NBA Player Point Prediction

Overview

This project aims to predict the points scored by NBA players based on various game-related statistics. By leveraging machine learning models, it analyzes historical data, identifies patterns, and predicts future performance. The primary goal is to create an accurate model that considers features such as field goals, rebounds, assists, and more.

Introduction

Predicting player performance in professional sports has become an essential tool for strategic decision-making. Accurate predictions can guide coaches, analysts, and fantasy sports enthusiasts in assessing a player’s potential and making informed decisions about team composition, game strategy, and overall performance.

In basketball, metrics like points scored, field goal percentage, assists, and rebounds correlate with future game outcomes. Understanding these relationships helps uncover trends and make informed predictions.

The dataset used contains performance data from basketball players across multiple past games, including:

Objective: Predict the number of points (PTS) a player will score in their next game based on historical performance data.
Factors such as player form, opponent strength, and team dynamics are considered to answer: How can we predict a player’s points in the next game based on their past performance, opponent, and other relevant factors?

By employing machine learning techniques like regression models and feature engineering, this project seeks to identify patterns indicative of future player performance.

Dataset Overview

The dataset was sourced using NBA-API, which fetches data from NBA.com. For this analysis, I am using LeBron James as an example.

Key Statistics

Introduction of Columns

Here are the columns included in the dataset:

  1. SEASON_ID: The season in which the game was played.
  2. WL: Win/Loss outcome of the game.
  3. MIN: Minutes played by the player in the game.
  4. FGM: Field Goals Made.
  5. FGA: Field Goals Attempted.
  6. FG_PCT: Field Goal Percentage (FGM/FGA).
  7. FG3M: Three-Point Field Goals Made.
  8. FG3A: Three-Point Field Goals Attempted.
  9. FG3_PCT: Three-Point Field Goal Percentage (FG3M/FG3A).
  10. FTM: Free Throws Made.
  11. FTA: Free Throws Attempted.
  12. FT_PCT: Free Throw Percentage (FTM/FTA).
  13. OREB: Offensive Rebounds.
  14. DREB: Defensive Rebounds.
  15. REB: Total Rebounds (OREB + DREB).
  16. AST: Assists.
  17. STL: Steals.
  18. TOV: Turnovers.
  19. PTS: Points scored by the player in the game.

Why This Dataset Matters

This dataset provides a rich, real-world representation of basketball player performance across seasons.

Predicting player points is more than analyzing past performance. It requires understanding nuances like game location, opponent strength, and player form. This project combines these factors to predict future performance, offering actionable insights for coaches, analysts, and enthusiasts alike.

Conclusion

This project harnesses the power of machine learning to enhance our understanding of player performance dynamics. Accurate point predictions can optimize game strategies, support fantasy sports decisions, and offer deeper insights into player and team performance.

By analyzing historical data, we aim to provide practical and impactful solutions for individual and team-level performance analysis.

Data Cleaning and Exploratory Data Analysis

Data Cleaning

To ensure the dataset was ready for analysis and accurately reflected the data-generating process, I performed the following data cleaning steps. Each step and its rationale are described below:

1. Identifying Home vs. Away Games

2. Extracting Opponent Teams

3. Standardizing Season Format

4. Mapping Win/Loss Outcomes

5. Handling Missing Data

6. Feature Engineering with Rolling Averages

7. Column Selection

Cleaned DataFrame Head

The head of the cleaned DataFrame is shown below, highlighting the transformed and newly added columns:

SEASON HOME OPP PTS AST STL TOV FG_PCT MIN_Roll PTS_Rolling AST_Rolling FG_PCT_Roll
2018 1 CHA 27 9 0 6 0.579 32 27 9 0.579
2018 1 WAS 23 14 1 3 0.55 33 25 11.5 0.5645
2018 1 SAC 29 11 2 4 0.409 33.6667 26.3333 11.3333 0.512667
2018 1 BKN 25 14 1 8 0.32 34.5 26 12 0.4645
2018 0 NYK 33 8 0 2 0.423 34.6 27.4 11.2 0.4562
2018 0 TOR 29 6 1 4 0.522 34.6 27.8 10.6 0.4448
2018 0 CHI 36 4 2 5 0.652 34.4 30.4 8.6 0.4652
2018 1 BOS 30 12 0 3 0.565 33 30.6 8.8 0.4964
2018 1 DEN 31 7 1 4 0.591 31.8 31.8 7.4 0.5506
2018 1 LAC 27 6 1 2 0.5 33.2 30.6 7 0.566

Univariate Analysis

Distribution of Points Scored (PTS)

The histogram of points scored (PTS) shows a distribution resembling a normal curve, with most values concentrated around the mean. This suggests that the player’s scoring performance is consistent, with fewer games at the extremes of very low or very high points scored. The visualization provides insight into the player’s typical scoring range and highlights their reliability as a scorer.

