২২ জুন, ২০২৬ · 12 blog.minRead · methodology
Logistic Regression for Football Predictions — The Foundation Every Model Needs
June 22, 2026 · 10 min read
Behind every football prediction model — from simple spreadsheets to Premier League forecasting engines — sits a deceptively powerful statistical technique. Logistic regression has been the workhorse of sports prediction for decades, and understanding how it works will change the way you think about match probabilities.
What Is Logistic Regression, and Why Does Football Need It?
Football matches have three possible outcomes: home win, draw, or away win. Unlike predicting a continuous number (like tomorrow's temperature), match results are categorical. You can't regress on "2-1" the way you'd regress on a price or a score. This is where logistic regression earns its name — it transforms raw numerical inputs into probabilities that always fall between 0 and 1.
The core idea is straightforward. You collect data about each match — team strength, home advantage, recent form, head-to-head records — and feed these features into a mathematical function. The function outputs the probability of each outcome. A model might tell you: "Manchester City have a 62% chance of winning, a 22% chance of drawing, and a 16% chance of losing." Those probabilities sum to 100%, and they're grounded in historical patterns, not gut feelings.
Logistic regression has been used in academic football prediction research since the 1990s. A 2018 study published in the Journal of Quantitative Analysis in Sports found that logistic regression models achieved 52-55% accuracy on three-way match outcomes (win/draw/loss) across Europe's top five leagues — a meaningful edge over the 33% baseline of random guessing, and often competitive with far more complex machine learning approaches.
The Sigmoid Function: Turning Numbers Into Probabilities
The mathematical heart of logistic regression is the sigmoid function (also called the logistic function). It takes any number — positive, negative, or zero — and squashes it into a value between 0 and 1:
p(x) = 1 / (1 + e−z)
where z is a weighted sum of your input features:
z = β₀ + β₁x₁ + β₂x₂ + β₃x₃ + ...
Each β coefficient represents how much influence a particular feature has on the outcome. A large positive β₁ for "home advantage" means playing at home significantly increases the probability of winning. A negative β₂ for "days since last match" (fatigue) means tired teams are less likely to win.
The sigmoid curve is S-shaped. When z = 0 (the features are neutral), the output is exactly 0.5 — maximum uncertainty. As z increases (strong positive signals), the probability climbs toward 1. As z decreases (strong negative signals), it drops toward 0. This natural mapping from "strength score" to "probability" is what makes logistic regression so intuitive for sports prediction.
A Worked Example
Suppose your model uses three features: Elo rating difference (x₁), home advantage (x₂ = 1 if home, 0 if away), and recent form measured as points per game over the last 5 matches (x₃). After training on 10,000 historical matches, you find:
- β₀ = −2.1 (intercept — baseline log-odds)
- β₁ = 0.003 (Elo difference — each 100-point gap adds 0.3 to z)
- β₂ = 0.45 (home advantage — adds 0.45 to z)
- β₃ = 0.35 (recent form — each additional point per game adds 0.35 to z)
Now you want to predict a match: a team with a 120-point Elo advantage, playing at home, averaging 2.4 points per game recently. Plugging in:
z = −2.1 + (0.003 × 120) + (0.45 × 1) + (0.35 × 2.4) = −2.1 + 0.36 + 0.45 + 0.84 = −0.45
p = 1 / (1 + e0.45) = 1 / (1 + 1.568) = 0.39
The model gives this team a 39% chance of winning — below 50%, meaning despite their advantages, the baseline intercept pulls the probability down. This might happen if the model was trained on data where home wins are less common than you'd expect, or if the Elo difference isn't large enough to overcome other factors.
From Binary to Three-Way: The Softmax Extension
Football doesn't have binary outcomes. You need to predict win, draw, or loss — three mutually exclusive categories. The standard approach is multinomial logistic regression, which uses the softmax function instead of the basic sigmoid.
Instead of computing a single score z, the model computes a separate score for each outcome:
- score(Win) = βW · features
- score(Draw) = βD · features
- score(Loss) = βL · features
The softmax function then converts these scores into probabilities:
P(outcome k) = escore(k) / (escore(W) + escore(D) + escore(L))
The denominator ensures all three probabilities sum to 1. Higher scores produce higher probabilities, and the exponential amplifies differences — a team that's slightly stronger on paper gets a disproportionately higher win probability.
