Screening model — how these markets were chosen and scored

The universe

Candidate markets were not brainstormed. They were derived from three sources already in this repo, then filtered against external evidence.

Sector concentration in the angel data

Raw sector labels in the CSV are the angel groups' own words, not a taxonomy (the source README says so explicitly). They were bucketed into families:

Read this as a lagging indicator, not a leading one. It says where capital was committed over roughly the last decade. A market that is crowded here is a market where the easy entry is gone. It is used below as a distribution signal (warm intros available) and a saturation penalty — not as a "this is hot" signal.

The entry signal — what decides who wins

Every record carries a red / amber / green bar. It is the first thing on the page, and it answers a narrower question than "should I enter":

How much of the outcome here sits inside an operator's control?

This replaced a label that said Screened out, which was the wrong claim to make. A screen is a statement about a market's conditions. It is not a verdict on the reader, and it does not know what the reader brings. Someone arriving with a channel, a licence, a balance sheet or twenty years in the trade can win a market this research shows in red — and being excellent at the work does not, on its own, get anyone into a market whose binding constraint is a kiln or a quota.

How the bar is derived

It is derived, never typed. Nothing is entered per record, so the signal cannot drift away from the record it describes: change the cut factor and the bar moves. Every input is printed on the page beside it, so a reader can disagree with the reading rather than only with the conclusion.

1. The binding constraint — the record's own cut factor, weighted by how much of it is executional. Distribution is the most beatable thing on the list: the product works and the channel is owned by someone else. Capital intensity is the least: the asset decides the market before any operating skill does.
2. How fragmented the field is — measured, tier A, from Statistics Canada's business counts. An industry of very small operators has share to take and no scale player to displace, which is exactly the condition under which execution decides. Applied to operating businesses only: on a software record the industry's counts describe the buyers, not the market being entered.
3. What it costs to be in the business — the subsector's structural profile.
4. The screen's own verdict on itself — eight records say plainly that they found no clean reason to stay out. A record that declined to find a kill cannot honestly sit in the red band, whatever its cut factor implies, so the flag floors it at amber.
5. A full study outranks all of it. Where the work has actually been done — four dimensions, three kill criteria, a 30-day test — the researched verdict sets the band directly.

The percentage is a score, not a success rate

Every record shows its signal as a percentage — "42% entry signal" — because that is how a reader wants to hold a number. It is a score out of 100 built from the cut factors above, and it measures one thing: how much of the outcome sits inside an operator's control. It is not the chance of succeeding, and two markets at 42 are not equally likely to work out. A market at 80 still has to be executed; a market at 20 has been entered successfully by people who brought something this screen did not assume.

Where two markets score the same, the ranked table separates them by how much sourced evidence each record carries — among markets the model cannot tell apart, the better evidenced one is the better bet. That tie-break is stated on the page rather than hidden in a sort.

It is not a probability of success

No survival probability is published per market and inventing one would be worse than saying so. The nearest honest thing is the measured rate at which new establishments in a sector are still trading after one, three, five and ten years, published by the US Bureau of Labor Statistics. Where that is available it is shown beside the bar and labelled as what it is: the observed outcome for everyone who opened a door in that sector, good operators and bad ones together, and for the whole sector rather than this market.

Scoring

Seven factors, 0–10 each, weighted. Higher total = more attractive entry.

These scores are analyst judgment, not measurement. They are calibrated to the cited evidence in each study, but two analysts would score them differently by ±1 on most factors. The scores rank; they do not prove. The studies prove.

Evidence tiering

Every number in these documents carries a tier:

• [A] — primary source: government statistic, company filing, association report.
• [B] — reputable secondary: named research firm, trade press, law-firm briefing.
• [C] — vendor or marketing-site figure. Directionally useful, self-interested.
• [UNVERIFIED] — analyst estimate or untested assumption. Not a finding.

Third-party market-size forecasts (Grand View, MarketsandMarkets, etc.) are tier [B] at best and are frequently reverse-engineered from each other. Where a study leans on one, it says so and gives a bottom-up cross-check.

What this model deliberately does not do

• No search-volume data. No keyword tool is wired into this repo. Every study states this and specifies the exact query set to run. Search volume is quoted nowhere; absence is declared rather than filled.
• No Amazon review mining. Four of the five markets are B2B and have no Amazon presence. Where review evidence exists it comes from G2 / Capterra counts and is cited as such.
• No primary customer interviews. Every study's 30-day test is built around getting them, because that is the evidence that is actually missing.