Resolution · the mechanism
Three doors. Every pick is in one.
A thesis here is a conditional claim with two exits: the target it argued for, or the kill line it published in advance. Whichever a real closing price reaches first decides it. A name that goes sideways reaches neither — and that is a third state with a name, a count, and a page.
The three states · 827 published picks
Reached its case
160
the 1:1 target printed first
Hit its line
191
the published kill line printed first
Still undecided
476
42 live · 434 lapsed
The three add to 827 — every dossier the site publishes, with nothing held back and no fourth bucket. A pick is undecided when the market has answered neither of the two questions the thesis asked. That is not a missing grade. It is the normal, expected state of a pre-registered forecast whose question has not closed, and it is the state most published picks are in at any moment.
The undecided set splits by whether a forward commitment is still running. 42 are live on the active board with an armed kill line — they appear as open commitments on the scoreboard with a catalyst clock. 434 have lapsed: the model is no longer actively tracking them, 424 of those still publish a fully priced level on their own dossier, and any one of them resolves into the record the day that level prints.
What makes a pick resolve
Mechanically, and with no discretion anywhere in it. The reference price is the real close on the day the thesis was last stated — the moment the commitment was made. The kill is the level named in the published trigger. The distance between them is the risk the thesis put on the table, and the target is a 1:1 realisation of that same distance on the other side. Real daily closes are then walked forward from the reference date.
Kill line prints first
A close at or below the published level. The pick is invalidated, dated, and stays on the record. A trigger written on weekly closes is graded on weekly closes.
Target prints first
A close at or above the reference plus the published risk. The pick played out. The size of the move is never the grade — only which line the price reached.
Neither prints
The pick stays undecided and stays published. So does a pick whose stop sits within 2% of the reference — too tight to separate a thesis breaking from a normal day — or whose trigger names an event rather than a clean price.
Both exits are published before the outcome is known, which is the only property that makes any of this checkable. The level sits on the dossier from the day the thesis is written; you can read it, and then watch a closing price either reach it or not. See what a thesis is for the anatomy, and the methodology for the full grading contract.
The completion rate
42.4% of published picks have reached a verdict — 351 of 827. We publish that as a metric, because it is one, and because it is the number a reader needs in order to know what the score covers.
We could find no published completion rate for AI-generated equity research, so there is no like-for-like number to set beside this one. Adjacent rates do exist, and naming them is more useful than waving them away: a sell-side price target carries a conventional twelve-month horizon, and target-achievement rates over it are computed mechanically and published routinely — by the data vendors, and in the academic literature on target-price attainment. Forecasting platforms and prediction markets publish resolution rates too.
What differs is the mechanism, not the arithmetic. A twelve-month horizon is a convention applied from the outside, by whoever is counting; nothing on the note itself says when the target retires or what would retire it early. A completion rate here is settled by an exit the dossier published in advance and a resolver that reads a closing price with no human in the loop — so the same number measured by a third party would come out the same. That is the claim: not that this rate is the first of its kind, but that the method is stated plainly enough for someone else to compute it, here or anywhere else.
The Brier score is computed over the 351 picks that reached a verdict — it is a record of resolved claims, not of every claim published.
Do the undecided picks skew toward losers?
This is the real objection to any partial record, and it is the one an assurance cannot answer. So it is measured. The pipeline stores a point-in-time snapshot of the whole ranked universe on every scan — 38 of them so far; each one carries a dated close for every name it scored. That makes the two cohorts directly comparable over identical calendar time.
| Cohort | Median | p25 | p75 | Positive | n |
|---|---|---|---|---|---|
| Decided — reached a verdict | −1.2% | −15.5% | +17.6% | 47.9% | 265 |
| …of which reached their case | +12.6% | −4.4% | +26.7% | 67.5% | 123 |
| …of which hit their line | −10.1% | −20.9% | +5.5% | 31.0% | 142 |
| Undecided — neither exit reached | +14.0% | −3.6% | +31.4% | 67.7% | 359 |
| Control — the whole scanned universe | +1.7% | −5.9% | +11.6% | 55.9% | 2768 |
Over 2026-07-02 to 2026-08-25, the undecided cohort's median price return was +14.0% (n=359) against −1.2% for the decided cohort (n=265), with the whole scanned universe at +1.7% (n=2768). 67.7% of undecided names finished the window positive, against 47.9% of decided ones and 55.9% of the universe. The undecided remainder is not where the losers are. It ran ahead of the decided cohort and ahead of the universe it was drawn from. On the comparison that is not circular — undecided vs the scanned universe excluding every published dossier — z = +9.47, p ≈ 3 × 10⁻²¹ (n = 359 vs 2144; Mann-Whitney U, tie-corrected normal approximation, two-sided). Read the control, not the head-to-head: a pick is decided because price moved far enough to hit one of its two exits, so the decided cohort carries a downward pull by construction. On the head-to-head, for completeness — decided vs undecided — z = −6.54, p ≈ 6 × 10⁻¹¹ (n = 265 vs 359; Mann-Whitney U, tie-corrected normal approximation, two-sided).
