Ranked Probability Score
0.2027
vs 0.2369 equal-probability baseline
14.4% lower error
Lower is better. RPS rewards well-calibrated probabilities across home, draw and away outcomes.
EPL model validation
Model Performance
A transparent record of what the FiveStat model gets right, how it compares with simple baselines, and where it still struggles.
Headline scorecard
Probability quality first
Backtest file updated 03 Nov 2026 at 11:03 GMT
These are historical out-of-sample backtest results, not a live betting record. Each fixture was predicted using only information available before that game.
1X2 accuracy
51.4%
vs 43.1% always-home baseline
+8.3 percentage points
How often the single most likely match outcome was correct.
Brier score
0.2124
Probability error score
Lower is better. Penalises confident probabilities when the predicted event does not occur.
Test sample
1,138
2023/24–2025/26 (combined)
3 fixtures were skipped where the model lacked sufficient prior information.
Pre-declared evaluation windows
Robustness and recent performance
Lower RPS is better
These windows are calculated together from the same frozen predictions. The full available sample remains the primary result; shorter windows are shown to reveal recent performance and possible model drift, not to replace the headline with whichever period scores best.
Full available window
2023/24–2025/26
0.2027
RPS vs 0.2369 baseline
14.4% lower error
- 95% bootstrap interval
- 0.1945–0.2112
- 1X2 accuracy
- 51.4%
- Predictions
- 1,138
Latest two seasons
2024/25–2025/26
0.2083
RPS vs 0.2345 baseline
11.2% lower error
- 95% bootstrap interval
- 0.1993–0.2176
- 1X2 accuracy
- 48.9%
- Predictions
- 759
Latest season
2025/26
0.2140
RPS vs 0.2322 baseline
7.8% lower error
- 95% bootstrap interval
- 0.2010–0.2281
- 1X2 accuracy
- 46.1%
- Predictions
- 380
Window comparison
RPS across fixed evaluation windows
Lower is better
Every point comes from the same chronological walk-forward run. Overlapping windows are shown for transparency, while the full sample remains the primary benchmark.
Supporting metrics
Other ways we test the model
How it was tested
Walk-forward, not in-sample
-
1
Train on the past
Ratings use only matches completed before the forecast date being predicted.
-
2
Freeze the forecast
Fixtures on the same date share a cutoff, so no result from that matchday can enter another prediction.
-
3
Score every prediction
Forecasts are compared with results using proper probability scores and simple benchmarks.
Known limitation
Draws are hard to rank first
22.0%average model draw probability
24.5%actual draw rate
The aggregate draw probability is reasonably close to the observed rate, but a draw was never the model's single highest-probability outcome in this test. That is why FiveStat publishes probability scores alongside headline classification accuracy.
Scope boundary
What this page does not claim
It is not yet an immutable live prediction record.
It does not claim realised betting profit or ROI.
It does not yet compare every forecast with closing market prices.
The next validation layer is to archive pre-match production forecasts and score them after results arrive. Until then, live and backtested performance remain clearly separated.
Want the mechanics?
Read the full methodology
See how attack and defence ratings, recent form, scoreline probabilities and season simulations are produced.