FootyTipsModel Track Record
Does the model beat the closing line?
Not on the evidence so far — and we publish that rather than hide it. Across 74 settled matches, the model's own pick drifted +0.45pp in our favour by kickoff. But simply backing the market favourite every time would have drifted +0.43pp. The part actually attributable to the model is +0.02pp — indistinguishable from zero at this sample size. Treat this model as a transparent, reproducible baseline, not a market-beater.
model pick's line movement
always backing the market favourite
excess attributable to the model
settled matches measured
Closing-line value compares the best price we showed 24 hours before kickoff with the de-vigged market consensus at kickoff. Sample is 74 of 104 settled matches that had a usable closing snapshot (median 2.68h before kickoff); the rest are excluded and counted rather than dropped quietly. These World Cup 2026 figures are reconstructed from stored odds snapshots; Premier League signals are recorded live from the opening round.
Across 22 World Cups (964 matches), our model called the correct result 56.6% of the time — versus ~33% for random guessing and ~45% for always backing the home/seeded side. Brier score 0.57 (lower is better; random ≈ 0.667).
World Cup backtest matches
World Cups tested
win/draw/loss hit rate
Brier score
Citable source note: this is a walk-forward backtest using only information available before each historical match. Historical match data comes from martj42/international_results; World Cup 2026 settlement is tracked separately after final scores are available.
18+ · Please gamble responsibly. Predictions are probabilistic, not guarantees. Availability of betting content may be restricted in your region.
World Cup 2026 Live Settlement
Settled matches
Correct W/D/L picks
Live hit rate
Live Brier
Flat 1u ROI
Current streak
World Cup 2026 results are settled after full time and include misses as well as hits. See the public prediction results ledger.
ROI is calculated as a flat 1-unit stake on the model's top W/D/L pick using the best pre-kickoff price available in the odds snapshot. 2 settled matches had no pre-kickoff market price and are excluded from ROI.
Hit Rate by Tournament
| World Cup | Correct result |
|---|---|
| 1930 | 78% |
| 1934 | 71% |
| 1938 | 67% |
| 1950 | 50% |
| 1954 | 73% |
| 1958 | 43% |
| 1962 | 63% |
| 1966 | 66% |
| 1970 | 63% |
| 1974 | 53% |
| 1978 | 53% |
| 1982 | 50% |
| 1986 | 60% |
| 1990 | 52% |
| 1994 | 54% |
| 1998 | 55% |
| 2002 | 50% |
| 2006 | 64% |
| 2010 | 52% |
| 2014 | 59% |
| 2018 | 56% |
| 2022 | 55% |
| All 22 editions | 56.6% |
How We Test — No Cheating
We replay every international match since 1872 (49,520 games) in date order, updating each team's Elo rating after every result. For each historical World Cup match, the prediction uses only the Elo available before kick-off — then we compare it to what actually happened. This walk-forward design means no future information leaks in: it reflects how the model would genuinely have performed at the time.
The exact same engine — Elo driven by real results, not hand-set numbers — powers our World Cup 2026 predictions. Method details on How It Works, with probability reliability shown in the calibration report.
Match results are sourced from the open-source martj42/international_results dataset (every international match since 1872); World Cup 2026 fixtures from TheSportsDB. The results dataset is released under a CC0 public-domain licence, so this backtest is fully reproducible. The full backtest — code and results — is open source: footytips-worldcup-backtest.
Premier League 2026/27 — live record
The same model now covers the Premier League 2026/27. Settling starts with the opening round on Aug 21 — every call will be published before kickoff and settled here. Premier League predictions →
Pre-season backtest (out-of-sample: params calibrated on 2019–24, tested on 2024-25 & 2025-26): 390/760 correct (51.3%), Brier 0.6083 — vs the de-vigged market baseline at 51.4% / 0.5951. The market is the strongest known baseline; matching it while staying fully reproducible is the point. Reproduce it yourself (open source) →
An Honest Note
56.6% correct at win/draw/loss is a strong, competitive level for international football — but it is not a crystal ball. Upsets are part of the game, and part of any honest model. That's why every page shows probabilities, not guarantees.
FAQ
How accurate are AI World Cup predictions?
FootyTips.io's AI model (Elo + Poisson) was backtested on all 22 World Cups from 1930 to 2022 — 964 matches — and picked the correct win/draw/loss result 56.6% of the time, versus ~33% for random guessing. At World Cup 2026 it has so far called 67 of 104 finished matches correctly (64%), settled publicly after every match. Professional football models typically land between 50–58%, so a mid-50s hit rate is a strong, honest level — these are probabilities, not guarantees.
How accurate are FootyTips predictions?
Backtested across 22 World Cups (964 matches), the model called the correct result — win, draw or loss — 56.6% of the time, well above the ~33% random and ~45% always-home baselines.
How is the backtest done without cheating?
We replay every international match since 1872 in date order, building each team's Elo rating as we go. For each historical World Cup game we predict using only the Elo available before kick-off, then compare to the actual result. No future information is ever used.
Does the hit rate mean predictions are guaranteed?
No. Around 56% correct at win/draw/loss is strong for international football, but upsets are inherent to the game. We publish probabilities, not certainties.
Cite This Data
Writing about this? Copy this citation (attribution appreciated):
Updated 2026-08-18 · based on 49,520 historical international matches from the martj42/international_results dataset.