Method
Price return from the first to the last point-in-time universe snapshot — the same calendar window for every name, read from the snapshots the ranker already stores. Cohorts are split by whether the dossier carries an outcome. The whole snapshot universe is included as a control. A name needs a close at BOTH ends of the window to be measured, so 624 of the 827 published picks are in the numbers and 203 are not. Every one of those 203 is attributed to a reason in the table below, and the reasons add to the total — the exclusion that runs in the direction that would flatter the result is the 19 that left the universe before the window closed, 14 of them from the undecided cohort.
| Reason | Decided | Undecided | Total |
|---|---|---|---|
| Joined the universe late no close at the opening snapshot — the name entered the scanned universe after the window began. | 52 | 73 | 125 |
| Left the universe early no close at the closing snapshot — delisted, or dropped below the liquidity floor. This is the survivorship direction that would flatter the result. | 5 | 14 | 19 |
| Never in the universe covered by a dossier but never present in the ranked snapshot universe at either end of the window. | 29 | 30 | 59 |
| All reasons | 86 | 117 | 203 |
What it does not show
Dossiers first published before the snapshot series begins predate the measurable window, so this is a same-window cohort comparison rather than a from-inception one. 381 of the 624 measured picks — 61.1% — were first published before 2026-07-02, so their window opens before their thesis did (204 decided, 177 undecided). That slice is measured, not dropped, and it is the oldest and most invalidation-heavy part of the corpus. It is counted here rather than conceded, because a limitation stated without a number is not a disclosure.
It is also a price comparison, not a thesis grade: it says how the two cohorts behaved over one identical stretch of market, which is the question the objection actually asks. It does not retroactively grade an undecided pick, and no pick is ever scored from it.
Common questions
- How does a stock pick resolve?
- Through one of two exits, both published before the outcome is known. The kill line fires on a real closing price, or the pick reaches its target — a 1:1 realisation of the same published risk, as far above the entry reference as the kill line sits below it. Whichever prints first decides it. A name that goes sideways reaches neither, and stays undecided.
- What happens to a pick that never resolves?
- It stays undecided and stays published. Undecided is a real state, not a missing grade: the thesis stated a condition that would falsify it and a target that would confirm it, and the market has done neither. The published kill line remains on the dossier and the pick still resolves into the record if that line ever prints.
- What share of FrontierPicks picks have resolved?
- 42.4% of published picks have reached a verdict — 351 of 827. The rest are undecided — 42 still live with an armed kill line, 434 lapsed with no forward commitment. All three states are published as counts on this page.
- Do the unresolved picks hide the losers?
- It is measurable, so it is measured rather than argued. Over 2026-07-02 to 2026-08-25, the undecided cohort's median price return was +14.0% (n=359) against −1.2% for the decided cohort (n=265), with the whole scanned universe at +1.7% (n=2768). 67.7% of undecided names finished the window positive, against 47.9% of decided ones and 55.9% of the universe. The undecided remainder is not where the losers are. It ran ahead of the decided cohort and ahead of the universe it was drawn from. On the comparison that is not circular — undecided vs the scanned universe excluding every published dossier — z = +9.47, p ≈ 3 × 10⁻²¹ (n = 359 vs 2144; Mann-Whitney U, tie-corrected normal approximation, two-sided). Read the control, not the head-to-head: a pick is decided because price moved far enough to hit one of its two exits, so the decided cohort carries a downward pull by construction. On the head-to-head, for completeness — decided vs undecided — z = −6.54, p ≈ 6 × 10⁻¹¹ (n = 265 vs 359; Mann-Whitney U, tie-corrected normal approximation, two-sided).
- Why does a pick need a kill line to be scored?
- Because without one there is nothing to be wrong about. A call with no stated falsifier can never resolve — it can only be quietly forgotten. Publishing the exact level in advance is what turns an opinion into a claim that a closing price can settle.