1X2
20/34
58.8%
20 wins · 14 losses · n=34
Fewer than 100 settled matches — not enough evidence yet.
Transparency
Track record
Every published prediction is recorded under the model version that made it. Live results are kept separate — never merged with development runs or backtests.
Settled · Public permanently · Never edited
In one sentence
As of Nov 4, 2026, 11:08 PM UTC, WinSight has publicly settled 688 matches and 1,375 match-result (1X2) forecasts.
A match is counted once. Forecasts count each settled match-result (1X2) call from a non-development model version. Void results are excluded, and the other markets are not added to these totals.
The current model has a 1X2 hit rate of 58.8% (n=34).
Hit rate is wins divided by wins plus losses, for that model version only. Versions stay separate and are never pooled.
Not enough data yet: with fewer than 100 settled matches, the numbers — high or low — don't yet say much about performance.
Download the settled record (CSV)
Each row is one settled match-result (1X2) forecast from a non-development model. Void results and the other markets are left out. Use of this file follows the site terms. It is not a separate open-data licence.
- Settled matches
- 688
- Settled forecasts
- 1,375
A match is counted once. Forecasts count each settled match-result (1X2) call from a non-development model version. Void results are excluded, and the other markets are not added to these totals.
Live results now lead this page. The registered backtest remains below as a fixed historical reference.
Issued prediction record
Overall hit rate
Current model
Issued prediction record
Overall hit rate
Earlier model
1X2
67/102
65.7%
67 wins · 35 losses · n=102
Issued prediction record
Overall hit rate
Earlier model
1X2
92/154
59.7%
92 wins · 62 losses · n=154
Issued prediction record
Overall hit rate
Earlier model
1X2
46/84
54.8%
46 wins · 38 losses · n=84
Issued prediction record
Overall hit rate
Earlier model
1X2
96/170
56.5%
96 wins · 74 losses · n=170
Issued prediction record
Overall hit rate
Earlier model
1X2
88/149
59.1%
88 wins · 61 losses · n=149 · 1 void
Issued prediction record
Overall hit rate
Earlier model
1X2
79/153
51.6%
79 wins · 74 losses · n=153 · 1 void
Issued prediction record
Overall hit rate
Earlier model
1X2
34/59
57.6%
34 wins · 25 losses · n=59 · 1 void
Issued prediction record
Overall hit rate
Earlier model
1X2
15/28
53.6%
15 wins · 13 losses · n=28
Issued prediction record
Overall hit rate
Earlier model
1X2
74/151
49.0%
74 wins · 77 losses · n=151
Issued prediction record
Overall hit rate
Earlier model
1X2
63/128
49.2%
63 wins · 65 losses · n=128
Issued prediction record
Overall hit rate
Earlier model
1X2
75/147
51.0%
75 wins · 72 losses · n=147
Issued prediction record
Overall hit rate
Earlier model
1X2
5/16
31.3%
5 wins · 11 losses · n=16
The rest of the settled record
Sign in to see the other settled markets and the weekly breakdown. The public page keeps the 1X2 hit rate and its sample size.
Backtest
Backtest · qualifying threshold
Current live model
Current model
Unchanged between versions
1X2
Registered 1X2 record
Earlier model
The current live model keeps the same match-result performance as the registered qualifying threshold, and only recalibrates BTTS probabilities. The backtest below is that qualifying threshold.
- 1X2 accuracy
- 50.54%
- n=17,320
- Multiclass Brier
- 0.6020
- Log loss
- 1.0071
- Calibration slope
- 0.940
Benchmarks
The same matches, measured against four benchmarks
2019/20–2025/26 · the same 16,322 matches with Bet365 closing data
Random pick of three33.3%
Always the home team43.0%
WinSight (walk-forward backtest)50.2%
Bet365 closing favourite52.4%
On the same 16,322 matches, WinSight trails Bet365 by 2.16 percentage points. Public forecasts rarely beat closing odds, and we make no such claim. These are backtest figures, not live results.
