Research through the 25 August 2026 close
Glossary · 55 defined terms
Archetypes, regime labels, conviction tiers. Each definition is referenced by the dossiers and the daily journal. Citable, dated, stable.
Quality balance sheet, secular tailwind, multi-year hold candidate.
Names with durable earnings power and a long runway. Sizing is generous, stop discipline lenient, holding period measured in quarters or years. Earnings beats compound rather than expire.
Mean-reverting earnings, regime-sensitive entry.
Names that win when a specific macro regime turns. Entry is gated on the macro setup, sizing scales with cycle conviction, exit is regime-state dependent rather than price-target dependent.
Highest-conviction name within an active theme cluster.
The cleanest pure-play within an established theme cluster. Sized higher than basket peers, but downgraded the moment the theme matures or saturates.
M&A, spin-off, restructuring, regulatory event.
Event-driven names where the catalyst is known and dated. Sizing is event-aware, exits are tied to event resolution rather than price action.
Pre/post-print setup with an explicit binary outcome.
Names whose price gates on a single quarterly release. Entry pre-print only on confirmed bullish setup; auto-exit windows around the print.
High-beta, short-interest-driven; hard sizing cap.
Squeeze-candidates and meme-style basket movers. Hard 1% sizing cap regardless of conviction, auto-trim on RSI > 82, no averaging down.
Cash-flow durable, low-beta, regime hedge.
Staples, utilities, dividend aristocrats. Used as a regime hedge during RISK-OFF; sized inversely to cyclical exposure.
Cross-asset proxy for a thematic risk (XLE / GLD / TLT / …).
An ETF or single-name macro expression of a thematic view, held to balance portfolio beta rather than for standalone alpha.
Risk-asset tailwind across credit + rates + breakevens.
Buy-thresholds eased, size multiplier 1.0, cash floor 5%. The model bids its strongest-graded names freely.
Mixed signals, no directional conviction.
Default neutral lean. Buy-thresholds neutral, sizing standard, position count capped tighter. Adds only on the strongest-graded setups.
Risk-asset headwind across the rate / credit / breadth complex.
Buy-threshold tightened, size multiplier 0.5, cash floor 30%. Only the strongest-graded names pass; the retail-squeeze archetype is switched off.
A single binary event is gating regime direction.
Used when one near-term event (CPI, NFP, Fed) will resolve regime ambiguity. Position sizing halved across the board until the event resolves.
Inflation expectations rising into a slowing-growth read.
Energy + materials hedges favoured; long-duration assets deprioritised. Cash floor raised; defensive sizing increased.
Recovery from RISK-OFF without full risk-asset confirmation.
A transitional read — partial easing of RISK-OFF tightening. Sizing graduated, buy-threshold loosened in steps.
The model's highest calibrated confidence — reserved for the cleanest setups.
The top tier. Reserved for the cleanest setup × balance sheet × narrative combinations. Sizing leans larger; stop discipline still firm.
Strong setup × narrative × balance-sheet alignment.
Standard target conviction for entry. The bull case is sourced and the invalidation trigger is concrete and near-term checkable.
Setup intact but at least one bull-case bullet is contested.
Probe-only conviction. Smaller sizing, tighter stop, accept higher kill rate.
Partial setup — one of three confirmation factors present.
Watchlist-grade or sub-1% probe only. Often paired with an explicit waiting condition (e.g. higher-low to confirm).
An internal read on each candidate, strongest to weakest, before sizing.
The model grades each candidate on narrative, balance sheet, and setup alignment — strongest to weakest. The strongest grade is reserved for names where all three clear independent thresholds. You mostly see its effect through the conviction tier published on the dossier.
The strongest quality grade — narrative, balance sheet and setup all clear.
Reserved for names where the bull case, the balance sheet, and the setup each clear an independent bar. It maps to HIGH or SUPREME conviction depending on the regime.
A passable read whose case is contested on at least one front.
The balance sheet has unresolved questions, the setup is partial, or the narrative is mature. Maps to LOW or MEDIUM conviction; probe-only sizing.
The one-paragraph bull-case summary on every dossier.
What the bet is, in plain English, anchored in sourced bullets from the bull-case body. Updated each time a dossier is re-analysed; the version date is published.
