ATP Tennis Match Predictor and Data Analysis

By M. Malik Tekin (mmtekin@umich.edu)

An analysis of ATP tennis matches to predict match outcomes and explore patterns in professional tennis. This project applies practical data science techniques to understand and predict tennis match results using historical ATP tour data.

Introduction

This project analyzes historical ATP tennis match data (after transformation, 6152 rows since every row from original dataset had to be split into 2 rows) to uncover patterns in professional tennis and build a pre‑match predictor of match outcomes. I explore how variables like player ranking, age, and Elo rating affect winning probability, then train a model that—using only these pre‑match stats—forecasts which player will win. Such insights can aid coaches, bettors, and tennis analysts in understanding match dynamics before a ball is struck.

Data Cleaning and Exploratory Data Analysis

Data Cleaning Process

My data cleaning process involved several sophisticated transformations to prepare the ATP match data for analysis:

  1. Elo Rating Computation
    • Implemented a dynamic Elo rating system for all players
    • Processed matches chronologically to maintain accurate rating progression
    • Initial rating set to 1500, following the standard Elo convention where 1500 represents an average skill level
    • Ratings updated after each match using the standard Elo formula: R_new = R_old + K * (Actual - Expected)
    • Used K-factor of 32 to balance rating responsiveness with stability
  2. Data Restructuring
    • Transformed the original match-centric format to a player-centric format
    • Each match was split into two rows to represent both perspectives (winner and loser)
    • This doubled my dataset size but provides balanced training data for my prediction model
  3. Feature Engineering
    • Converted tournament dates to datetime format
    • Standardized player attributes (rank, age, hand, rank points)
    • Created symmetric features for both players (p1 and p2)
    • Added binary outcome variable ‘did_p1_win’
  4. Final Dataset Structure Key features in transformed dataset:
    • Tournament info: date, surface, draw_size, level, round
    • Player 1 stats: rank, age, hand, rank_points, elo
    • Player 2 stats: rank, age, hand, rank_points, elo
    • Outcome: did_p1_win (binary)

This transformation approach ensures my model can learn from both winning and losing perspectives, while maintaining the temporal nature of the Elo rating system.

Univariate Analysis

My univariate analysis revealed several interesting distributions in the tennis match dataset:

  1. Player Rankings Distribution

The distribution of player rankings shows a significant skew towards higher-ranked players (ranks 1-49), with 2,778 matches compared to 1,831 matches for players ranked 50-99. This makes sense as my dataset focuses on major tournaments rather than challenger events. This is just good to know for understanding the dataset.

  1. Age Distribution

The age distribution reveals that most players are between 25-29 years old, with a steep drop-off after age 29. This suggests that players’ physical prime typically lasts until around age 29. This makes it even more impressive that someone like Novak Djokovic is still playing at the highest level at age 37.

  1. Court Surface Distribution

Hard courts dominate the dataset, followed by clay, with grass courts having the fewest matches. This distribution will be important when evaluating my model’s performance across different surfaces. This could also mean that the model is biased towards hard courts, which is something to keep in mind.

  1. Player Handedness

The vast majority of players are right-handed, which aligns with general population statistics (roughly 90% of the population is right-handed and this gave 86%).

Bivariate Analysis

My bivariate analysis focused on understanding how different factors relate to match outcomes:

  1. Age Difference and Win Rate

This analysis reveals that:

  1. Handedness and Win Rate

Surprisingly, my analysis shows no significant advantage for left-handed players, contrary to common tennis wisdom. This suggests that handedness might not be a crucial feature for match prediction. This does go against the common belief that left-handed players have an advantage over right-handed players due to them being able to hit forehands to their opponents’ backhands. This plot seems to suggest that there is no advantage to being left-handed which is important to note.

  1. Ranking Difference and Win Rate

The relationship between ranking difference and win probability is strong and clear:

This strong correlation makes ranking difference a potentially powerful predictor for my model.

Interesting Aggregates

I analyzed how player ranking tiers interact with court surfaces to influence win rates:

This pivot table reveals several fascinating patterns about the relationship between player ranking and surface-specific performance:

  1. Top 10 Players’ Dominance:
    • Elite players (Top 10) maintain impressive win rates >70% across all surfaces
    • Highest win rates on grass (77%) and clay (76%)
    • Relatively lower win rate on hard courts (71%), suggesting surface might be a slight equalizer.
    • Possible reason for that is that most players play on hard courts and are therefore more familiar with the surface.
  2. Mid-Tier Performance (11-50):
    • Players ranked 11-25 show consistent performance (~60% win rate) across surfaces
    • Rankings 26-50 maintain win rates around 48-51%, with grass courts showing slightly better outcomes but not significantly so.
  3. Lower Rankings Impact:
    • Players ranked 100+ struggle significantly with ~40% win rates across all surfaces
    • The win rate is evenly distributed across all surfaces which means that the surface is not a significant factor for lower ranked players as compared to the top players (grass being higher for top players)

This analysis suggests that while top players maintain their advantage across all surfaces, the impact of ranking on win probability varies very little by surface type. Hard courts, being the most common surface, show slightly more competitive matches across ranking tiers, possibly due to players’ greater familiarity with the surface.

Handling Missing Values

My data cleaning process was designed to minimize the impact of missing values:

  1. Elo Ratings
    • New players automatically assigned 1500
    • Ratings evolve naturally through match play
    • No imputation needed as system is self-contained
  2. Player Attributes
    • Rank and rank points: Used as-is (missing values indicate unranked players)
    • Age distribution before and after imputation:

As shown in the distributions above, age imputation had minimal impact on my dataset since the vast majority of players had their ages recorded. The distributions before and after imputation are nearly identical, confirming that missing age values were not a significant concern for my analysis.

