1. Model Architecture
The system employs a multi-layered data processing pipeline:
Pipeline: D → F(D) → M(F) → T(M) → P2. Model Specifications
| Model Identifier | Golsinyali AI |
| Version | v2.1 |
| Training Data Window | 24 months |
| Training Dataset | ~120,000 matches |
| Cumulative Analysis Count | 50,000+ |
| Model Update Frequency | Weekly |
3. Data Sources and Feature Engineering
AMarket-Based Features
Opening/Closing odds differential (ΔO)Line movement velocity (dL/dt)Market Consensus Index (MCI)Sharp money indicator
BPerformance Features
Exponential Moving Average (EMA-5, EMA-10)Expected Goals (xG) differentialHome/Away performance coefficientGoal Difference Momentum (GDM)
CHistorical Features
Head-to-head win probability (H2H-WP)League position delta (ΔPos)Seasonal trend vectorMatch Importance Weight (MIW)
4. Mathematical Framework
Confidence score calculation uses a normalized distance function:
Where O represents the observed odds value, μ is the optimal range mean, and σ is the range standard deviation.
The prediction is activated when the confidence score exceeds the threshold and the odds value is within the acceptable range.
5. Prediction Types and Acceptance Criteria
Match Result (FT)
P(MS) = f(O_home, O_away) where O ∈ [1.40, 2.00]Favorite identification based on European odds system
Acceptance: 1.40 ≤ O_fav ≤ 2.00Over 2.5 Goals
P(O2.5) = g(L_ou) where L ∈ [2.5, 3.5]Goal expectation depends on Over/Under line
Acceptance: 2.5 ≤ Line ≤ 3.5BTTS (Both Teams to Score)
P(BTTS) = h(O_draw, L_ou) where O_x ≤ 4.0 ∧ L ∈ [2.5, 3.75]Dual-condition acceptance criterion
Acceptance: O_draw ≤ 4.00 ∧ 2.50 ≤ L ≤ 3.756. Backtesting Results
Model performance was evaluated on a 24-month out-of-sample test set:
Overall Performance
Confusion Matrix (Normalized)
Calculated from last 10,000 predictions
7. Confidence Score Calibration
The model produces well-calibrated probabilities. The table below compares predicted confidence ranges with observed success rates:
| Predicted | Observed | Sample Count | Δ |
|---|---|---|---|
| 70-75% | 72.3% | 2,847 | +2.3% |
| 75-80% | 77.8% | 3,521 | +2.8% |
| 80-85% | 82.1% | 2,198 | +2.1% |
| 85-90% | 86.4% | 1,102 | +1.4% |
| 90-95% | 91.2% | 332 | +1.2% |
8. Feature Importance Analysis
Feature importance scores calculated using permutation importance method:
9. Methodological Limitations
- [1]The model does not incorporate real-time events (injuries, red cards, weather conditions)
- [2]Predictions are statistical probabilities; no deterministic outcome guarantee
- [3]Market manipulation and insider trading are outside model scope
- [4]Performance may decrease in smaller leagues due to data insufficiency
- [5]The model detects correlations; it does not make causal inferences
10. References and Methodology
- [1]Kelly, J. L. (1956). A New Interpretation of Information Rate. Bell System Technical Journal.
- [2]Štrumbelj, E., & Vračar, P. (2012). Simulating a basketball match with a homogeneous Markov model.
- [3]Dixon, M. J., & Coles, S. G. (1997). Modelling Association Football Scores. Applied Statistics.
- [4]Constantinou, A. C., et al. (2012). Profiting from an inefficient association football gambling market.
11. Corporate Information
Tarsier Vision LTD
UK Company #14646033
This system is developed and operated by Tarsier Vision LTD (UK Company #14646033).