The explicit kill criterion published with every pick — formally, the invalidation trigger.
A pre-committed price, news, or event that, if it fires, strips the conviction and marks the pick broken. Functions as a public commitment, not a soft warning.
Dated events the dossier expects to move the name.
Earnings dates, regulatory deadlines, sector catalysts, FOMC, CPI. Each entry is dated and references its source. The next-30-days slice is shown on the dossier page.
A group of correlated names that move together on shared catalysts.
Themes are treated as primary, not afterthoughts. A name's basket co-movement is read alongside its standalone setup; cluster confirmation can lift a single-name conviction. Examples: Quantum computing, AI datacenter infrastructure, Squeeze & momentum setups.
Independent corroboration of a thesis by basket peers moving on the same catalyst.
If the theme leader and at least two basket peers all confirm the same direction on the same headline, the system upgrades signal weight. Cluster failure on the same headline downgrades it.
A confirmation pattern: price prints a low above the prior low on expanding volume.
The confirmation condition for retail-squeeze setups: the gap itself is not the signal, and the setup is unconfirmed until a higher-low holds above the catalyst base.
The price-structure read: moving averages, RSI, levels, basing pattern, volume.
Moving averages, RSI, basing pattern, volume profile, levels. The dossier's 'Setup & Price Structure' section captures it in plain English.
A published call — thesis, conviction tier and kill line — that can be graded.
The unit FrontierPicks is scored on. A pick is a dossier's thesis plus the conviction tier it stakes and the kill line it publishes in advance. It resolves exactly two ways — played out, or invalidated — and every resolution stays on the scoreboard, wins and losses alike. A call with no published kill line is not a pick; it can never be graded, only quietly forgotten.
Every name the model considers each day: held positions, watchlist, and the top-100 momentum screen.
Held positions + watchlist + the top-100 momentum names that pass the liquidity screen. Refreshed every premarket. The breadth keeps narratives from being missed.
The current regime call's mapping to sizing / threshold settings.
Each regime (Risk-on, Choppy, Risk-off, …) maps to a specific set of settings — buy-threshold, size multiplier, max exposure, and cash floor. The journal records the active tier daily.
The FrontierPicks publishing principle: the reasoning, the names, and the record are all public.
Dossiers, regime calls, the names held and circling, archetype/regime classifiers — all published with source links and dated, the day each call is made.
The names FrontierPicks holds now plus the names it's circling — research attributes only.
The held positions and watchlist published as names with conviction, archetype, theme, thesis and a kill line, the day each call is made. A live, reasoned book instead of a logo wall published a quarter late.
The public scoreboard: every thesis resolved as played-out or invalidated, dated and scored.
Each thesis ships a falsifiable kill criterion; when it resolves it lands in the ledger as played-out or invalidated, dated, with whether the published trigger fired. Strictly non-monetary: it scores whether each published claim held up, leaving aside trade timing, sizing and P&L entirely.
The share of resolved picks that played out (vs were invalidated).
Among picks that have resolved, the fraction marked played-out rather than invalidated. Reported overall and broken out by conviction tier, archetype, and the macro regime in force at resolution.
The forecast-accuracy score on the FrontierPicks conviction calls, decomposed into reliability, resolution and the question set's own uncertainty.
Mean squared error between the probability a conviction tier implies (SUPREME≈0.9 … LOW≈0.5) and the binary outcome. Decomposed (Murphy) into calibration, resolution and uncertainty. For external context, Tetlock's Good Judgment Project superforecasters scored ≈ 0.08 on geopolitical questions, and published forecasting evals put the strongest LLMs near ≈ 0.10 — both on different question sets, so they are reference points, not a like-for-like comparison. The FrontierPicks measured score lives on /track-record/.
Whether stated conviction matches reality — does SUPREME actually play out far more often than LOW?
A model is calibrated when its conviction tiers play out at the rates they imply. FrontierPicks publishes the gap between claimed and observed play-out rates per tier (the reliability diagram), so confidence is checkable rather than asserted.
Whether staking conviction adds information over forecasting the base rate (1 − Brier ∕ uncertainty). Zero means the tiers don't yet separate outcomes.
