Machine Learning Models for Predicting Olympic Medal Outcomes

Data Preprocessing & Pipeline Construction

The proposed pipeline integrates two datasets covering 128 years of Olympic history (1896–2024). Text-based columns were cleaned by removing leading and trailing whitespace and converting values to title case. Duplicate records were removed based on the combination of NOC, Year, Season, Host City, Medal, and Event.

Categorical target variables were transformed using one-hot encoding. Numerical features were standardized using StandardScaler, while missing values were handled using KNNImputer. To address class imbalance in the training data, the Synthetic Minority Over-sampling Technique (SMOTE) was applied.

Evaluated Models & Hyperparameters

The framework evaluates 13 machine learning architectures across ensemble, tree-based, linear, non-linear, and neural network approaches:

Ensemble & Tree-Based Models

• XGBoost: learning_rate = 0.001–0.1, n_estimators = 100–500, max_depth = 3–10
• LightGBM: learning_rate = 0.01–0.3, n_estimators = 100–300, num_leaves = 31–100
• Gradient Boosting: learning_rate = 0.01–0.1, n_estimators = 100–200
• Random Forest
• AdaBoost
• Decision Tree
• Extra Trees

Linear & Non-Linear Classifiers

• Logistic Regression
• Polynomial Logistic Regression: polynomial degree = 4, max_iter = 1000, class_weight = balanced
• Support Vector Machine (SVM): RBF kernel, C = 0.1–10, gamma = scale / auto
• K-Nearest Neighbors (KNN): k = 5–30
• Naïve Bayes

Neural Network

• Loss Function: Cross-Entropy Loss
• Optimizer: Adam
• Learning Rate: 0.0001–0.001
• Early Stopping: patience = 5
• Maximum Epochs: 100

Optimization & Validation

Model hyperparameters were optimized using GridSearchCV with 5-fold cross-validation (CV = 5). The complete dataset was divided into an 80:20 training–testing split, with model optimization performed on the training set and final performance evaluated on the held-out test set.