  1. Tournament Information
    • All selected fields (date, surface, level, round) are complete
    • Draw size is consistently recorded

The transformation process ensured that the final dataset contains no missing values that would impact the prediction task.

Framing a Prediction Problem

I have formulated my analysis as a binary classification problem to predict tennis match outcomes.

Prediction Task Details

Evaluation Metric

I chose accuracy as my primary evaluation metric for several reasons:

  1. The dataset is perfectly balanced due to my transformation approach (each match appears twice with players swapped)
  2. False positives and false negatives are equally important in tennis match prediction
  3. Accuracy provides an intuitive interpretation for stakeholders: “What percentage of matches did we predict correctly?”

Alternative metrics like F1-score or AUC-ROC were considered but not chosen since:

Time of Prediction Considerations

The model only uses information that would be known before a match starts:

This ensures the model makes realistic predictions that could be used in practice, without relying on in-match statistics or post-match data.

Baseline Model

For my initial prediction approach, I implemented a simple decision tree classifier (with max_depth=3) as a baseline model. This choice provides an interpretable starting point while incorporating both ranking and player characteristics.

Model Architecture

Features

The baseline model uses four features:

  1. Quantitative Features (2)
    • Player 1 ATP Ranking (p1_rank)
    • Player 2 ATP Ranking (p2_rank)
  2. Nominal Features (2)
    • Player 1 Handedness (p1_hand)
    • Player 2 Handedness (p2_hand)

Preprocessing Pipeline

All preprocessing steps are encapsulated in a single sklearn Pipeline:

  1. For Quantitative Features:
    • Missing values imputed using median strategy
    • Features standardized to zero mean and unit variance
  2. For Nominal Features:
    • One-hot encoding applied to handedness
    • Unknown categories handled gracefully with ‘ignore’ strategy

Performance

Evaluation

This baseline model’s performance is relatively weak, suggesting significant room for improvement. The modest improvement over random guessing (only 8.57 percentage points) indicates that:

  1. The selected features, particularly player handedness, may not be strong predictors of match outcomes
  2. The simple decision tree structure may be insufficient to capture complex patterns in tennis match outcomes
  3. Important predictive features from our exploratory analysis (like age and surface) are not yet incorporated

This baseline provides a clear starting point for improvement in our final model, where I can incorporate additional features identified in our exploratory analysis and potentially use more sophisticated modeling techniques.

Final Model

To improve upon the baseline model’s performance, I developed a more sophisticated approach using gradient boosting and carefully engineered features that capture the complex dynamics of tennis matches.

Feature Engineering

I engineered several new features that capture different aspects of match dynamics:

  1. Player Differentials
    • Rank Difference (rank_diff): Captures skill gap between players
    • Age Difference (age_diff): Reflects physical and experience disparities that are important as we saw in the exploratory data analysis
    • Ranking Points Difference (points_diff): Another measure of recent performance
    • Elo Difference (elo_diff): Dynamic rating system that updates after every match
  2. Tournament Context
    • Round Ordinal (round_ord): Encoded match stages (R128→1 to F→7)
    • Tournament Level (tourney_level_ord): Encoded tournament prestige (Grand Slam→5 to Tour→0) as nerves probably play a role in the performance of the players
    • Seasonal Features (month_sin, month_cos): Cyclical encoding of tournament month to capture seasonal effects if any. This captures grass, clay, and hard court seasons.
  3. Playing Conditions
    • Surface: One-hot encoded court types (Hard, Clay, Grass)
    • Player Handedness: Encoded left/right-handed matchup dynamics

Feature Preprocessing Pipeline

The model employs a sophisticated preprocessing pipeline with different strategies for each feature type:

  1. Numeric Features (rankings, differences):
    • Missing value imputation
    • Standard scaling (zero mean, unit variance)
  2. Quantile Features (points difference):
    • Missing value imputation
    • Quantile transformation to normal distribution
  3. Ordinal Features (round, level, seasonal):
    • Passed through as already properly encoded
  4. Categorical Features (surface, handedness):
    • One-hot encoding with unknown category handling

Model Selection and Training

I chose the Histogram-based Gradient Boosting Classifier for several reasons:

  1. Efficient handling of mixed data types
  2. Ability to capture non-linear feature interactions
  3. Strong performance on tabular data

Hyperparameter Tuning

Used 5-fold cross-validation with GridSearchCV to optimize:

Best parameters found:

{
    'max_iter': 100,
    'max_depth': 10,
    'learning_rate': 0.01,
    'l2_regularization': 0
}

Performance Improvement

The final model achieved significant improvements over the baseline:

This improvement suggests that the engineered features successfully capture important match dynamics that the baseline model missed.

Performance Visualization

Confusion Matrix

The confusion matrix shows balanced performance across both positive and negative predictions, indicating the model isn’t biased toward either outcome.

Feature Impact Analysis

The most influential features in predicting match outcomes were:

  1. Ranking difference (capturing skill gap)
  2. Elo difference (recent performance)
  3. Tournament level (match importance)
  4. Surface (playing conditions)

This aligns with tennis expertise, where ranking and recent form are typically the strongest predictors of match outcomes.

Model Limitations

While my model shows good improvement, some limitations remain:

  1. Doesn’t capture player-specific surface preferences such as Rafael Nadal’s clay court dominance or Roger Federer’s grass court ability
  2. Cannot account for injuries or fatigue
  3. May not fully capture momentum from recent tournament performance

These limitations suggest potential areas for future model improvements, such as incorporating player-surface win rates or recent match history features.

This project was created as part of EECS 398: Practical Data Science at the University of Michigan.