Methodology — Best Ideas

Transparency is a feature. This page explains what Best Ideas builds and why — in plain terms. The exact weights, bands and cutoffs are proprietary and documented in our internal methodology; the summary here is deliberate, not an omission.

Best Ideas builds a scored, weighted portfolio out of the stocks a chosen set of 13F filers actually hold. Conviction is how many of them hold a name and how big a position it is for them — never a dollar total, so the largest fund in the set cannot out-vote the rest.

The caveat that governs the whole page. A 13F filing is a quarterly list of long US positions. It has no shorts, no cash, no options and no non-US listings, and it arrives up to 45 days after the quarter it describes. Every book here is therefore a lagged one, built from those lists and our own fundamentals.

Conviction: breadth that cannot be bought by cloning a book

Conviction blends two things, both measured as ranks across the candidate names rather than in dollars: breadth — how many of the selected filers hold the company — and depth — how large the position is inside the books that hold it.

Breadth is not a head count. Two filers can hold nearly the same book, and counting them as two independent votes would let a set of near-identical funds manufacture consensus. So each filer's vote is deflated by how much of its book it shares with the others in the set: two identical books together carry about one manager's vote. The deflation is deterministic — it does not depend on sampling, or on the order the managers were picked.

Because both halves are ranks rather than dollars, the score is fund-size-neutral. That is the same principle behind the "match the managers' sectors" target, where each filer's own sector mix gets one vote instead of the mixes being averaged by assets under management.

Candidates are grouped by company, not by ticker line. Two share classes of one issuer count once — one idea, one slot in the book, one position cap — and a filer holding either class counts once. The ticker shown is the class those filers hold more of by value.

Scoring: percentile ranks, and an honest price for missing data

Conviction sits in a blend with quality, value, growth, momentum and an optional earnings-quality lens, each of which the reader weights. Every factor is percentile-ranked across the candidate set first, so factors on entirely different scales combine without one of them dominating by unit.

A name missing an input is scored neutrally on it rather than being dropped or assumed the worst. On its own that would make thin coverage a free pass to mid-pack, so a name's blended score is then scaled down in proportion to how much of the blend actually saw real data. The scaling never invents a value; it prices the absence, and each row reports what share of its factor inputs were real so a reader can see which names were marked down.

Conviction is exempt from that discount, because it is always real: it comes from the filings themselves.

Fair-value upside enters the value factor only when the valuation engine trusts its own number. A serve the model has flagged, or a low-confidence headline, is a disclosed model view rather than a value signal, and it is scored neutrally instead of being laundered into the blend as evidence.

The value-trap discount

A name carrying our value-trap flag has its blended score cut sharply before selection. The flag is an empirical loser fingerprint — the cheap, high-margin, low-research profile that underperformed through the study window — and the discount is deliberately blunt rather than finely tuned: it is de-selection pressure, not a score adjustment pretending to precision. The flag is printed on the row it acts on, so a reader sees the discount instead of inferring it.

Selection: sector quotas, or none at all

Sector targets are the reader's choice, and there are three.

Agnostic takes the best-scoring names and nothing else. Managers aims at the average of the selected filers' own sector mixes, one filer one vote. Benchmark aims at real index constituent weights — the S&P 500 or the Nasdaq-100 as filed — or at a broad cap-weighted mix of covered US large caps as a disclosed proxy. We offer exactly the constituent data we hold; other indices are not ingested, and offering them would mean inventing a benchmark.

Under a target, slots are apportioned across sectors and each sector's best-scoring names take them; the delivered weights are then scaled to the target, so the mix matches it rather than only the name counts. The result always reports both the aim and the achievement, because a position cap or floor can move the delivered mix off the target and a reader should be able to see by how much.

Sectors are GICS, mapped from SEC industry codes where a GICS label is missing. A name we cannot place in a sector at all gets no slot under either constrained mode — a sector quota over a non-sector would be noise dressed as neutrality — and the number of names that fell out for that reason is published with the book rather than absorbed.

A filer that has stopped filing publicly is excluded from a build outright, by name, in the response: a dead book must never be blended as though it were current. A filer merely running a quarter behind is included, and its lag is reported beside it.

The response also states what share of each filer's 13F dollars actually mapped into our universe. The built book reflects the mapped slice, and that share belongs next to the result rather than in a footnote nobody can compute.

The point-in-time backtest, on the copier's convention

The backtest rebuilds the strategy from the as-filed books at every stored quarter and chains the returns. Each quarter's book goes on at the 13F filing deadline for that quarter — the earliest date on which somebody copying these filers could actually have acted — and is held until the next rebuild. That is the copier's convention, and it is what keeps the exercise from crediting the strategy with a quarter it could not have traded.

Factor ranks run on our stored point-in-time metric archive wherever a snapshot exists from on or before the date a quarter was applied; one basis per quarter, never mixed. Quarters the archive does not cover fall back to today's data, and that look-ahead is disclosed first in the notes, which also name exactly which quarters ran true point-in-time.

The page prints one word for that basis beside the return. It says verified only when every quarter ranked on filings knowable at the time and the construction itself carries no look-ahead: a benchmark sector target steers by today's index membership at every historical rebuild, so a benchmark-mode backtest is never verified however clean its factor snapshots are.

Known limitations of the backtest, disclosed with every run: it is long-only and priced names only, and a name we cannot price at an application is dropped with the weights renormalized; a holding that delists mid-period is floored at a total loss; a quarter with no application contributes no return rather than being bridged by the previous book; fact values are the latest restated vintage, so restatement bias is present; and the universe is survivors-only.

What this is not

Nothing here is investment advice, and no causal claim is made from any filer's holdings. A book built from 13F filings is a lagged, long-only, US-listed view of what a set of managers reported — not what they hold today, and not a fund's returns.

Sources

SEC bulk Form 13F data sets (holdings and filed dates); SEC XBRL company facts and our licensed end-of-day price store for the fundamentals and momentum; index constituent files for the benchmark targets. All are public regulatory data or licensed market data — see the data-sources page for licensing notes.