MatchOracle

Deep Ensemble EPL Prediction Engine

13 base learners · Dixon-Coles statistical model · 376+ features · 8 data sources · NLP sentiment analysis

---

An advanced EPL match prediction system built on real data from 8 sources spanning 20 seasons (~7,600 matches). Features a 5-layer deep stacking ensemble with walk-forward backtesting that consistently outperforms market odds.

Metric	Result	vs Market
3-Way Accuracy	60.2%	+4.6% (55.6%)
RPS Score	0.1163	+11.2% skill
ELITE Tier (>70%)	82.2% acc	90 matches
Best Season	66.8%	2023-24

Ensemble Architecture · Features · Data Sources · Results · Getting Started

Highlights

5-Layer Ensemble Architecture Layer 0: Dixon-Coles statistical model Layer 1: 13 base learners (HGB, XGB, LightGBM, CatBoost, RF, MLP) Layer 2: 4 meta-learners with isotonic calibration Layer 2.5: Binary classifier boosting Layer 3: Best ensemble selection (stacking vs weighted avg vs market-fused)	376+ Engineered Features Elo + Glicko-2 + Pi-Ratings Rolling form (6 windows), H2H, momentum Market intelligence (Shin probabilities, odds movement) Poisson goal decomposition, xG-based metrics Manager tenure, GK quality, rest days, scoring patterns
NLP Sentiment Analysis RoBERTa transformer (70%) + keyword fallback (30%) Dual news sources: NewsAPI + Google News RSS Aspect-based: injury risk, morale, tactical disruption 30 live sentiment features injected at prediction time	Smart Model Caching 5 automated retraining checks (data hash, age, integrity) <2 second predictions after initial training Interactive CLI with arrow-key fixture selector Auto-generated HTML dashboard with Plotly/Chart.js

---

Ensemble Architecture

Layer 0: Dixon-Coles (goals + xG variants)
    │
Layer 1: 13 Base Learners
    │   HGB, HGB-Agg, HGB-Deep, XGBoost, LightGBM, CatBoost,
    │   Random Forest, Extra Trees, DeepMLP, MLP-Wide,
    │   Logistic Regression, Bagging-HGB, Vote-HGB3
    │
Layer 2: 4 Meta-Learners (Meta-LR, Meta-MLP, Meta-HGB, Meta-XGB)
    │
Layer 2.5: Binary Classifier Boosting (4 dedicated HGB models)
    │
Layer 3: Best Ensemble Selection
        Stacking vs Weighted Avg vs Binary-3Way vs Market-Fused

---

Data Sources

#	Source	Coverage	Key Data
1	football-data.co.uk	20 seasons	Results, shots, corners, bookmaker odds
2	Understat	11 seasons (2014+)	Match-level xG, xGA
3	Club Elo	20 seasons	Historical Elo ratings
4	Open-Meteo	20 seasons	Weather at stadium GPS
5	Football-Data.org	Live season	Standings, fixtures, H2H
6	API-Football	Live season	Injuries, player ratings
7	NewsAPI	Live	Team news for NLP sentiment
8	Google News RSS	Live	Fallback news source

---

Walk-Forward Results

5-season walk-forward backtesting (no future data leakage):

Season	Accuracy	RPS	vs Market
2020-21	55.0%	0.1314	—
2021-22	61.8%	0.1080	+17.6%
2022-23	60.0%	0.1189	+9.2%
2023-24	66.8%	0.1040	+20.6%
2024-25	57.1%	0.1194	+8.9%

Confidence Tier Breakdown

Tier	Confidence	Accuracy	Matches
ELITE	>70%	82.2%	90
VERY HIGH	60-70%	66.7%	96
HIGH	50-60%	62.7%	83

376+ Feature Groups

Group	Features	Description
Elo Ratings	6	Home/away Elo + differential
Pi-Ratings	8	Home/away attack/defense ratings
Rolling Form	40+	6 rolling windows (3-20 matches)
Head-to-Head	20+	Historical H2H results and trends
Market Intelligence	15+	Implied probabilities, odds movement, Shin
Momentum	25+	Streaks, acceleration, velocity
Glicko-2	10	Rating + uncertainty + volatility
Poisson	15+	Attack/defense decomposition
Contextual	10	Derby flags, distance, title contender
Injuries	12	Per-team injury impact
Manager	6	Tenure, new manager bounce
GK Quality	6	Clean sheets, consistency
Sequence Patterns	12	Encoded result sequences
Scoring Patterns	8	Early goals, comebacks
Rest Days	6	Fatigue/freshness flags

---

Getting Started

git clone https://github.com/abailey81/MatchOracle.git
cd MatchOracle

# Setup
chmod +x setup.sh && ./setup.sh
# Or manually:
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# Configure API keys
cp .env.example .env  # Edit with your keys

# Run predictions
python predict.py --fdo-key YOUR_KEY --apif-key YOUR_KEY --news-key YOUR_KEY

---

Project Structure

MatchOracle/
├── predict.py                  # Interactive CLI entry point
├── dashboard.py                # HTML dashboard generator
├── data/
│   ├── generator.py            # Real data pipeline (8 sources, 20 seasons)
│   └── api_client.py           # Rate limiter, circuit breaker, caching
├── features/
│   ├── engine.py               # 376+ features across 24 groups
│   └── sentiment.py            # RoBERTa NLP sentiment analysis
├── models/
│   ├── run_pipeline.py         # 5-layer ensemble pipeline
│   ├── dixon_coles.py          # Statistical model (1997)
│   └── model_cache.py          # Smart caching with retraining detection
├── requirements.txt
├── setup.sh
└── .env.example

---

MIT License

Built with scikit-learn, XGBoost, LightGBM, CatBoost, and RoBERTa