Backtest — not live results

NBA

The live NBA model (v2) was tested walk-forward: each game is priced from earlier games only. The 2008–09 to 2017–18 seasons were used to choose the version. A later holdout, and a 2025–26 season test, were scored after that. Headline figures below are the completed 2025–26 season — the same window the site already cites for calibration — plus an earlier holdout window for green hit rate.

2025–26, 1,322 games: the model picked the winner in 895 of 1,322 (67.7%). Games it gave 70–75% won 100 of 136 (73.5%). Green picks (65%+) won 536 of 697 (76.9%).

A separate, already published Super test on 2025–26 — the one quoted on How the lights work — looked at 162 Super games: they won 63%. That test knew the real line-ups. Live Super still depends on the injury report being right, which is why a value pick with uncertain key players is held at its normal light.

In the older SBR-matched window (2018–19 through 16 January 2023, 5,613 games), green picks (65%+) won 2,139 of 2,938 (72.8%). Green is a confidence light, not a value flag.

Sources: NBA v2 walk-forward predictions joined to results; Super hit rate as already published on How the lights work.

NFL

The live NFL model is v2. It was scored walk-forward on 2016 through early 2026. 2016–22 (1,896 games) was the validation window used to choose the version, so it is not the headline. The clean test is the 2023–26 holdout (919 games, including 65 games of 2026 so far).

Holdout 2023–26: the model picked the winner in 592 of 919 (64.4%). Games it gave 70–75% won 91 of 123 (74.0%). Green picks (65%+) won 316 of 441 (71.7%).

For comparison, the 2021–2025 seasons already mentioned on the site (a mix of validation and holdout): 915 of 1,420 (64.4%) of all picks, and 132 of 177 (74.6%) in the 70–75% band. That all-picks figure is 64.4%, which the site has rounded to 65%. The file-traced numbers are the ones in this table.

Source: NFL v2 walk-forward predictions. 2016–22 was used to choose the version, so the holdout is the headline.

NHL

Walk-forward on official results, one game date at a time. Settings were chosen on 2021–22 and 2022–23; 2023–24 to 2025–26 are the clean test (4,192 games). Lights below use the current public cut-off (green at 65%+), applied to the same walk-forward probabilities.

Clean test 2023–26: the model picked the winner in 2,414 of 4,192 (57.6%). Green picks won 609 of 867 (70.2%). 2025–26 was the weakest season in the test (751 of 1,394, 53.9%), and that season stays in the table.

Source: NHL walk-forward re-run on official results. Green rates use the current 65% cut-off, not the older 60% column on the sport Stats tab.

Premier League

Walk-forward, match by match from earlier days’ results only (2020–21 used to warm up the ratings). 2021–22 and 2022–23 were used to choose the settings, so they flatter the model a little. 2023–24 to 2025–26 (1,140 matches) are the clean test. Green rates below use the current public cut-off of 55%+, recomputed from the walk-forward chances. (The walk file’s own light column still uses the older 60% cut-off; we do not quote that as the live green rule.)

Clean test 2023–26: the model was right in 602 of 1,140 (52.8%). Green picks (55%+) won 274 of 406 (67.5%). 2025–26 was a poor season for all picks (185 of 380, 48.7%) and is kept in the test. With three possible results, about half right is a normal season; draws are the hardest to call.

Source: Premier League walk-forward re-run. Match totals match the published backtest file; green rates use the current 55% cut-off.

Championship

Past-season re-run with the same settings as the Premier League model, never tuned on Championship data. 2019–20 was used only to warm up the ratings. 2021–22 to 2025–26: 1,294 of 2,760 (46.9%) of all picks. The green column stored in the published backtest file is 347 of 579 (59.9%). We could not find a walk-forward CSV to confirm which cut-off produced that green column, so we report it as stored rather than re-colour it. Live Championship green on the site is 55%+.

With three results and a draw about one match in four, under half right is normal for this league. 2025–26 was the weakest of the five seasons and stays in the total.

Source: the published Championship backtest file. The green column is reported as stored; live Championship green is 55%+.

Tennis

Walk-forward re-run of ATP and WTA main-draw matches, January 2021 to November 2026. Ratings were built match by match from earlier results only. The final step (a logistic regression) was fitted on 2009–2020 and never saw these matches.

30,189 matches: the model picked the winner in 19,669 of 30,189 (65.2%). Green picks (70%+) won 8,017 of 10,225 (78.4%). Orange (60–70%) won 5,765 of 9,103 (63.3%); red (under 60%) won 5,887 of 10,861 (54.2%). In tennis the favourite wins often, so a red light really does mean a close match. The model’s average stated chance was 66.0%, against 65.2% actual — close, and slightly cautious.

Live tennis results show vs-close only on events that had an Australian bookmaker price.

Source: the published tennis walk-forward backtest file, the same figures shown on the tennis Stats tab.

What this page is not

It is not the live record, not a promise of future results, and not betting advice. This page shows hit rates only — not staking, units or ROI. The official member-facing record, once you are looking at tips, remains the timestamped Results page and its CSV.

Live results → · How it works → · How the lights work → · Live calibration →