Trust · Methodology
How Navurex turns match data into a probability.
The core probabilities are produced by Navurex model code. AI is used to explain supplied model/data signals in language—not to invent the core 1X2 probability.
01 · Inputs
Recent results are weighted by competition.
Navurex uses up to five recent completed matches for the core recent-form layer. Prediction Engine V2 also scans the cached recent history to build up to five home-only matches for the home team and five away-only matches for the away team. Newer matches receive more weight, while a friendly still influences the model far less than a competitive league match.
02 · Rating
Five rating dimensions.
V1 first builds the canonical team rating from recent form (30%), attack (20%), defence (20%), goal balance (10%) and league strength (20%) when standings are usable. Contextual V2 then applies a conservative adjustment from time-decayed recent performance, home/away venue form and opponent strength when that opponent can be matched to the current league table. The contextual rating is capped so a small sample cannot replace the canonical rating.
League strength itself uses current position, points per game and goal difference per game. At the beginning of a season, standings may be unavailable or less informative.
03 · Probability
Ratings become a 1X2 distribution.
Navurex converts the two overall ratings into home-win, draw and away-win weights. The draw weight rises when the teams are closer together. Published values are normalized to 100% and guardrails prevent broken/NaN inputs from leaking into the interface.
Model Confidence is not the same as Data Quality. Confidence reflects how separated the model outcomes are; Data Quality describes whether the underlying match sample is strong enough to publish confidently.
04 · Data quality
The model can say “not enough data.”
A prediction is unavailable when the effective sample is below one match or one side has no competitive recent match. The V2 completeness gate also checks provider availability, venue-specific sample size, league-table availability and opponent-strength coverage. Stronger samples can be labelled High or Medium; smaller samples are explicitly marked Limited data. Friendlies remain visible as context but contribute only 25% model weight.
05 · Simulation
10,000 deterministic Poisson simulations.
Expected goals blend competition-weighted recent scoring, opponent defending and the current contextual Navurex attack/defence ratings. Navurex then runs 10,000 seeded Poisson simulations. The same match and same model inputs produce the same simulation sequence for the same model version, which makes results reproducible rather than visually random on every refresh.
06 · Corner Intelligence
Corners use a separate fixture-statistics model.
Corner Intelligence does not reuse the 1X2 probability as a shortcut. Navurex reads recorded Corner Kicks from up to five recent completed matches per team, applies the same competition weighting principle, and calculates weighted corners-for and corners-against profiles. A team projection blends its own recent corner generation (55%) with the opponent's recent corner allowance (45%). Smaller samples are shrunk toward the pooled recent-match baseline so one unusual fixture cannot dominate the public number.
The displayed total-corner percentages use a Poisson-style count approximation around the projected total. If either team has too little recorded corner coverage, Navurex marks the module Limited or Unavailable rather than filling missing statistics with zero.
07 · V3 research layer
V3 is testing a broader evidence model before promotion.
V3 Shadow evaluates up to 20 completed matches with recency decay, keeps the latest five as a separate short-term signal, adds schedule fatigue and evidence-based mean reversion, compares common opponents to bridge teams across competitions, reduces the league-level prior, and can calibrate against no-vig bookmaker consensus.
Streaks are not treated as due to reverse. A rebound/cooling adjustment only appears when short-term performance materially diverges from the longer baseline and supporting evidence such as fatigue is present.
V3 remains shadow-only by default. Its pre-match probabilities are frozen and compared with V2 using Brier score, log loss, calibration and pick accuracy before any production switch.
Frozen history
Predictions should not change after the result.
Before kickoff, Navurex can save the current prediction and simulation as a versioned snapshot. After kickoff, a valid stored snapshot becomes the source of truth for public performance tracking. Older matches without a snapshot may be reconstructed only from data available before that kickoff and are identified separately in the model pipeline.
View model performance →Model scope
Core rating vs. match context
Navurex separates repeatable model inputs from external match context so a missing injury report or late lineup change is never silently presented as if it were already inside the rating.
Included in the core rating
Recent form, time decay, home/away venue form, opponent strength when available, attack, defence, goal balance and league strength when usable. These inputs are structured and repeatable. Contextual V2 is blended conservatively with the V1 baseline rather than replacing it outright.
Not guaranteed in the core rating
Injuries, confirmed lineups, tactical changes, transfers, weather and head-to-head history may be incomplete, delayed or absent. They should be treated as separate context until Navurex can verify and model them consistently.
Current limitations
Navurex does not currently guarantee complete injury, confirmed-lineup, tactical, transfer, weather or head-to-head context in the core rating. Model inputs can be incomplete or delayed. Probabilities are estimates, not certainties, and are not betting advice.