Model
How the Leadgap score is built, and how it is behaving on live data right now. A heuristic ranking, not a proof of causality or a forecast.
- Scored now
- —
- Candidate rate
- —
- Median mark captured
- —
- Model version
- heuristic-v5
Score composition
Score = 100 × the product of six factors, each between 0 and 1. Because they multiply, one weak factor pulls the whole score down. Bars show each factor’s live average on the 4h window.
- Perp lag (residual magnitude)—
The gap against two of the perp’s typical moves over the window. Capped at 1.
- Implied-move magnitude—
1 when the odds-implied move beats the perp’s by 25%, 0.42 in line, 0.12 when the perp led.
- Mapping confidence—
How directly the event names or clusters with the perp.
- Volume / liquidity—
log₁₀ of event volume over 6, capped at 1.
- Odds movement—
1 for moves ≥ 0.8 pts, 0.65 for ≥ 0.3 pts, otherwise 0.3.
- Sanity damping—
Pulls down implausibly large odds or implied moves.
Live signals by score band · 4h
Candidates require a score of at least 28, a clear side, less than 85% catch-up, measured odds leadership and a gap above estimated costs. Missing evidence cannot qualify.
n = 0 live comparisons. Historical calibration — how often the mark later closed the gap — is research-only and not published until the archive is large enough.
How odds become a move
Price-threshold markets — “above $X on a date”, “reach $X by a date” — are options on the perp itself. With the strike fixed, a change in Yes probability maps to the price move that would explain it, given time to expiry and an assumed volatility (for a digital, ln S = ln K + σ√τ·Φ⁻¹(p)). Odds that drift as expiry nears are not read as price moves. Markets priced below 5% or above 95%, or resolving within the window, are left out: there a one-tick wobble or time decay would read as a large move. Clear negative outcomes such as “will not reach” use the complement probability (1 − Yes). Compound conditions and unsupported negations are left out.
Supported events only. The selected question must name the underlying and express a supported price threshold with an expiry. News, mentions, macro outcomes and relative-performance questions are excluded. The parent event title cannot establish a link.
Model v5 adds strict eligibility, measured timing and quote evidence. Saved alert settings remain intact; changing the model does not create an alert crossing. Historical v4 replay remains separate.
Known limitations
- Volatility assumptions and the threshold pricing model can be wrong. A supported price relationship does not establish mispricing.
- Timing compares minute returns over up to 60 minutes at lags from −5 to +5 minutes. It needs 12 pairs, 80% coverage, correlation ≥ 0.4 and a 0.1 advantage over zero and opposite lags. The 1m and 5m windows cannot qualify. Correlation does not establish causality.
- Upstream feeds can be delayed or interrupted. Observations older than 90 seconds are dropped rather than scored, so coverage can thin out.
- The residual is gross. The separate $100 notional / 30-minute benchmark estimates spread, depth, taker fees and adverse funding. It is research evidence, not an executable order or profit forecast.
- Any calibration will be measured on a small, evolving sample and can shift as the model or coverage changes.
New to the gap? Read the guide or take the Signals tour.