Match & team data
Available information is collected before kickoff and bound to the correct fixture identity.
Inside ScoreProb
How ScoreProb turns football data into probabilities
ScoreProb is a probability system, not a conventional tips engine. It combines validated pre-kickoff football information with engineered statistical features and trained models to publish probabilities that can later be measured against real outcomes.
362 engineered inputs Pre-kickoff only Certified publication Historical auditing
From football data to published probability
The public methodology shows the process and safeguards. Exact model weights, transformations and training implementation remain proprietary.
Data validation
Invalid, missing or non-finite information is handled explicitly rather than silently becoming a normal value.
Feature engineering
Raw information is converted into structured football and statistical signals suitable for model inference.
362-feature state
The certified feature vector captures strength, form, venue, scoring, defending, timing, stability, schedule and other context.
Model inference
Frozen model logic turns the validated feature state into probability outputs.
Probability outputs
Separate probability questions are published for 1X2, Over/Under 2.5 and Both Teams to Score.
Certification checks
Fixture identity, pre-kickoff timing and publication lineage are checked before an output becomes a certified public record.
Settlement & measurement
After the match, the stored probability can be evaluated with accuracy, calibration, Brier score and other performance metrics.
What ScoreProb analyzes
The 362-input vector is organized into football and statistical families. The examples below describe the type of evidence considered; they do not reveal model weights or interaction formulas.
Strength & ratings
Relative team strength, ELO-related information and strength differences.
ELO Power score ELO difference
Recent form
Recent-match points, results and performance patterns over short rolling windows.
Recent 5 Recent 10 PPG
Venue context
Home and away performance is separated where the venue context carries information.
Venue PPG Win rate Goals for per game
Attack & scoring
Scoring rates, goals per game, multi-goal frequencies and recent attacking behavior.
Scoring rate Goals for per game Scored 2+ rate
Defense & conceding
Conceding rates, goals allowed, clean sheets and defensive consistency.
Conceding rate Goals against per game Clean-sheet rate
Goal profiles
Over/Under tendencies, scoring thresholds and distributions around total goals.
Over 1.5 rate Over 2.5 rate Over 3.5 rate
BTTS & scoring failure
Both-teams-to-score behavior, clean-sheet rates and failed-to-score patterns.
BTTS rate Clean-sheet rate Failed-to-score rate
Volatility & stability
Variation in results, goal difference and performance stability across samples.
Standard deviation Stability score Result entropy
Halftime & match state
First-half and second-half behavior, lead retention and response when trailing.
First half Second half Retention
Goal timing
When goals tend to occur and how scoring behavior changes across match periods.
First half Second half Goal
Schedule & recovery
Rest intervals, recent workload and schedule context available before kickoff.
Rest days Previous 7 days Previous 14 days
Competition context
Competition-level historical context and sample availability where it is statistically usable.
Competition Sample Available
Market context
The certified feature universe can include validated market-context and odds-derived signals among many other inputs. ScoreProb probabilities themselves are model outputs, not bookmaker odds.
Probability Bookmakers History
Examples, not an exhaustive formula
Three probability questions, not one universal score
A shared certified feature state can feed separate prediction problems. Each output answers a different football question.
362 362-feature state
Home, Draw and Away form one 1X2 probability distribution.
Over 2.5 and Under 2.5 describe the probability of the match total crossing the 2.5-goal line.
Yes and No describe whether both teams are expected to score at least once.
How to read a probability
Illustrative example
An illustrative 62% Home probability does not mean the home team is certain to win. It means the model assigns Home a 62% probability given the information available before kickoff.
A well-calibrated probability system should see events assigned around 60% occur around 60% of the time across a sufficiently large sample.
See real calibration → Home 62% Draw 23% Away 15%
62% 23% 15%
Margin and uncertainty help describe the shape of 1X2
They are interpretation aids, not extra outcomes. Historical performance by margin and uncertainty is published on the Performance page.