In practice, one outcome (usually "Loss") is set as the reference category with coefficients fixed at zero. The model then learns how much each feature pushes the probability toward Win or Draw relative to Loss. This avoids a mathematical problem called "non-identifiability" — without a reference point, multiple different coefficient sets could produce the same probabilities.
Choosing the Right Features
A logistic regression model is only as good as its inputs. Football prediction research has identified several feature categories that consistently improve model performance:
Team Strength Metrics
- Elo ratings: The single most predictive feature in most studies. A 2019 analysis of 150,000 matches found Elo alone achieved 50.2% accuracy on three-way outcomes — higher than any other single feature.
- League position / points per game: A simpler proxy for team strength, though noisier early in the season.
- xG (Expected Goals):strong> Teams that consistently outperform their xG may be genuinely clinical or simply lucky. Using xG-based strength estimates instead of actual goals produces more stable predictions.
Match Context
- Home advantage: Historically worth 0.4-0.5 goals in European leagues, though this has declined in recent years. At the 2022 World Cup, home advantage was effectively zero (no host nation effect for Qatar's group stage). The 2026 World Cup's three-host format makes this even more complex.
- Rest days: Teams playing 2 matches per week show measurably lower win rates. The coefficient for rest days is typically small but statistically significant.
- Travel distance: Relevant for international tournaments and continental competitions. Teams crossing multiple time zones historically underperform by 3-5%.
Form and Momentum
- Points per game (last 5-10 matches): Captures current trajectory without overreacting to a single result.
- Goals scored/conceded per match: More granular than points — a team winning 1-0 every week is different from one winning 4-3.
- Head-to-head record: Some matchups produce persistent patterns. Italy historically struggles against certain tactical setups; South American teams tend to overperform against European opposition in knockout rounds.
How the Model Learns: Maximum Likelihood Estimation
Logistic regression doesn't "memorize" outcomes — it finds the coefficient values that make the observed data most probable. This process is called Maximum Likelihood Estimation (MLE).
Imagine you have 5,000 historical matches with known outcomes. For each match, your model produces probabilities P(Win), P(Draw), P(Loss). MLE asks: "What coefficient values would make these 5,000 outcomes collectively most likely?"
The math involves computing a "log-likelihood" score — essentially, for each match, you take the log of the probability your model assigned to the actual outcome. If the model said "60% Win" and the team won, the contribution is log(0.6) = −0.51. If the model said "60% Win" but the team lost, the contribution is log(0.4) = −0.92 — a worse score, reflecting the incorrect prediction.
The optimizer (typically gradient descent or Newton-Raphson) iteratively adjusts the coefficients to maximize the total log-likelihood across all matches. Convergence usually happens within 50-200 iterations for a well-scaled dataset.
This is also the origin of "log loss" — the evaluation metric that measures how well-calibrated your probabilities are. A perfect model has log loss = 0. A model that always predicts 50/50 has log loss ≈ 0.693 (the natural log of 2). Lower is better.
Evaluating Your Model: Beyond Simple Accuracy
Saying "my model is 52% accurate" tells you something, but not enough. Football prediction models need to be evaluated on two dimensions: discrimination (can it tell strong teams from weak ones?) and calibration (are the probabilities reliable?).
Accuracy
The simplest metric: what fraction of matches did the model's highest-probability outcome match reality? For three-way football outcomes, anything above 50% is genuinely useful — bookmakers' implied probabilities from closing odds typically achieve 50-53% accuracy on the 1X2 market.
Log Loss
Measures how "surprised" the model is by actual outcomes. A model that says "80% Win" and loses gets heavily penalized; one that says "55% Win" and loses takes a smaller hit. Log loss is the gold standard for evaluating probability predictions because it directly measures calibration. A well-calibrated football model typically achieves log loss between 0.95 and 1.05 on major leagues.
Brier Score
The mean squared error between predicted probabilities and actual outcomes. For a match where the model predicted P(Win)=0.6, P(Draw)=0.25, P(Loss)=0.15 and the home team won, the Brier score contribution is (0.6−1)² + (0.25−0)² + (0.15−0)² = 0.16 + 0.0625 + 0.0225 = 0.245. Lower is better, with 0 being a perfect prediction. A good football model scores between 0.18 and 0.22 on the Brier metric.