Distribution of Minutes Played (MIN)

The histogram of minutes played (MIN) displays a distribution that appears to be centered around a moderate range, with most values clustered near the mean. This suggests that the player’s playing time is relatively consistent across games, with fewer instances of extremely low or high minutes played. The visualization offers insight into the player’s typical role in the team, indicating a steady involvement in games and contributing to an understanding of their workload and stamina.

Bivariate Analysis

Relationship Between Assists and Points Scored

The scatter plot visualizing the relationship between assists (AST) and points scored (PTS) shows a slightly negative trend, where higher assist numbers tend to correspond with slightly lower point totals. This suggests that as the player focuses more on facilitating plays and passing the ball, they score fewer points themselves. The visualization highlights the trade-off between playmaking and scoring, reflecting the player’s role as more of a facilitator in the offense.

Points Scored by Home/Away

The box plot comparing points scored (PTS) at home versus away games shows similar median values for both situations. However, when the player is at home, there are a couple of outliers where the points scored are notably higher than the rest of the games. This suggests that while the player’s scoring performance is consistent whether playing at home or away, there are certain home games where their performance spikes significantly, possibly due to factors like crowd support or matchups.

Interesting Aggregates

Win Percentage by Home/Away

The table below shows the win percentage for the player in home and away games. The win percentage is calculated as the ratio of wins to the total number of games played in each setting (home/away).

HOME/AWAY Win Percentage (%)
Away 54.44
Home 61.46

This suggests that the player performs better in home games, with a higher win percentage compared to away games.

Imputation

Since the number of rows remains the same before and after filtering out the missing values, imputation is not needed in this case.

Framing a Prediction Problem

Prediction Problem and Type

The problem at hand is a regression problem where we are predicting the number of points scored (PTS) by LeBron James in each game based on various features such as minutes played, field goal percentage, assists, etc. Since we are predicting a continuous numerical value (points scored), this is a regression task.

Response Variable

The response variable, or the target we are predicting, is the points scored (PTS) in each game. We chose PTS as the target because it represents a key performance metric in basketball, and understanding how different features (like minutes played, shooting percentage, assists, etc.) influence scoring can provide valuable insights into a player’s performance.

Evaluation Metric

To evaluate our model, we are using Mean Squared Error (MSE). MSE is chosen because it penalizes larger errors more heavily, which is useful when dealing with continuous variables like points scored. It is a standard evaluation metric for regression models, allowing us to assess how well the model’s predictions align with the actual values. In our case, lower MSE values indicate better model performance, as the predicted points are closer to the true points scored.

Information Known at the Time of Prediction

At the time of prediction, we would know the following features for a given game (based on the input data):

We would not have access to the PTS for the game we are predicting, but we do have access to all the other features listed above.

Model Training

In this regression task, we train the model using the following features:

By using these features, we aim to predict the points scored by LeBron James in a given game while ensuring that the features selected are known at the time of prediction and do not include future game information.

Baseline Model

Model Description

For this prediction problem, I am using a Ridge Regression model. Ridge regression is a regularized linear regression model that helps prevent overfitting by penalizing large coefficients. This is useful because our dataset may have some multicollinearity (when independent variables are highly correlated), and regularization helps improve the generalization of the model.

The model is built in a Pipeline that includes a ColumnTransformer to preprocess the data. Specifically, the preprocessing consists of:

  1. One-Hot Encoding for the OPP (opponent) column to convert the categorical feature into numerical values.
  2. Standard Scaling for the continuous numerical features to normalize them so that they have zero mean and unit variance. This step is important for models like Ridge Regression that are sensitive to the scale of the data.

The final model then fits a Ridge Regression model to the processed features.

Features in the Model

The model includes the following features:

  1. Quantitative Features (Numerical):
    • MIN: Minutes played in the game.
    • FGM: Field goals made.
    • FGA: Field goals attempted.
    • FG_PCT: Field goal percentage.
    • FG3M: 3-point field goals made.
    • FG3A: 3-point field goals attempted.
    • FG3_PCT: 3-point field goal percentage.
    • FTM: Free throws made.
    • FTA: Free throws attempted.
    • FT_PCT: Free throw percentage.
    • OREB: Offensive rebounds.
    • DREB: Defensive rebounds.
    • REB: Total rebounds.
    • AST: Assists.
    • STL: Steals.
    • TOV: Turnovers.
    • MIN_Roll: Rolling average of minutes played over the last 5 games.
    • FG_PCT_Roll: Rolling average of field goal percentage over the last 5 games.
    • FT_PCT_Roll: Rolling average of free throw percentage over the last 5 games.
    • PTS_Rolling: Rolling average of points scored over the last 5 games.
    • AST_Rolling: Rolling average of assists over the last 5 games.

