Methodology

This page is the constitution of Longview's research process. Its hash is pinned in the chain's genesis event; any change to it requires a signed, public METHOD_CHANGE event. Version 1, 2026-07-24.

What a thesis is

A Longview thesis is a falsifiable, long-horizon (5–10 year) claim about a company positioned inside an emerging industry. Every thesis must contain, before it can be signed: a thesis statement of at most three sentences; evidence of inflection already visible in filings; a business-quality assessment; a financial base cited to SEC filings with as-of dates; a scenario valuation; at least three pre-registered kill conditions; a risk register; its position within the theme; and an explicit benchmark statement. A thesis missing any section cannot enter the chain — the format is the quality gate.

Scenario valuation — ranges, never targets

We publish present-value ranges, not point targets. Each scenario is fully parameterized: a five-year revenue growth path, a terminal margin or exit multiple, annual share dilution, and a 10% discount rate. The formula is deliberately simple:

  • Year-5 enterprise value = TTM revenue × (1 + growth)⁵ × exit multiple (or × terminal FCF margin × P/FCF multiple)
  • Year-5 equity = enterprise value − net debt (held flat — a conservative simplification)
  • Year-5 price = equity ÷ diluted shares grown at the dilution rate
  • Present value = year-5 price ÷ 1.10⁵

Scenario probabilities must sum to one and include a loss case where the risk profile demands one. Every thesis publishes its full sensitivity grid (growth × exit multiple) so a reader can find the cell they believe in rather than accept ours.

Publication gates. A thesis publishes only if the base case implies a 5-year IRR of at least 15% per year, and the probability-weighted expected value is at least 1.3× the drafting price. Below those bars the market's own ~8–10% does the job with less risk, and the work becomes a watchlist note instead.

Grading. Ranges are graded at the 1-year checkpoint (did the price land inside the published range?) and at the 5-year horizon. The running in-range percentage is our calibration score, published as prominently as returns. Honest ranges — not ranges wide enough to never miss; the range widths are published too.

Entry prices — no lookahead, by construction

The entry price for every call is the first market close strictly after the publication timestamp. Daily data cannot see intraday, so a thesis published on a trading day gets the next day's close — entry can only ever be later than publication, never earlier. This rule is enforced in code and pinned by a regression test on the exact path that runs.

Benchmarks

Every thesis is compared against SPY and QQQ over the identical window, and the headline number is the equal-weight basket of all open theses versus QQQ — the number a skeptic would compute. Version 1 uses price-return series for both sides of the comparison (dividends excluded from theses and benchmarks alike); a total-return upgrade will arrive as a signed METHOD_CHANGE.

The scoreboard rules

Every call we have ever published stays on the board permanently — closed and killed theses are graded to their close date and never removed. Forbidden forever: cherry-picked windows, "best picks since inception," excluding "old methodology" calls, annualizing anything under a year.

Theme selection

A theme must pass five gates before any thesis in it can publish: a structural driver with a 5–10 year runway; revenue inflection already visible in filings (narrative alone fails); at least three listed pure-plays with majority revenue exposure; a trackable adoption metric checkable quarterly; and live disagreement in the market. Themes that fail are published too, with the failing gate named.

Reviews and corrections

Every open thesis is reviewed within ten days of its quarterly filing: kill conditions checked, the theme's adoption metric updated, and the review signed into the chain. A triggered kill condition forces a close with a post-mortem. Corrections are signed CORRECTION events — struck through, dated, never deleted.

The role of AI

AI drafts research sections from filing data, with citations. A human reviews, edits, and signs every thesis; nothing enters the chain or touches the scoreboard without human sign-off.

The chain

Every event is Ed25519-signed and hash-linked to its predecessor: chain_hash = sha256(prev_hash | payload_sha256 | type | ts_utc | seq). Payloads are content-addressed JSON. The full chain and every payload are published at /chain.json; verification runs in your browser at /verify or offline against the same data. What it proves: the record is append-only. What it doesn't prove: that we're right — the forward scoreboard does that, either way.

Methodology v1.1 — the engine (adopted 2026-07-29)

Version 1.1 adds an evidence-graded selection engine around the original gates. Every rule below names its evidence; anything popular that failed rigorous tests is listed as decoration and banned as justification. Items marked rolling out are adopted policy whose computation is being wired into the daily tick; they apply to all new calls as they land.