The Brier skill score asks whether the conviction tiers add information over a model that simply forecasts the overall play-out rate every time. A positive score means the tiers discriminate; zero means they do not yet. The FrontierPicks skill score sits at the base-rate line on the current sample — a statement about sample size, not a result in either direction. See /skill-or-luck/ and /track-record/.
How much the conviction tiers separate outcomes from the base rate — the discrimination term in the Brier decomposition. Higher is better.
Not to be confused with a thesis resolving played-out or invalidated. This is the discrimination term in the Murphy decomposition: the degree to which SUPREME, HIGH, MEDIUM and LOW actually sort outcomes apart instead of all reverting to the base rate. Near-zero resolution means the tiers don't yet separate — confidence may track reality (good calibration) while conviction still carries no extra information.
The plot of predicted versus observed play-out rate per conviction tier; the diagonal is perfect calibration.
The artifact FrontierPicks renders on /track-record/. Each dot is one conviction tier plotted at its implied probability against the rate at which its theses actually played out; the diagonal line marks perfect calibration. Dots above the line are overconfident, below it underconfident, and the spread between dots is what resolution measures.
How often theses play out overall, before any conviction is staked — the benchmark a skill score has to beat.
The unconditional play-out rate across every resolved thesis. A forecaster who simply predicted this number every time would score the uncertainty term and nothing more; the base rate is therefore the floor that staking conviction must clear to demonstrate skill. The uncertainty term in the Brier decomposition derives directly from it.
A claim is only meaningful if some observation could prove it wrong — the principle behind every dossier's kill line.
Popper's criterion: a claim that no possible outcome could refute explains nothing. Every FrontierPicks dossier ships a pre-committed kill line precisely so the pick can be marked broken in public, and that falsifiability is what makes the scored record possible — a claim you can't disprove can't be Brier-scored either.
Judging a record only by its survivors — the distortion an append-only ledger, where nothing is ever retired, is built to defeat.
When losers are quietly dropped and only winners are counted, any track record looks skilled. It is why most 'AI beats the market by X%' claims are unfalsifiable. The FrontierPicks structural answer is an append-only ledger in which every published pick is in exactly one of three states — it reached its case, it hit the kill line it published in advance, or it is still undecided — and all three are published as counts. Nothing is retired, and the invalidated picks stay on the board. Whether the undecided remainder skews toward losers is measured against the universe those picks were drawn from; see /how-picks-resolve/.
A simulated, non-live account: nothing to sell, no book to pump, no order to front-run — so the incentive that bends most research isn't present.
The FrontierPicks pipeline runs on a simulated account by design. Because no real capital changes hands, there is no position to talk up, no order flow to front-run, and no product to sell off the back of a call. The conflict of interest that distorts most sell-side and influencer research is structurally absent, which is part of why the record can be published in full.
Splitting the Brier score into calibration, resolution and uncertainty (Brier = calibration − resolution + uncertainty).
Allan Murphy's identity decomposes the Brier score into three readable parts: calibration (how far claimed confidence sits from observed rates), resolution (how much the tiers separate outcomes), and uncertainty (the irreducible difficulty of the question). A low calibration term paired with near-zero resolution is exactly what a skill-≈0 record looks like — confidence that tracks reality, conviction tiers that don't yet add information. FrontierPicks surfaces all three on /track-record/.
Theme state: attention is arriving for the first time — participation widening, headline flow fresh. Carries the widest theme cap the sizing rules allow.
The earliest of the four states a theme can be in, and the one that buys the most room: an accelerating theme is allowed the largest share of the book across its member names. It has to be dated rather than asserted — a named catalyst, a cluster of coverage that did not exist a fortnight ago, a class of buyer that was not there before. It is falsifiable in the ordinary way: if the attention thins while price holds, the read moves to maturing and the cap tightens with it.
Theme state: well known and still working, but the flow behind it has moderated — coverage repeats rather than expands. The cap tightens.
The story is understood, the names still get a bid, and the headlines have stopped being new — target revisions on maintained ratings rather than fresh initiations. What separates it from saturated is that participation is still there: a maturing theme is being held, a saturated one is being distributed. It is the most common state in the corpus, which is what you would expect of a book built out of narratives that have already been discovered.
