Manchester City vs Chelsea Prediction, Fair Odds & Result
Model answer: Home at 58%. Home is 58%, draw 24% and away 18%. These are clearly labelled sample figures until the production pipeline is connected.
One URL, complete history
Final — result reviewed
The original prediction is preserved beside the final score of 1–2.
Automatic post-match audit
Demo MissThe result is recorded without rewriting the forecast.
The locked top call was home win at 58%. The verified result was away win at 1–2.
- Final score
- 1–2
- Actual outcome
- Away win
- Probability assigned to result
- 18%
- Single-match Brier
- 0.355
- Latest settlement
- 2026-05-03T18:30:00.000Z
A single match cannot prove or disprove calibration. The score joins every other settled row on the public accuracy ledger.
Calibrated 1X2 view
Home, draw and away probabilities
- H · Manchester City
- 58%
- D · Draw
- 24%
- A · Chelsea
- 18%
| H — Manchester City | D — Draw | A — Chelsea |
|---|---|---|
| 58%probability | 24%probability | 18%probability |
| 1.72fair odds | 4.17fair odds | 5.56fair odds |
Calibration mirror
What does this confidence mean in reality?
58% is the largest probability, not certainty. The other two outcomes still carry 42% combined, and the top two outcomes are separated by 34%.
- Settled forecasts
- 2
- Mean forecast
- 57.5%
- Observed frequency
- 100.0%
- Calibration gap
- 42.5%
These labelled sample rows were already settled before this demo forecast was published. They are not evidence of production performance.
Open the current full accuracy ledgerHistorical probability shapes
Probability Neighbours
Before publication, 8 settled forecasts with the closest full 1X2 probability shapes were selected from all published leagues and model versions. Actual outcomes were 5 home wins, 2 draws and 1 away wins. Their mean prediction was home 50.0%, draw 26.5% and away 23.5%; mean class-averaged Brier score was 0.1429. Mean Total Variation distance was 11.5% and the maximum was 28.0%. Fewer than 8 records were available inside the 10% Total Variation window, so the deterministic search expanded to the nearest eligible records; the furthest selected record was 28.0% away. This is a small sample below 30 records, so it is descriptive evidence rather than proof of model skill or a promise about this match.
As of
A distance of 5% means five probability points must be redistributed to turn one 1X2 split into the other. Match results, scores, Hit/Miss labels and Brier scores never participate in neighbour selection. This historical view does not change the current prediction.
- Settled neighbours
- 8
- Selection scope
- all published leagues and model versions
- Mean distance
- 11.5%
- Maximum distance
- 28.0%
- Average Brier
- 0.1429
What actually happened
Observed H/D/A outcomes
- Home win
- 562.5%
- Draw
- 225.0%
- Away win
- 112.5%
Swipe horizontally to inspect probabilities, result, distance and verdict.
| Historical record | Published H / D / A | Actual result | TV distance | Verdict |
|---|---|---|---|---|
| Inter vs BolognaSerie A | H 59%D 25%A 16% | H · Home winFinal score 2–0 | 2.0% | Demo Hit |
| Enyimba vs Kano PillarsNigeria Premier Football League | H 56%D 27%A 17% | H · Home winFinal score 2–1 | 3.0% | Demo Hit |
| Arsenal vs West Ham UnitedPremier League | H 61%D 23%A 16% | H · Home winFinal score 2–0 | 3.0% | Demo Hit |
| Barcelona vs Real BetisLa Liga | H 67%D 20%A 13% | H · Home winFinal score 3–1 | 9.0% | Demo Hit |
| Roma vs FiorentinaSerie A | H 48%D 29%A 23% | D · DrawFinal score 1–1 | 10.0% | Demo Miss |
| Bandari vs Kakamega HomeboyzFKF Premier League | H 45%D 30%A 25% | H · Home winFinal score 1–0 | 13.0% | Demo Hit |
| Athletic Club vs VillarrealLa Liga | H 34%D 30%A 36% | D · DrawFinal score 0–0 | 24.0% | Demo Miss |
| Everton vs Newcastle UnitedPremier League | H 30%D 28%A 42% | A · Away winFinal score 1–2 | 28.0% | Demo Hit |
Publication proof
Forecast identity
- Published
- Model version
- FPAI-LGBM 0.8.8
- Competition
- PREMIER-LEAGUE
- Confidence
- 58%
- Record SHA-256
- Not issued for demo data
Demo notice: the timestamp, model version and figures on this preview are sample data, so no production record fingerprint is issued. Official records are generated automatically at least 24 hours before kick-off.
Kito calibration challenge
Make your call before the whistle
Set your own home, draw and away split. It stays only in this browser, saves the match to your private desk, and can be scored beside the locked model after the result.
Loading your device-local call…
Your selected probabilities never leave this device · anonymous aggregate use counters respect Global Privacy Control and Do Not Track · excluded from the public ledger, accuracy, APIs and structured data · not betting advice
Interactive counterfactual
What would have to change?
Move three transparent assumptions and watch the probability shape respond. This is an educational sensitivity test, not a second model and never an edit to the demo record.
Nearest illustrative flip
5-point gridFind the smallest move from the neutral official baseline that puts another outcome at least 0.5 percentage points ahead.
This boundary is a sensitivity illustration—not a new prediction. Missing-information stress stays at 0 and the official record never changes.
- Top-two gap
- 34 pts
- Uncertainty
- High
- Official record
- Unchanged
Illustrative Scenario v1. Method: apply visible log-probability shifts, then normalise with softmax; the stress control raises temperature toward a flatter split. Boundary search uses only team balance and draw pressure on a 5-point grid, holds stress at zero, and requires the target to lead by at least 0.5 percentage points.
Why the model leans this way
Locked probability record
- 01
The locked home probability was 58%.
- 02
The locked draw probability was 24%.
- 03
The locked away probability was 18%.
Grounded record explainer
Ask Kito about this probability
Kito can read only this labelled demo record. It cannot browse for hidden context, change the probabilities, suggest a stake, or promise a result.
Instant evidence mode is deterministic and makes no external model request. Optional AI response planning remains off until a server-side OpenAI key and budget switch are configured. When enabled, the model can only choose a full or compact view of server-authored evidence.
This is a labelled demo record. The top call is home win at 58%. Across a sufficiently large set of genuinely comparable 58% forecasts, that outcome should occur about 58 times in 100—not every time. The other two outcomes still hold 42% combined probability.
- Locked probability splitHome 58% · Draw 24% · Away 18%
- Model versionFPAI-LGBM 0.8.8
- Publication time2026-05-01T08:00:00Z
- Probability neighbours8 prior settled records · all published leagues and model versions
Locked feature context
Model-generated explanation
The published top call was Home. The final result was Away at 1–2, so the forecast was recorded as a miss. This row remains visible as part of the complete demo record.
Generated only from the locked pre-match feature snapshot; maximum 120 words and no change to the official probabilities.