1X2 calibration curve
Historical backtest · not live performanceObserved result frequencyPerfect calibrationMarker area = sample size
↑ Observed more often than predicted · ↓ Observed less often than predicted
- 0%–10% range · 1,369 outcome probabilities · predicted 6.9% · observed 9.0% · observed minus predicted +2.1 pp
- 10%–20% range · 6,187 outcome probabilities · predicted 16.0% · observed 16.1% · observed minus predicted +0.1 pp
- 20%–30% range · 20,610 outcome probabilities · predicted 25.5% · observed 26.1% · observed minus predicted +0.6 pp
- 30%–40% range · 9,603 outcome probabilities · predicted 34.4% · observed 34.3% · observed minus predicted −0.1 pp
- 40%–50% range · 6,414 outcome probabilities · predicted 44.7% · observed 43.9% · observed minus predicted −0.9 pp
- 50%–60% range · 4,065 outcome probabilities · predicted 54.5% · observed 53.2% · observed minus predicted −1.2 pp
- 60%–70% range · 2,177 outcome probabilities · predicted 64.5% · observed 64.4% · observed minus predicted −0.1 pp
- 70%–80% range · 1,056 outcome probabilities · predicted 74.4% · observed 72.9% · observed minus predicted −1.5 pp
- 80%–90% range · 406 outcome probabilities · predicted 83.8% · observed 79.8% · observed minus predicted −4.0 pp
- 90%–100% range · 73 outcome probabilities · predicted 92.9% · observed 90.4% · observed minus predicted −2.5 pp
Predicted probability (%) →
Where the model places its probabilities
Outcome-probability samples
Across 17,320 matches, the model produced 51,960 outcome probabilities. 58.1% fall within 20%–40%; only 0.92% are 80% or higher. This shows how often the model assigns each level of confidence, not whether a prediction was correct.
- 0%–10% probability · 1,369 probability samples
- 10%–20% probability · 6,187 probability samples
- 20%–30% probability · 20,610 probability samples
- 30%–40% probability · 9,603 probability samples
- 40%–50% probability · 6,414 probability samples
- 50%–60% probability · 4,065 probability samples
- 60%–70% probability · 2,177 probability samples
- 70%–80% probability · 1,056 probability samples
- 80%–90% probability · 406 probability samples
- 90%–100% probability · 73 probability samples
Registered qualification record
Methodology
How the record is calculated
Scoring
Each settled prediction is counted market by market. A WIN adds one hit and a LOSS adds one miss. Correct score is scored against the model's Top 1 pick, not any of its Top 3 displayed picks.
VOID handling
POSTPONED and CANCELLED fixtures are recorded as VOID. VOID rows are disclosed in sample details but excluded from both the numerator and denominator, so they cannot inflate or depress the hit rate.
Weekly cut
Weeks start Monday and follow the fixture's UTC kickoff, matching the weekly breakdown on this page. A percentage is wins ÷ (wins + losses); summaries are recomputed from counts, never by averaging weekly percentages.
Model history
Every result remains attached to the model version that produced the prediction. Versions are reported as separate cohorts. The live record begins with production predictions from the current model; development datasets and walk-forward backtests are labelled separately.
Market benchmark
The reference scale above is historical and never pooled into the live record. WinSight and Bet365 use the exact same 16,322 fixtures and closing-price coverage; live bookmaker comparison will remain absent until it can be matched fixture by fixture.
Terms
The four markets, explained
Compact tables retain common football abbreviations. Each market and its scoring rule are defined below.
- Match result (1X2)
- The result after regular time: home win, draw, or away win. “1”, “X”, and “2” are the compact labels for those three outcomes.
- Total goals — over/under 2.5 (O/U 2.5)
- Whether both teams combine for at least 3 goals, or finish with 0–2 goals.
- Both teams to score (BTTS)
- Whether each team scores at least once during the match.
- Correct score
- The exact final score. WinSight displays three likely scores for context, while the record scores only the Top 1 pick.