    Total Quantitative Features: 21

  2. Nominal Features (Categorical):
    • OPP: Opponent team, which is one-hot encoded during preprocessing.

    Total Nominal Features: 1

  3. Ordinal Features:
    • HOME: Whether the game was played at home (1 for home, 0 for away). This feature is treated as nominal in the model since it represents categorical data with no inherent order.

    Total Ordinal Features: 1

Model Performance

To evaluate the performance of the model, we used the Mean Squared Error (MSE), which is a standard metric for regression tasks. The MSE was computed on both the training and testing datasets:

These results suggest that the model is performing reasonably well, with the training MSE being quite low, indicating that the model has learned well on the training data. However, the test MSE is higher, which could indicate that the model is not generalizing perfectly to unseen data. The cross-validation MSE provides a good balance, showing that the model performs moderately well across different splits of the data.

Model Evaluation

While the model is performing decently, I believe there is room for improvement. Here are a few reasons why:

  1. Model Regularization: While Ridge Regression is helpful in reducing overfitting, we could explore more sophisticated models such as Random Forests, Gradient Boosting Machines (GBMs), or even Neural Networks to capture non-linear relationships that Ridge Regression might miss.

  2. Feature Engineering: The rolling averages provide useful information, but we may benefit from creating additional features or interactions between existing features, such as player efficiency ratings or adjusted shooting percentages based on opponent defense.

  3. Hyperparameter Tuning: The regularization strength (alpha) in Ridge Regression was not optimized. Hyperparameter tuning via techniques such as Grid Search or Randomized Search could help find a better model configuration.

  4. Evaluation Metrics: We only used MSE as the evaluation metric. While it’s a good standard, using other metrics like R-squared, Root Mean Squared Error (RMSE), or even mean absolute error (MAE) could give us more insights into model performance.

In conclusion, the current model is a solid baseline, but there is still potential for improvement. Further experimentation with more advanced models and additional feature engineering could enhance the prediction accuracy.

Final Model

In the final model, I utilized a combination of transformed and scaled features. The key transformations and additions include:

  1. Quantile Transformation on the following numerical features: MIN (minutes played), REB (rebounds), AST (assists), STL (steals), TOV (turnovers), FGA (field goals attempted), and FTA (free throws attempted). The purpose of the QuantileTransformer was to reduce the influence of outliers and skewed distributions. These features are vital for predicting points (PTS), as they represent critical aspects of a player’s performance. Applying quantile transformation to these variables ensures that extreme values do not disproportionately affect the model, allowing it to focus on general trends rather than individual outliers.

  2. Standard Scaling was applied to all numerical features, ensuring that they were centered around zero with unit variance. This standardization is particularly beneficial for models like Random Forest, as it makes the features comparable and avoids giving undue weight to certain features based solely on their scale.

  3. One-Hot Encoding was applied to the OPP (opponent) column. By converting this categorical variable into binary features, the model can properly handle the categorical nature of the opponent variable, enabling it to learn the relationships between the opponent type and the target variable (PTS).

These features were chosen because they capture key aspects of the data-generating process for basketball games. Minutes played, rebounds, assists, and other statistics are well-known indicators of player performance, and transforming these features ensures that the model can handle a variety of distributions in the data effectively.

Modeling Algorithm and Hyperparameters

The RandomForestRegressor was chosen for this task because it is a powerful ensemble learning method capable of capturing complex, non-linear relationships in the data. Random forests are particularly useful when dealing with a mix of numerical and categorical features, as they do not require the data to be linearly separable and can naturally handle missing data and outliers.

Hyperparameter Tuning:

To improve the performance of the Random Forest model, GridSearchCV was used for hyperparameter optimization. The grid search tested several combinations of hyperparameters to identify the best model:

The best combination of hyperparameters identified by GridSearchCV was:

These hyperparameters were selected because they resulted in the lowest Mean Squared Error (MSE) during cross-validation, indicating that this combination of parameters best balanced model complexity and generalization.

Model Performance

The performance of the final model was evaluated using Mean Squared Error (MSE), which measures the average squared difference between the actual and predicted values.

Baseline Model Performance:

The Baseline Model performance was as follows:

Performance Comparison:

Conclusion

The final model, incorporating Quantile Transformation, Standard Scaling, and One-Hot Encoding, along with RandomForestRegressor and optimized hyperparameters, does not show a clear improvement over the baseline model. In fact, it exhibits signs of overfitting, as seen from the higher Test MSE and Cross-Validation MSE. Despite the more complex feature engineering and hyperparameter optimization, the simpler baseline model appears to generalize better, with a lower test and cross-validation error. This suggests that for this particular problem, a simpler model might be more effective in predicting player points.