Testing theprototype.
A historical reconstruction of 115,933 stored fixtures. The prototype and a simple competition-season baseline are scored on the same matches. This is evidence about this snapshot, with the limits below.
Restored from the 11 October backup. Current updates are not connected.
- Model
- poisson-proto-1
- Evaluated fixtures
- 115,933
- Competition-seasons
- 639
- Cutoff (UTC)
- 2026-10-18 02:30:00 UTC
- Evidence
- Retrospective reconstruction
Free on MatchDaySync
Stats Studio
Choose a competition, a window and a metric, then compare up to six clubs.
The model beside its baseline.
Lower values are better for every score below. The result Brier score sums three squared probability errors per fixture, on a 0–2 scale. Binary goal-threshold scores use a 0–1 scale. These are scoring rules, not financial returns or a reliability guarantee.
| Score | Prototype | Season baseline |
|---|---|---|
| Result Brier | 0.630 | 0.651 |
| Result log loss (natural log) | 1.061 | 1.076 |
| Three or more goals · Brier | 0.256 | 0.246 |
| Both clubs score · Brier | 0.259 | 0.248 |
| Goal error · mean absolute error per side | 0.945 | 0.950 |
In this snapshot, the prototype's result Brier score is lower than the baseline. Its three-or-more-goals score is higher than its baseline, and its both-clubs-score result is higher than its baseline. A lower score is better.
Result probabilities below 1e-15 are floored only when calculating log loss: 30 prototype and 0 baseline cases. The probabilities and Brier scores are unchanged.
From the GKM network
Ghana travel guides and planning tools
Guides, tools and roadmaps for visiting, living and investing in Ghana. Every price is published.
What entered the test.
Chronology comes first
Only valid scored finished matches with a recorded kickoff strictly before each target kickoff, within the same competition-season. Equal kickoff batches are predicted before any result in the batch is added. Unknown kickoffs are excluded from both fitting and evaluation.
The model keeps its existing minimum of 5 earlier matches per club and 40 per competition-season. It restarts each season. Its formulas and 0–7 score grid are unchanged.
The comparison is deliberately simple
Same-prefix competition-season home/draw/away, over2.5 and BTTS frequencies, and home/away goal averages. Exactly the same evaluated fixtures as the prototype.
The baseline does not use club strength. It establishes whether extra model complexity helps on these same fixtures; the table reports the result either way.
- Stored full-time rows
- 170,549
- Unknown kickoff · excluded
- 15,499
- Input or scoring-baseline shortfall
- 39,113
- Numerical failure · excluded
- 0
- Invalid score or identity · excluded
- 0
- At/after cutoff · excluded
- 4
- Known, scored and before cutoff
- 155,046
- Actually evaluated
- 115,933
Advertise with MatchDaySync
Reach readers who check the numbers
MatchDaySync pages are built around fixtures, tables and club statistics. Ask about a placement.
Probability and observed frequency.
These ten probability bands cover all 115,933 evaluated fixtures for each outcome. Compare the mean model probability with the proportion that occurred. Empty bands stay empty; small bands provide less evidence. This is descriptive calibration on the stored snapshot, without a claim about future calibration.
Home win
| Band | N | Mean probability | Occurred |
|---|---|---|---|
| 0–10%* | 4,112 | 6.2% | 19.1% |
| 10–20%* | 10,825 | 15.5% | 27.1% |
| 20–30%* | 16,954 | 25.3% | 32.6% |
| 30–40%* | 20,994 | 35.1% | 39.5% |
| 40–50%* | 21,218 | 44.9% | 44.3% |
| 50–60%* | 17,338 | 54.8% | 51.3% |
| 60–70%* | 12,261 | 64.6% | 57.9% |
| 70–80%* | 7,497 | 74.5% | 62.6% |
| 80–90%* | 3,632 | 84.3% | 71.9% |
| 90–100% inclusive | 1,102 | 93.5% | 80.6% |
* Upper edge excluded; the final band includes 100%.
Draw
| Band | N | Mean probability | Occurred |
|---|---|---|---|
| 0–10%* | 3,064 | 6.9% | 14.1% |
| 10–20%* | 21,450 | 16.5% | 21.6% |
| 20–30%* | 73,452 | 25.0% | 26.7% |
| 30–40%* | 16,268 | 33.0% | 30.0% |
| 40–50%* | 1,398 | 43.4% | 31.5% |
| 50–60%* | 223 | 53.6% | 26.9% |
| 60–70%* | 59 | 63.3% | 28.8% |
| 70–80%* | 15 | 72.9% | 40.0% |
| 80–90%* | 4 | 83.9% | 50.0% |
| 90–100% inclusive | 0 | – | – |
* Upper edge excluded; the final band includes 100%.
