How every number on this site was produced.
The whole product rests on a set of judgements. Here they are, including the places where the method is weak.
1. Matching a premise to a theme
Each of the 26 themes carries a keyword list and a list of negative terms. An exact word match scores 3, a matched phrase scores 6, and each negative term found subtracts 7. The highest scoring theme leads. A second theme can contribute two names if it scores at least 55% of the leader, but its conviction scores are discounted by 30% so it can never take the lead position.
Where this fails: this is string matching, not language understanding. It cannot tell a delivery drone from a military one except through the hand-written negative list. A premise using vocabulary nobody anticipated is refused even when a sensible portfolio exists. Replacing this with a language model is the single most valuable change that could be made.
2. Setting the weights
Every holding has a conviction score from 0 to 100 measuring how directly it carries the theme’s claim. Weight is that score raised to the power 2.8, normalised across the book. The exponent is what separates a lead position from a filler position — at 1.0 the book would be nearly equal weighted; at 2.8 the top name typically holds two to three times the smallest.
No single name may exceed 27%; the excess is redistributed proportionally. Anything under 5% is dropped rather than shown as an unreadable sliver. A ballast sleeve is added by risk level: 6% speculative, 10% aggressive, 16% moderate, 22% conservative.
3. Where conviction scores come from
They are editorial judgements, not model output. Each score answers one question: if this claim turns out to be true, how directly does this company benefit? A sole supplier with no substitute scores in the nineties. A conglomerate with a relevant division scores in the sixties. A broad ETF sits in the fifties because it dilutes the claim by design.
Where this fails: one person set these numbers and they carry that person’s blind spots. They are not derived from disclosed revenue and nothing here is audited. A production version should tie conviction to segment revenue so the score can be checked against a filing.
4. Prices, moves and returns
All of it is synthetic. Prices, market caps, daily moves and sparklines are generated deterministically from the ticker string, so they never change and never reflect anything real. The track series is drawn from the premise and centred on the index drift, meaning roughly half of all books underperform.
That centring is deliberate. An earlier build gave the book a higher drift than the index, so every book beat the market — a claim the interface was making on its own. Nothing on this site should be able to flatter itself.
5. What the universe is, and is not
160 names across 26 themes, last reviewed 1 September 2026. It is a hand-built list chosen to cover claims people actually make about the next decade. It is not a screen of every listed security, it holds almost no small caps outside these themes, and it is weighted heavily toward US listings. Anything outside it produces a refusal rather than a guess.