TOUCHLINE TIPS

Season 2026/27

Are the calls any good?

Two kinds of evidence, kept apart. The live record counts only predictions locked before kickoff. The backtest replays past seasons with the same rules and is useful, but it is not the same as calling games in advance.

Live record

Started with Matchweek 9, 2026/27

10Predictions locked before kickoff
6/6Super Tips settled
50%Hit rate, against 65% expected
0.212Brier score (lower is better); typical-match baseline 0.190
3/4Passes that would have won

By confidence

20%30%40%50%60%70%80%90%Strong: landed 100% of 3 (plausible 44–100%); model expected 72%Strongn = 3100%Lean: landed 0% of 3 (plausible 0–56%); model expected 59%Leann = 30%
Live Super Tips by confidence. Dot: what actually happened. Band: plausible range for that sample size. Dark tick: what the model expected.

Calibration

20%30%40%50%60%70%80%90%50–60%: landed 0% of 2 (plausible 0–66%); model expected 57%50–60%n = 20%60–70%: landed 50% of 2 (plausible 9–91%); model expected 66%60–70%n = 250%70–80%: landed 100% of 2 (plausible 34–100%); model expected 73%70–80%n = 2100%
Live Super Tips grouped by stated probability. Dot: what actually happened. Band: plausible range for that sample size. Dark tick: what the model expected.

Is the record luck or a signal?

If every locked probability were exactly right, this is how many of the 6 settled Super Tips should have won. 3 of 6 is within the normal range for these probabilities: no sign of a problem, and no sign of brilliance either.

0 of 6: 0.1%01 of 6: 1.8%12 of 6: 9.0%9%23 of 6: 23.3%23%Actual34 of 6: 33.3%33%45 of 6: 24.9%25%56 of 6: 7.6%8%680% of the time
3 of 6 Super Tips won. 3 of 6 is within the normal range for these probabilities: no sign of a problem, and no sign of brilliance either.

Small sample. 6 settled Super Tips is far too few to judge the model: with this many picks, a hit rate can swing by 20 points or more on luck alone. The plausible-range bands above show how wide the uncertainty still is.

By market

MarketSettledWonHit rateModel expectedPlausible range
Team goals100%58%0%–79%
Double chance4375%70%30%–95%
Match result100%55%0%–79%

Week by week

RoundFixturesWithout a callSuper TipsPassesWonLostPushVoidPending
Matchweek 91006433000Scorecard

Model versions

VersionRoundsPredictionsSettled tipsWonFirst locked
tl-poisson-v1.1Matchweek 910632026-10-30 23:10 UTC

Every prediction keeps the model version it was locked with. A new version only applies to rounds locked after it was promoted.

Backtest: the same rules on past seasons

840 matches, 2024/25 to Matchweek 8 of 2026/27

Did confident calls land more often?

Strong calls landed 80% of the time against 74% expected. Moderate and lean calls landed within 3 points of what the model said.

40%50%60%70%80%90%Strong: landed 80% of 245 (plausible 74–84%); model expected 74%Strongn = 24580%Moderate: landed 63% of 243 (plausible 57–69%); model expected 64%Moderaten = 24363%Lean: landed 58% of 168 (plausible 51–66%); model expected 61%Leann = 16858%Pass: landed 63% of 184 (plausible 55–69%); model expected 65%Passn = 18463%
Simulated Super Tips by confidence, walk-forward. Dot: what actually happened. Band: plausible range for that sample size. Dark tick: what the model expected.

Does kickoff time matter once team strength is known?

Mostly no. 16:00–18:59 kickoffs drifted from the model; with 4 slots tested, that is a hypothesis to track, not a rule.

20%30%40%50%60%70%Before 14:00: landed 48% of 65 (plausible 36–60%); model expected 42%Before 14:00n = 6548%14:00–15:59: landed 43% of 414 (plausible 38–48%); model expected 44%14:00–15:59n = 41443%16:00–18:59: landed 32% of 161 (plausible 25–39%); model expected 40%16:00–18:59n = 16132%19:00 or later: landed 44% of 200 (plausible 37–51%); model expected 43%19:00 or latern = 20044%
Home-win rate by kickoff slot, out of sample. Dot: what actually happened. Band: plausible range for that sample size. Dark tick: what the model expected.

Where the model has skill

Skill compares the model’s Brier score with simply using how often each outcome happens. Positive means better than the base rate. Only markets with clear skill on the earlier validation seasons can become Super Tips.

MarketMatchesModel saidHappenedSkillEligible
Home to win84043%41%11.5%Yes
Draw84023%26%0.9%No
Away to win84034%32%10.9%Yes
Home win or draw84066%68%10.9%Yes
Away win or draw84057%59%11.5%Yes
Over 2.5 total goals84056%55%0.3%No
Under 2.5 total goals84044%45%0.3%No
Under 3.5 total goals84066%69%-0.7%No
Over 1.5 total goals84079%80%0.3%No
Both teams to score84056%56%0.0%No
At least one team fails to score84044%44%0.0%No
Home over 1.5 goals84046%47%5.9%Yes
Away over 1.5 goals84039%38%4.1%Yes
Home not to score84022%24%6.2%Yes
Away not to score84027%26%4.2%Yes

Match-result log loss: model 1.009 against 1.084 for base rates (lower is better). Model tl-poisson-v1.1, half-life 240 days.

What we changed, and why it matters. v1 made every market with positive validation skill eligible. On the test seasons, total-goals and both-teams-to-score markets showed ~0% skill and the selected Under 3.5 picks were overconfident (predicted 72%, landed 68%): picking the largest gap to the base rate in a low-skill market mostly selects model error. v1.1 (used live from 2026/27 Matchweek 9) requires at least 3% validation skill. This change was made AFTER seeing the v1 test results, so the 2024/25–2026/27 figures are no longer a clean holdout for v1.1. The live weekly record is the clean test. No odds were recorded for the backtest, so it says nothing about betting profit: a high hit rate on short-priced outcomes can still lose money.