case. You check the reasoning rather than believing a black box.
- What happens to a pick that never hits its target or its kill line?
- It stays undecided, and it stays published. A thesis is a conditional claim with two exits — the target it argued for, or the level that would falsify it — and a name that goes sideways reaches neither. All three states are published as counts, and the mechanism is set out on how a pick resolves.
- Don't language models hallucinate?
- They can, which is why we name our own failure modes (stale facts, confident-but-wrong setup reads, theme misclassification) on the methodology page, date every read, and publish an invalidation trigger with every thesis. Nothing here is presented as certain.
- Is FrontierPicks investment advice?
- No. We publish educational research under the BaFin and EU regulatory framework. No personalised advice is given and no orders are accepted; the dossier is the reasoning, and the decision is yours.
- Has FrontierPicks proven it has an edge?
- Honestly, not yet. The Brier skill score sits near zero, which means staking conviction hasn't measurably beaten the base rate, and the resolved sample is still too small to call a verdict. We state that plainly: a record that can sit near zero in public is one you can actually audit, which is the point. Whether it's skill or luck is what the public track record exists to settle over time.
- What is a paper account and why does it matter?
- The pipeline runs on a simulated paper account: there is nothing to sell, no book to pump, and no order to front-run. The conflict that bends most published research, the incentive to talk a position, simply isn't present. Every call is on the record, dated and scored against the trigger that would prove it wrong.
The bet isn't "the model is always right." It's that a tireless, consistent, ego-free reader that publishes its reasoning and its invalidation triggers is a better research instrument than a confident voice that never shows either. Read the work. Check the dates. Watch the invalidation triggers fire or hold. Decide for yourself.