Away win
| Band | N | Mean probability | Occurred |
|---|---|---|---|
| 0–10%* | 10,835 | 6.1% | 12.7% |
| 10–20%* | 22,210 | 15.3% | 19.8% |
| 20–30%* | 25,753 | 25.0% | 26.1% |
| 30–40%* | 21,839 | 34.8% | 31.0% |
| 40–50%* | 15,423 | 44.7% | 36.9% |
| 50–60%* | 9,753 | 54.6% | 43.5% |
| 60–70%* | 5,620 | 64.5% | 49.7% |
| 70–80%* | 2,844 | 74.3% | 57.3% |
| 80–90%* | 1,317 | 84.2% | 64.0% |
| 90–100% inclusive | 339 | 93.8% | 70.5% |
* Upper edge excluded; the final band includes 100%.
Three or more goals
| Band | N | Mean probability | Occurred |
|---|---|---|---|
| 0–10%* | 984 | 6.5% | 39.5% |
| 10–20%* | 4,170 | 16.0% | 39.5% |
| 20–30%* | 11,284 | 25.7% | 42.1% |
| 30–40%* | 20,999 | 35.4% | 45.3% |
| 40–50%* | 27,233 | 45.1% | 49.1% |
| 50–60%* | 24,661 | 54.8% | 53.3% |
| 60–70%* | 16,172 | 64.5% | 57.4% |
| 70–80%* | 7,402 | 74.3% | 61.9% |
| 80–90%* | 2,545 | 84.0% | 67.9% |
| 90–100% inclusive | 483 | 93.0% | 79.1% |
* Upper edge excluded; the final band includes 100%.
Both clubs score
| Band | N | Mean probability | Occurred |
|---|---|---|---|
| 0–10%* | 652 | 5.6% | 43.1% |
| 10–20%* | 2,784 | 16.0% | 45.7% |
| 20–30%* | 9,094 | 25.9% | 45.9% |
| 30–40%* | 22,249 | 35.6% | 48.9% |
| 40–50%* | 34,335 | 45.1% | 52.0% |
| 50–60%* | 29,566 | 54.6% | 54.8% |
| 60–70%* | 13,366 | 64.1% | 59.2% |
| 70–80%* | 3,336 | 73.6% | 63.2% |
| 80–90%* | 529 | 83.2% | 66.4% |
| 90–100% inclusive | 22 | 91.4% | 72.7% |
* Upper edge excluded; the final band includes 100%.
Look inside the aggregate.
The score table above and calibration bands keep the complete evaluation population. This directory shows each competition-season with eligible fixtures, so a good aggregate cannot hide every poor subgroup. Small samples remain visible.
Free on MatchDaySync
Find matches by your own rules
Set conditions on club statistics and list the matches that meet them.
From the GKM network
The Dental Protocol Learn Hub
An editorial oral-wellness library of in-depth, evidence-based articles.
Advertise with MatchDaySync
Advertise with MatchDaySync
Place your brand beside stored match data, tables and tools.
What this evidence cannot establish.
- Retrospective reconstruction from the current stored results, not a record of forecasts published at the time.
- Original ingestion/publication times and historical corrections are unavailable; contemporaneous data availability is not proven.
- Unknown kickoffs and invalid observations are excluded, so the evaluation population differs from the live season aggregate input population.
- No lineups, injuries, odds, betting returns, prospective validation or reliability guarantee is inferred.
- All source competition types meeting the existing model minimums are included; the aggregate is not a professional-league-only result.
Reproduce this exact calculation
Evaluation chronological-known-kickoff-1. Generated 2026-10-18T02:45:17.169Z. The read-only repository script scripts/analysis/evaluate-forecast.cjs accepts an explicit UTC cutoff. Reproduction requires the same underlying stored rows, not just the same model version.
Source SHA-256: 3c7aaf1eb7d13a8982c0447df0a7f06710abe0543c467c692cc5c513a1cd6876
- src/lib/forecast.ts
- 23346541458742a1d7bb6b5495506d73ddd1672a91dc1ec3a4b4368806e00674
- src/lib/forecast-evaluation.ts
- 824620c4fb5b40714ee3060cafba0deb2ed05a367dea02cd929092bc9495e001
- scripts/analysis/evaluate-forecast.cjs
- 2728c948af825c183907c4820984c70bbab3f1860b96d52c257585e8e0a4c696
More statistics
How to read these figures
The full definitions are in the glossary and the methodology.
- What is a sample?
- The number of matches behind a figure. A percentage or an average is counted only over matches that have that measurement stored, and a figure from a handful of matches is marked as a small sample.
- What is a rate?
- The share of measured matches in which something happened, such as both teams scoring. It describes the matches counted and is not a forecast.
- Why can two tables differ?
- Season views count one competition-season. Last 5, 10 and 20 windows follow each club across stored seasons in its competition, and a home or away filter applies after the window is chosen.
- Is a missing figure counted as zero?
- No. A missing measurement is left out of that figure's sample and is never treated as zero.
Statistics only, not advice. Not affiliated with any league or club.