Strong separation Lower uncertainty
55% 27% 18%
Tight contest Higher uncertainty
38% 34% 28%
What happens before publication?
A ScoreProb probability becomes public only after the required pre-kickoff checks pass.
✓ Available before kickoff
✓ Required data validated
✓ Feature state assembled
✓ Model inference completed
✓ Fixture identity verified
✓ Kickoff still in the future
✓ Certified output published
What can change — and what cannot
Before the next certified publication
- New completed matches can change recent form and historical statistics.
- Venue, schedule and other pre-kickoff context can evolve as new valid information becomes available.
- A newer certified pre-kickoff probability can become the current public probability.
After certification
- The stored historical probability record is not rewritten after the result.
- Its fixture and prediction identity remain bound to the certified publication.
- Historical performance is measured against the record that existed before kickoff.
Probability is not certainty
80% can still lose If an event is truly an 80% event, failure roughly 20 times in 100 is normal over a large enough sample. High probability means more likely, not guaranteed.
80% 20%
Data integrity safeguards
Future match information cannot be used in a certified pre-kickoff probability.
Missing information is not silently converted to zero unless a defined feature legitimately represents zero.
The prediction must remain bound to the same fixture, teams and kickoff context it was generated for.
Invalid numeric values are rejected instead of quietly entering a certified model state.
Settled results are attached to previously published records rather than recreating probabilities after the match.
Certified outputs retain the identifiers and lineage needed to audit which model record was actually published.
Transparent about methodology. Proprietary about implementation.
ScoreProb publishes enough information to understand and audit the product without publishing enough detail to reproduce the commercial model.
Public
- Probability definitions and model outputs
- Broad feature families and approved fixture-level feature values
- Historical accuracy, calibration and performance
- Certification, settlement and integrity rules
Proprietary
- Exact feature weights and coefficients
- Transformation equations and normalization parameters
- Feature interaction and ensemble logic
- Training code, hyperparameters and model binaries
- Private validation thresholds and internal source mappings
Methodology explains the process. Performance tests the claims.
Compare predicted probability with observed frequency on the Performance page.
Inspect historical 1X2 accuracy by margin and uncertainty bands.
Performance is settled against stored certified records, not reconstructed after results.
Frequently asked questions
Is ScoreProb an AI model?
ScoreProb uses trained statistical prediction models over engineered pre-match features. We describe it as a probability model rather than relying on 'AI' as a marketing label.
Are ScoreProb probabilities bookmaker odds?
No. ScoreProb publishes model probabilities. The feature universe may include validated market-context or odds-derived signals where available, but the published ScoreProb percentages are model outputs rather than bookmaker prices.
What does 70% mean?
It means the model assigns the event a 70% probability given the certified information available before kickoff. It does not mean the event is guaranteed.
How many statistics does ScoreProb analyze?
The current certified feature vector contains 362 engineered inputs. They span multiple football and statistical families; the exact weighting and interactions remain proprietary.
Does ScoreProb change probabilities after kickoff?
The historical certified pre-kickoff record used for performance measurement is not rewritten after kickoff or after the result. A later public live state is a different concept from rewriting the certified historical probability.
Why can a 75% probability still lose?
Because 75% still leaves 25% probability for other outcomes. Football contains irreducible uncertainty, and good probability models express that uncertainty rather than pretending it does not exist.
How is ScoreProb performance measured?
The Performance page uses settled certified pre-kickoff records and reports pick accuracy, calibration, Brier score, log loss, probability-strength bands and other diagnostics.
What are Brier score and calibration?
Brier score measures squared probability error. Calibration asks whether events assigned a given probability occur at roughly that frequency over a large sample.
How often is the model updated?
The public site shows the latest certified pre-kickoff probability available for a fixture. A newer certified state can replace the current public view before kickoff, while earlier certified records remain preserved for audit.
Does historical accuracy guarantee future performance?
No. Historical performance describes how previously published probabilities behaved. It is evidence about past model behavior, not a guarantee of future outcomes.