Cross-Validation
Never evaluate on training data — that's circular reasoning. The standard approach is k-fold cross-validation: split your historical data into 5 or 10 folds, train on k−1 folds, test on the remaining fold, and average the results. For time-series football data, use chronological splits (train on seasons 1-4, test on season 5) rather than random splits, because you want to predict the future, not the past.
Where Logistic Regression Falls Short
No model is perfect, and logistic regression has known limitations in football prediction:
- Linear decision boundaries: Logistic regression assumes the relationship between features and outcomes is linear (on the log-odds scale). In reality, some interactions are non-linear — the effect of home advantage might depend on team quality, or fatigue might matter more for weaker squads.
- Feature interactions must be engineered manually: If home advantage depends on team quality, you need to create an explicit "home × Elo" interaction term. Tree-based models (random forests, gradient boosting) discover these automatically.
- Draw prediction is inherently difficult: Draws are the hardest outcome to predict in football. Most logistic regression models assign draws probabilities of 20-30%, but the actual draw rate in top leagues is 25-28%. The model often "rounds down" draw probabilities in favor of clear win/loss predictions.
- Cannot capture match dynamics: A logistic regression model trained on pre-match features cannot account for a red card in the 10th minute, a tactical substitution, or a momentum shift. It's a snapshot at kickoff, not a live model.
Logistic Regression vs. Other Approaches
How does logistic regression stack up against more sophisticated methods? The honest answer: surprisingly well, especially for its simplicity.
| Method | Typical Accuracy | Pros | Cons |
|---|---|---|---|
| Logistic Regression | 50-55% | Interpretable, fast, well-calibrated | Linear only, manual feature engineering |
| Random Forest | 51-56% | Handles non-linearity, robust to outliers | Poor probability calibration, harder to interpret |
| Gradient Boosting (XGBoost) | 52-57% | Best raw accuracy, handles complex patterns | Overfits easily, requires careful tuning |
| Neural Networks | 52-58% | Can learn arbitrary patterns | Black box, needs large data, overfits on small datasets |
| Poisson Regression | 50-54% | Models goals directly, good for scorelines | Assumes independence of goals, limited features |
The accuracy gains from moving to complex methods are often 1-3 percentage points. That sounds small, but in the betting market, a 2% edge translates to significant profit over thousands of bets. The trade-off is that logistic regression gives you clear, interpretable coefficients ("home advantage is worth 0.45 in log-odds"), while XGBoost gives you a black box that's harder to trust or debug.
Practical Tips for Building Your Own Model
If you want to build a logistic regression football prediction model, here's a practical roadmap:
- Start with Elo ratings: Use a public Elo system (like ClubElo or World Football Elo Ratings) as your primary feature. It's the single strongest predictor and requires zero engineering.
- Add home advantage as a binary feature: Even a simple 0/1 indicator for home/away improves accuracy by 1-2%.
- Include recent form (5-10 matches): Points per game or goals scored/conceded ratio over the last 5-10 matches captures current trajectory.
- Scale your features: Logistic regression converges faster when features are on similar scales. Standardize Elo differences to a 0-1 range or z-scores.
- Use chronological cross-validation: Train on past seasons, test on the most recent season. Random splits inflate accuracy by leaking future information into training.
- Check calibration plots: Bin your predictions by probability (e.g., all 60-65% predictions) and compare the predicted rate to the actual outcome rate. A well-calibrated model's dots should fall on the 45-degree line.
- Compare to bookmaker closing odds: If your model can't beat the implied probabilities from closing odds (typically 50-53% accuracy), it's not adding value. This is your benchmark, not random guessing.
Key Takeaways
- Logistic regression maps features to probabilities using the sigmoid function, making it the natural choice for predicting categorical outcomes like match results.
- For three-way football outcomes (win/draw/loss), multinomial logistic regression with softmax produces calibrated probability distributions across all three categories.
- Elo ratings are the single strongest feature for football prediction, achieving ~50% accuracy alone on three-way outcomes. Adding home advantage and recent form pushes this to 52-55%.
- Evaluate with log loss and Brier score, not just accuracy. A well-calibrated model with 51% accuracy can be more valuable than an overconfident one at 53%.
- Logistic regression remains competitive with far more complex methods. The 1-3% accuracy gap to gradient boosting is often not worth the interpretability cost, especially for understanding why a prediction was made.
- Start simple, then iterate. A logistic regression with 3-4 well-chosen features outperforms a neural network with 50 poorly-engineered ones. Feature quality matters more than model complexity.