The last saved data is shown below. Scores and statuses may change at the next refresh.
FLY · ANALYSIS
Prediction analysis
See the data used, how predictions are tested and what their recorded results show.
How does the model work?
For football, the model estimates attack, defence, home advantage and weighted past results. It derives home, draw and away probabilities and possible scorelines.
With limited data, estimates move toward league averages to reduce overconfidence. Percentages are withheld when data is insufficient.
Methodology and data tiers →Probability and data quality
A 70% probability does not guarantee a win in one match. It describes a frequency over many comparable matches.
A/B/C tiers describe data coverage. They are not win probabilities or guaranteed success rates.
Injury, confirmed lineup and weather data are not yet model inputs. The analysis does not invent knowledge of them.
ACTUALLY PUBLISHED PREDICTIONS
Current version results
Model dc-v2 · Last 90 days
95% confidence interval: 43.1%–51.3%.
Check probability calibration
| Predicted probability | Matches | Observed hit rate |
|---|---|---|
| 0–40% | 106 | 39.6% |
| 40–50% | 254 | 47.2% |
| 50–60% | 126 | 45.2% |
| 60–70% | 57 | 61.4% |
| 70–80% | 14 | 50.0% |
| 80–100% | 5 | 80.0% |
Evaluation includes only completed matches with predictions published before kickoff and then locked. Results from different model versions are kept separate.
TEMPORAL VALIDATION
Previous-season comparison
688 matches · 11 chronological folds · Spain, Germany.
Each fold trains only on matches completed earlier. This is a retrospective test on selected leagues, not the website's live performance across all sports.
| Variant | Matches | 1X2 | Log loss | Brier |
|---|---|---|---|---|
| Current model | 688 | 51.7% | 0.9961 | 0.5932 |
| Stronger shrinkage | 688 | 51.3% | 1.0014 | 0.5969 |
| Faster recency weighting | 688 | 51.7% | 1.0004 | 0.5962 |
| Probability calibration | 688 | 51.7% | 1.0025 | 0.5969 |
| Past-result frequencies | 688 | 41.7% | 1.0825 | 0.6560 |
| Equal probabilities | 688 | 41.7% | 1.0986 | 0.6667 |
Log loss penalizes the probability assigned to the observed result; Brier is the sum of squared errors across the three outcomes. Lower is better for both.
Live model weights are not automatically replaced before a candidate is confirmed on later, prospective matches.
Data updates
In-play probabilities are shown only when that match's score is less than two minutes old. Delayed scores are not presented as knowledge of the current match state.
Free football, NBA and tennis feeds can update slowly. Always check the score's observation time.
Data and update intervals →