The evidence - 661 financings, and how many were flagged first

These are real issuers and real financings, shown on a 90 day delay. Everything below closed on or before the cutoff. Subscribers receive tonight's names at 6:00 AM ET.

Read this before the numbers.

The table measures recall: how many financings were flagged in advance. Precision, how often a flag is followed by a financing, is the number an allocator actually underwrites, and it is now measured separately and shown in the two blue tiles: every day an issuer first crossed the surfacing threshold with at least 180 days of history after it, against a base rate for issuers the engine never flagged. Both figures are floors from the deterministic layer alone; no language model ran in the replay.

Filter it yourself

Every headline number above recomputes from whatever you select, so nothing on this page is a figure you have to take on trust.

Sector

Exchange

Outcome

Warning given

Score at first flag

Reached the brief

Loading.

Issuer	Sector	Flagged	Financing	Warning	Score	Brief	Filing

The full sample

What the twelve months looked like.

These charts cover the complete 813-event in-sample set from the underlying study, which is slightly larger than the delayed subset in the table above. The two agree closely: the full sample caught 70.5% at a 69 day median, the delayed subset 70.3% at 73 days.

Two histograms of lead time. The naive 180-day lookback saturates at the search boundary; the clock-restarted measure does not. — **Why 73 days and not 159.** A naive 180 day lookback reports a much larger number, but 125 of the catches were already above threshold on the first day searched, so it measures coverage of chronically distressed issuers rather than prediction of a specific raise. Restarting the clock at each issuer's previous financing removes that artefact. Only the restarted measure is reported.

Nightly funnel from filings through to names emailed. — **A median night.** 66 issuers filed something relevant, 57 cleared the trigger gates, 29 scored at or above 45, 18 entered the dossier pool, and 4 reached the brief after the 30 day repeat suppression. Capacity, not the threshold, is the binding constraint.

Catch rate by sector. — **Sector is not uniform.** Life Sciences is strongest and Natural Resources weakest, on the smallest sample. If a rubric is calibrated against your own history, that is where calibration has the most to correct.

Cumulative share of financings flagged, by days before the raise. — **How the warning accumulates.** Share of financings already flagged at 30, 60 and 90 days before the announcement.

Outcome of all 1,024 identified financing events. — **Every event, including the excluded ones.** 1,024 financing events were identified. 211 were excluded by point-in-time checks, leaving 813 in sample.

Method

How it was run.

422 simulated nightly runs across the twelve months to 28 July 2026, each scoring the full 676 issuer universe. Point in time throughout: the cutoff on every run is the run date minus one day, universe membership is re-decided on each simulated date, and only XBRL facts whose filed date precedes the cutoff are visible. Staleness is judged as of the simulated day rather than from today. That last detail matters more than it sounds: an earlier draft that judged staleness from the present reported a catch rate 23 points lower.

Why both figures are floors

No language model ran anywhere in this backtest. Only the deterministic layer was replayed: runway arithmetic, 8-K item codes, form matching, float math. In production a text pass adds signals that can only raise a score, never lower it. Both the catch rate and the lead times are therefore lower bounds rather than estimates. One further day of lead is given away on every event by the cutoff rule.

Stated limitations, in full

Precision is a cohort estimate, not a per-name promise. It is measured over threshold crossings with 180 days of history after them, so the newest crossings are excluded, and a base rate for never-flagged issuers is shown beside it. It says nothing about any particular name.
Survivorship bias. The universe is as of today, so a delisted issuer's raise is invisible to the replay.
Event detection is imperfect. 1,363 Item 1.01 filings were read and 517 named a financing instrument. Hand-checking a random sample found roughly 85 to 90% precision. Removing that channel entirely moves the full-sample catch count from 573 to 552.
Coverage gaps. 18 of 390 business days had no published daily index.
Scope. Equity and equity-linked financings only. Debt papered as a credit or loan agreement is not counted.
The rubric was not calibrated. Weights were reasoned from financing mechanics, not fitted to any investor's revealed preferences. The full rubric is published.
Float is a lagging proxy for market cap. It is measured at fiscal Q2 and can be materially stale for an issuer whose price has moved.

What this establishes is narrow and worth stating plainly: a rubric written from financing mechanics, with no calibration against anyone's revealed preferences, would have flagged seven in ten subsequent raises at a median of just over two months, using nothing but free public filings. It does not establish that those are the right names for you. Which signals matter, and how much, depends on whether you lead equity, write converts, or take down shelf paper.

Read the rubric

Filter it yourself

What the twelve months looked like.

How it was run.

Why both figures are floors

Stated limitations, in full

Request the brief