Free 1X2 experiment · Runs locally
Recalibrate football prediction probabilities
Fit a mapping on one settled archive. Freeze it. Then learn whether it improves Brier score, log loss and reliability on a different, untouched football test set.
Canonical publication record
Abstract
A deterministic browser experiment that fits temperature, one-vs-rest Platt and one-vs-rest isotonic mappings on a football 1X2 calibration partition, freezes the parameters, then scores them against raw probabilities on a separate test partition.
- Author and publisher
- Football Proof AI
- Technical report
football-prediction-recalibration/1.0.0- Published
- Last modified
- Release status
- Current release
- Review status
- Editorial technical note; not externally peer reviewed
- Version history
football-prediction-recalibration/1.0.0: Initial public release.
- Immutable artifacts
- football-recalibration-calibration-example-v1.csv
sha256:38be26bb9966d2be671ce508cf3116ad646da21f36ae8cb694f1845411abec24 - football-recalibration-test-example-v1.csv
sha256:07f96aa13294be1686abac39ffbf1e51143dfe5c5c87e5d7dc5e2bf78527ca7a
- football-recalibration-calibration-example-v1.csv
Two locked partitions · Formula v1
Fit on calibration. Open the test result once.
Both CSVs stay on this device. The calibration file fits three mappings; the test file never changes their parameters. Choose one candidate before the test scoreboard is calculated.
This browser lock records intent in the exported passport. It is not an independent preregistration service or historical timestamp. Choosing again after inspection turns the test set into tuning data.
Calibration CSV
Parameters come from these accepted rows and nowhere else.
Untouched test CSV
Outcomes here score fixed mappings; they never refit them.
Check both partitions, declare one candidate mapping, then run the local experiment. The example is synthetic and proves no model skill.
Direct answer
Recalibration is a trained model layer, not a cosmetic curve
To recalibrate football predictions honestly, estimate the mapping on past predictions and outcomes, select the method without reading the final test scores, and evaluate the frozen mapping on later matches. If test outcomes influence the mapping or method choice, the reported test improvement is no longer untouched evidence.
Input contract
Both partitions use complete, pre-match 1X2 records
A row is fitted or scored only when all three probabilities, one settled outcome and valid publication and kick-off times pass the same strict audit. Invalid rows stay visible instead of disappearing from the denominator.
- match_id
- Unique fixture identity; overlap between files invalidates an external-validation claim
- published_at
- Zoned ISO 8601 time when the probability split was available
- kickoff_at
- Zoned ISO 8601 kick-off strictly after publication
- p_home / p_draw / p_away
- One complete distribution summing to 1 or 100
- outcome
- Settled H, D or A result used only in its declared partition
Download the synthetic calibration CSV and untouched test CSV. They demonstrate the contract and must never be cited as model performance.
Three mapping families
Flexibility rises from temperature to isotonic
- Temperature scaling
- Fits one positive T to soften or sharpen all three log probabilities. It preserves class ordering and has the fewest degrees of freedom.
- OvR Platt
- Fits a logistic slope and intercept for home, draw and away separately, then renormalises the three outputs to sum to one.
- OvR isotonic
- Fits a monotone PAV mapping per outcome, linearly interpolates between blocks, applies a disclosed floor and renormalises.
More flexible is not automatically better. Isotonic can closely follow a small calibration set and still degrade the untouched test set. Temperature scaling is restrictive but often easier to estimate. Publish both partitions instead of selecting the most flattering training fit.
The test-set firewall
Looking at every method and choosing the winner spends the test set
This lab records one candidate method before calculation and labels it on the scoreboard. The other methods remain a sensitivity analysis. If their test results change the deployment choice, create a new, later test period before making an external claim.
- 01
Keep IDs disjoint
One fixture in both files leaks its outcome into fitting and testing. The lab reports every overlapping match ID.
- 02
Move forward in time
A later test period better matches deployment and exposes drift. A non-forward holdout is labelled honestly.
- 03
Choose before scoring
Method selection, hyperparameter search and stopping rules belong before the final test is read.
- 04
Retest after decisions
Once a test result changes the workflow, it becomes development evidence—not a reusable final exam.
Read the complete scoreboard
Lower probability error matters even when the top pick does not change
Brier score measures squared error across home, draw and away. Log loss strongly penalises assigning tiny probability to the observed result. ECE-10 compares confidence with frequency in ten bins but changes with the bin design. Top-pick hit rate ignores most probability movement and therefore remains supplementary.
First use the calibration and reliability lab to diagnose raw probabilities. Use this page only when you have enough independent data to fit a mapping and still preserve a test. For time-ordered evaluation, follow the walk-forward validation protocol.
Interpretation boundary
A lower test loss supports this mapping on this supplied period
It does not prove future accuracy, profitable odds, historical publication or archive completeness. Team, league and time dependence can make small differences unstable, while model drift can make yesterday's mapping obsolete.
- No universal winner.The right method depends on sample size, misspecification and genuinely unseen performance.
- No same-sample claim.A better fit on the calibration partition is expected and is not external evidence.
- No timestamp proof.SHA-256 identifies the bytes; it does not establish when either archive existed.
- No betting advice.Calibration quality alone does not establish expected value, staking edge or profit.
Complete the evidence chain
Diagnose. Recalibrate. Monitor.
Treat probability correction as one versioned model component. Preserve its fit window, frozen parameters, test fingerprint and later drift evidence together.
Primary and technical sources
Methods behind the lab
- Zadrozny and Elkan (2002): transforming classifier scores into multiclass probabilities
- Niculescu-Mizil and Caruana (2005): empirical comparison of probability calibration methods
- Guo et al. (2017): temperature scaling and modern calibration
- scikit-learn: calibration curves and cross-validation boundary
Cite this tool as: Football Proof AI. “Football Prediction Probability Recalibration Lab.” Version football-prediction-recalibration/1.0.0, 14 July 2026, https://footballproofai.com/tools/football-prediction-recalibration-lab.