The funnel

Listed universe → themes with filing-proof of inflection (five gates, unchanged) → pure-plays only → a quality gate → the valuation gates → the board. The quality gate runs before valuation (rolling out): gross profits over total assets in roughly the top third of the market (Novy-Marx 2013 — the single best-evidenced quality signal), and return on invested capital of at least 15% in seven of the last ten years (Mauboussin's persistence base rates: elite profitability persists, growth does not — Chan, Karceski & Lakonishok 2003 shows long-term growth persistence is no better than chance, so we never underwrite a decade of extrapolated growth).

Red-flag vetoes

Any one of these blocks a compounder tag regardless of the story, each computable from filings: trailing three-year net share issuance above 2%/yr (Pontiff & Woodgate 2008); asset growth persistently outrunning gross-profit growth (Cooper, Gulen & Schill 2008); accruals above 10% of assets (Sloan 1996); three-year median cash conversion below 80%. A catastrophic-decline profile — drawdown over 60% plus an issuance spike plus negative free cash flow — can never be actionable (rolling out).

Two kinds of winners

Bessembinder (2018): roughly 4% of stocks created all net wealth over Treasury bills; the median stock lost money. So the engine hunts two distinct classes and refuses to blur them. Compounders — elite, persistent profitability at a defensible price; the greatest compounders in history returned ~13.5%/yr held for decades, so the edge is holding through boredom, not finding rockets. Moonshots — small companies ($100M–$2.5B) bought cheap (P/S ≤ 1.5, gross margin ≥ 30%, positive free cash flow) with open-ended upside: the actual profile of historical 10-baggers (Yartseva 2025 — median starting cap ~$348M, P/S 0.6), not lottery tickets. Lottery traits are a measurable tax, and each is a veto: top-decile single-day spike in the last month (Bali, Cakici & Whitelaw 2011), sub-$5 price, OTC listing, dilution above 5%/yr, top-decile idiosyncratic volatility. Moonshots only work as a basket: small positions, many names, ten-year holds, no stop-losses — a 50%+ drawdown along the way is expected, not a thesis break.

TAM discipline

Any thesis leaning on a total addressable market must compute the terminal market share the valuation implies (rejected above ~20–25% for a non-leader), never model >20% revenue growth beyond year five without named evidence, and check that implied shares summed across our own covered names in a theme stay below 100% — the Big Market Delusion test (Cornell & Damodaran 2020).

Timing: valuation decides what, trend decides when

Momentum is the one technical signal that survived rigorous testing (Jegadeesh & Titman 1993; expect roughly half the published premium going forward — McLean & Pontiff 2016). So a call may only move from Watch to Actionable when price crosses our line and the tape is not a falling knife: 12-1 momentum positive or price above its 10-month average (rolling out). Escalations are suppressed in momentum-crash conditions — index in top-decile volatility and below its 10-month average (Daniel & Moskowitz 2016). Decoration, banned as justification in any thesis: RSI, Fibonacci levels, Elliott waves, chart patterns, moving-average crossovers as alpha (Sullivan, Timmermann & White 1999), and sell-side price targets — only ~38% of 12-month targets are ever met (Bradshaw, Brown & Huang 2013).

Portfolio construction (published as method, never advice)

Concentration with conviction has evidence (best-ideas literature); concentration without rules is how skew kills. Our published sizing method: no position above 10% at cost; positions above 5% may not sum past 40%; no more than 25% in one correlation cluster (pairwise 252-day correlation above 0.6 = same cluster); moonshots capped at 1–2% each inside a basket of many; sizing anchored at quarter-to-half Kelly because full Kelly is ruinously sensitive to estimation error (MacLean, Thorp & Ziemba). Rebalancing by bands, not calendar; winners trimmed reluctantly. Every call already carries pre-registered kill conditions — the sell rules are sealed before the buy case can be.

Forecast discipline

Fair values are bands, never points; the actionable price is set at the conservative end. Scenario assumptions are graded against base rates (rolling out: a reference-class table built from EDGAR revenue histories). Every probabilistic claim in a thesis gets a resolution date and a Brier score in the sealed record (rolling out) — calibration is measurable and trainable (Tetlock). And every signal the engine uses is published with its post-publication decay haircut — anomalies lose roughly 26% of their edge out-of-sample and 58% after publication (McLean & Pontiff 2016); anything whose rolling edge hits zero gets retired, publicly.