Home › Open data › Calibration tables
How well the TACTIX model’s probabilities match what happened, as aggregated tables you can download and check. Each row is one cell: a market, a confidence tier, a tenth of the probability scale and a calendar month, with the number of settled predictions behind it, the mean probability the model gave, the share that landed, the Brier score and the calibration error. The predictions are the ones the scoreboard grades, settled and graded the same way.
The file holds derived results only: no match, date, competition, team or single result appears in it, and a cell behind fewer than 10 matches is left out, so no row can be traced to one fixture. Fixture data: API-Sports.
Latest snapshot : 11404 rows from 7873023 settled predictions across 246 markets, kick-off months 2025-08 to 2026-09. Download the latest snapshot (CSV).
| Column | Type | Definition |
|---|---|---|
month | YYYY-MM | Calendar month of kick-off (UTC) of the matches in the cell. A month only: no day, no time. |
market | text | The TACTIX market key the predictions were made for (for example match_winner or btts). |
confidence_tier | text | The TACTIX confidence tier of the probability bucket: gold (0.90 and above), silver (0.80 to 0.90), bronze (0.70 to 0.80), untiered (below 0.70). |
probability_bucket | text | The tenth of the probability scale the predictions fall in, lower bound included (0.7-0.8 means 0.70 up to but not including 0.80; 0.9-1.0 includes 1.00). |
predictions | integer | Number of settled, graded TACTIX predictions in the cell (voids and ungraded predictions are not counted). |
matches | integer | Number of distinct matches behind those predictions. Cells with fewer than the published minimum are left out of the file. |
mean_predicted | decimal 0-1 | Mean probability the TACTIX model gave the predictions in the cell. |
observed_frequency | decimal 0-1 | Share of the predictions in the cell that landed. |
brier | decimal 0-1 | Mean Brier score of the cell: the mean of (probability minus outcome) squared, outcome 1 when the prediction landed and 0 when it missed. Lower is better. |
ece | decimal 0-1 | Calibration error of the cell: the absolute gap between mean_predicted and observed_frequency. The prediction-weighted mean of this column over any set of rows is that set’s expected calibration error. |
| Snapshot | Rows | SHA-256 |
|---|---|---|
| 2026-10-26.csv | 11404 | 681366e847df0f4701a365efd9823d6d85ea26129f9e9e35dd855608edbfe0db |
The manifest: index.json.
Each weekly snapshot is published once and never revised; corrections appear in a later snapshot with a changelog line. A snapshot covers the kick-off months that ended before the week it was written in; a prediction graded later appears in the next snapshot.
Published by TACTIX under the Creative Commons Attribution 4.0 International (CC-BY-4.0) licence. Attribution: TACTIX, with a link to this page. The licence covers these derived tables; fixture data: API-Sports, not part of the file.
Plain text:
TACTIX (2026). TACTIX Open Calibration Tables, Snapshot 2026-10-26 [Data set]. https://tactix.football/data/calibration
BibTeX:
@misc{tactix_calibration_20261026,
author = {{TACTIX}},
title = {TACTIX Open Calibration Tables, Snapshot 2026-10-26},
year = {2026},
howpublished = {\url{https://tactix.football/data/calibration}},
note = {Weekly snapshot 2026-10-26, published once and never revised}
}If you use the data, please link back to https://tactix.football/data/calibration. How the model works: methodology. The record as a page: scoreboard. Site terms.
Predictions are probabilities, not certainties. 18+, please gamble responsibly (BeGambleAware.org).