Can Machine Learning Beat the Odds? A Deep Dive into Football Match Prediction

Overview

This work builds a data‑driven football match prediction system for the five major European leagues (2010‑2015). It extracts two groups of features: (1) team‑profile vectors (17 continuous dimensions per side covering strength, recent form, head‑to‑head history, home advantage, offensive/defensive metrics) and (2) bookmaker odds (initial win/draw/loss odds from 17 bookmakers, yielding 51 dimensions). The resulting dataset contains 1,339 training matches (after removing extreme outliers) and a 365‑match test set for the English Premier League.

Prediction Models

Linear and Classical Non‑Linear Models

Logistic Regression (LR) achieved 38.18% accuracy on the Premier League test set. Support Vector Machine (SVM) improved accuracy to 51.23%.

Ensemble Non‑Linear Models

Random Forest consistently reached >53% accuracy across most leagues; the French Ligue 1 showed lower performance due to higher competition entropy.

Deep Neural Network (DNN) with Unsupervised Pre‑training

A greedy layer‑wise unsupervised pre‑training followed by supervised fine‑tuning was applied to learn high‑level embeddings from the raw 68‑dimensional input (17 team features + 51 odds features). The DNN ensemble further increased prediction accuracy, achieving up to 54.55% on the Premier League.

Score Prediction

Two approaches were explored:

Modeling each team’s goal count with a Poisson distribution and estimating the λ parameters from the learned features.

Formulating score prediction as a multi‑class classification problem (e.g., 5 × 5 possible scorelines) using a Softmax output layer.

Betting Strategy Analysis

For a single‑match fixed‑odds bet (stake = 2 CNY), profitability requires the inverse of prediction accuracy to be lower than the average winning odds: profit > 0 ⇔ 1/accuracy < average_odds Empirical analysis shows that only predictions with model‑estimated probabilities < 0.4 or ≥ 0.9 satisfy this condition, covering roughly 20‑55% of matches depending on the league. The best‑performing SVM model on the Premier League test set yields a profit‑positive interval at p < 0.4 and p ≥ 0.9.

Limitations

Profitability intervals have been validated only on test sets of ~300 matches; broader generalization requires larger datasets.

Betting coverage ratios are modest (e.g., 20% for the Premier League, 7% for Ligue 1).

Odds volatility leads to unstable profit rates across different data splits.

Extension to Financial Time‑Series

The same pipeline—signal mining, feature extraction, and predictive modeling—can be transferred to stock price prediction. Relevant signals include price series, technical indicators (MACD, KDJ), sentiment data, and macro‑economic events. Deep models such as DNN, LSTM, and Transformers are well‑suited for capturing temporal dependencies in financial data.

Future Work

Expand the dataset to include additional leagues (e.g., Chinese Super League) and cup competitions (Champions League, domestic cups).

Enrich features with dynamic odds changes, lineup information, player fatigue, match importance, and news sentiment.

Develop adaptive betting strategies that exploit probability intervals while improving stability.

Explore additional prediction targets such as exact goal counts, upset probabilities, and in‑play odds.

Conclusion

The proposed system, based on team‑profile and odds features, achieves up to 54.55% accuracy for win/draw/loss prediction in the Premier League. While the current models can guide single‑match betting and score prediction, further data collection, feature engineering, and strategy refinement are required to build a stable, profitable betting framework.