Theme state: coverage has gone mainstream and late-cycle and the marginal new buyer has thinned — everyone who would act on the story already has.
Late-cycle attention: retail-facing and retrospective coverage, sell-side targets clustered at or below the price, a beat that no longer moves the tape. The story can still be true and the business can still be compounding — saturation is a statement about who is left to buy, not about the fundamentals. It carries a much tighter theme cap than maturing, and the state that follows it, if the price structure goes, is dead.
Theme state: the narrative failed or its price structure broke. The sizing rules stop the model adding to the theme rather than treating it as cheap.
The terminal state, and the one most often misread. It labels the NARRATIVE and the price structure, not the company: a theme can be dead while the business behind it grows, which is exactly the case the dossiers keep having to spell out. What ends a theme is a catalyst that fired and produced nothing, or a level that was published in advance and then lost. Because a dead theme takes no further exposure, this is the one state that changes what the model can do rather than only how much.
The name is on the live book right now: a published thesis, a conviction tier, and a kill line that is armed.
The smallest of the four states. A held name has cleared the liquidity screen, been graded, and had a falsifiable kill line published in advance — which is what makes it scoreable later. The dossier says what the bet is and what would end it; it does not say what was traded, at what level, or in what size, and no surface on this site does.
Circled but not on the book: the thesis is written and the kill line published, and the setup has not cleared.
A watchlist name is fully researched — the same dossier, the same conviction tier, the same pre-committed kill line as anything held. What is missing is the setup, which is why so many watchlist theses carry an explicit waiting condition. Publishing them is the point: a call that only becomes visible after it works is not a record, it is a highlight reel.
The name has come off the live book, and its dossier and its published line stay up.
Nothing is retired when it stops being held. The thesis, the conviction that was staked and the kill line that was published all remain on the page, and the pick keeps resolving into the record — reached, hit its line, or still undecided — exactly as it would have if the name were still on the book. That is the structural answer to survivorship bias: the exit changes the status pill and nothing else.
Researched, published, and not currently being tracked forward — by far the most common state in the corpus.
The universe is far larger than the book: every name that clears the liquidity screen can earn a dossier, and most of them are never held. A dormant dossier is a real published read with a real date on it, left standing rather than deleted. It is also why the corpus is mostly dormant names — the alternative, publishing only what is owned, is the thing this site exists not to do.
A published pick that reached neither exit — not its 1:1 target, not its own kill line. A normal expected state, not a missing grade.
Every published pick stands in exactly one of three doors: it reached its case, it hit the line it published in advance, or it is still undecided. The three add up to the published total by construction, and that identity is what makes the record checkable instead of a survivor set — an undecided pick is counted in public rather than quietly dropped. A thesis is a conditional claim with two exits, and a name that goes sideways reaches neither; a resolution market sits in the same state before its question closes.
Undecided, with no forward commitment: the published line still stands and still resolves if it fires, but the name is not being tracked toward it.
The undecided set splits in two, because collapsing the halves would be its own evasion. A live pick is still on the active board with an armed kill line. A lapsed one is not — attention has moved on — and yet the line it published stays on the dossier and lands in the record if it ever prints. The commitment outlives the attention, which is the only version of this that is honest: the alternative is to stop counting a claim the moment it stops being interesting.
The explicit kill criterion published with every pick — formally, the invalidation trigger. A pre-committed price, news, or event that, if it fires, strips the conviction and marks the pick broken. Functions as a public commitment, not a soft warning.
The one-paragraph bull-case summary on every dossier. What the bet is, in plain English, anchored in sourced bullets from the bull-case body. Updated each time a dossier is re-analysed; the version date is published.
An internal read on each candidate, strongest to weakest, before sizing. The model grades each candidate on narrative, balance sheet, and setup alignment — strongest to weakest. The strongest grade is reserved for names where all three clear independent thresholds. You mostly see its effect through the conviction tier published on the dossier.
The model's highest calibrated confidence — reserved for the cleanest setups. The top tier. Reserved for the cleanest setup × balance sheet × narrative combinations. Sizing leans larger; stop discipline still firm.