Geo / SKILL / Ensemble
Geo-kinetic priors and optional XGBoost residuals produce xGF / xGA only.
Modes: geo (primary), residual, or ensemble. Reports show the
skill_mode that was used — L2 never prices 1X2.
Quant stack
This page explains how the WDI pre-match stack thinks — layer contracts, core pricing algorithms, and the publish surface. The website is explanatory and read-only: it does not run the engine in the browser.
Pricing chain (final matrix) Cθ(Fν(λ), Gν(μ)) →IPFcurrent →1X2 / EV opening →COM-Poisson+Frank invert (ν=1 then re-fit at final ν) · physics · Copula · current IPF
Two compute layers feed an immutable publish layer. Boundaries are intentional — so fundamentals never silently become prices.
Geo-kinetic priors and optional XGBoost residuals produce xGF / xGA only.
Modes: geo (primary), residual, or ensemble. Reports show the
skill_mode that was used — L2 never prices 1X2.
Match physics synthesizes intensities (λ, μ); COM-Poisson ν shapes scoring tails on the final joint matrix. Frank Copula couples the sides; IPF soft-anchors to current de-vigged margins; tri-track EV selects ANCHOR / ALPHA / ROCKET.
Board, per-match JSON, and posters are frozen artifacts. Free: 1X2 bands + track presence. Pro: picks, scorelines, λ/μ, and fundamentals context.
The six ideas that define WDI pricing — what each does, and what it deliberately does not do.
Links home and away goal counts into one joint scoreline matrix — not two independent Poissons glued together. Dependence strength θ is shared between opening-market invert and the final matrix build.
Does not: invent 1X2 from XGBoost directly.
On the final matrix, asymmetric ν controls how fat or thin each side’s scoreline tails are under chaos, fatigue, or cup context — before the Copula couples the sides. Opening invert and the joint matrix share the same COM-Poisson marginal helper; when market-anchored, λ/μ are re-fit under the final ν so margins stay consistent.
Does not: replace the Copula joint step.
Shin removes bookmaker overround on 1X2; Power de-vig covers Asian handicap and totals (and is the Shin fallback). Cleaned market probabilities feed IPF and EV comparison — model 1X2 itself still comes from the Copula matrix, not from the book sheet alone.
Does not: claim every book is a sharp reference, or replace model pricing.
When opening 1X2 + totals imply λ/μ (COM-Poisson+Frank invert; first pass ν=1), physical modifiers are η-blended toward that market base. After final ν is set, λ/μ are re-calibrated under that ν so invert margins match the Copula build.
Does not: hard-overwrite λ/μ with raw book prices.
Iterative Proportional Fitting gently reshapes the Copula matrix toward de-vigged current 1X2 and totals margins when those lines are present — soft blend, audited weight. Opening is used for λ/μ anchoring; current is used for IPF.
Does not: discard the model matrix for a pure market sheet, or IPF from opening alone.
ANCHOR is the production spine; ALPHA hunts medium-volatility value; ROCKET is high-volatility research. Confidence stars are a recommendation index — not a stake %.
Does not: show Kelly % or bankroll advice on this site.
These modifiers shape λ/μ before Copula pricing. They prove the stack is not a pure market classifier — and when an opening anchor is present they are η soft-coupled, not hard overrides.
Applied upstream in publish (flight / altitude) before L1 sees the match.
Floor at 1000. When travel_km > 0 is applied here, that side sets
travel_distance_via_elo so L1 tax does not re-count km.
Separate L1 channel: timezone, rest, and age logistics. Distance km is included only when Virtual Elo did not already own that side’s flight distance.
League-aware home boost on intensity. When market-anchored, blended with η instead of full override.
Tanh-bounded growth / decay on λ and μ so extreme priors cannot explode the scoreline grid.
Error-exploitation multiplier, squad-depth elasticity, and recovery ψ — favorite pressure and fatigue nuance.
Spear / shield counters from formation pairs. Often off for club leagues until formation ingest is trusted.
Physics explains intensity — it does not replace Frank Copula, IPF, or tri-track selection.
Three research weights on the same fair matrix — not three equal stake recommendations.
Confidence stars are a recommendation index, not a stake instruction. Half-Kelly exists only as an internal ceiling for stars and portfolio caps — never shown as a bet size.
find_ev.Published fixtures often carry soft-book closing prices. CLV figures stay labeled soft-book until a Pinnacle (or mixed) reference is present on the artifact. Hub and Report surfaces show an Odds chip so the reference is never ambiguous.
Under the default calibration policy (display_only / ADR-0015),
on-screen 1X2 bars may use calibrated margins while ANCHOR / ALPHA / ROCKET value bets
always price from the joint score-matrix margins.
Published reports expose pricing_1x2 (matrix) beside display
prob_*, plus a physics_audit block
(market-anchored flag, η, favorite side).
Top scorelines use the full 16×16 joint. Any UI heat grid is an 8×8 slice — under fat tails the visible mass can sum to less than 1 without implying a bug.