Predicting the 2022 FIFA World Cup Champion Using Machine Learning Models

The project tackles a classification prediction problem by analyzing pre‑2020 FIFA World Cup match results sourced from Kaggle, retaining only essential attributes such as home team, away team, goals scored, and match outcome (win, loss, draw).

Initial data cleaning removes irrelevant columns, and additional features are engineered, including the number of tournament appearances, win counts, win rates, and average goals per match for both teams, resulting in an enriched dataset stored as tr_data_after.csv .

Data preprocessing applies z‑score standardization, producing play_score_normal.csv , which is then used to train several models: neural network, logistic regression, support vector machine, decision tree, and random forest. Initial accuracies hover around 60% with notable over‑fitting in tree‑based models.

Error analysis reveals high bias due to limited data, especially the scarcity of draw instances (199 records), and the mismatch between binary classifiers and the three‑class outcome space. Consequently, draw records are removed, and models are retrained.

After refinement, model performances improve modestly; for example, logistic regression achieves 62% test accuracy, while decision tree and random forest suffer from over‑fitting, showing high training but low test accuracy.

To further boost performance, a deep neural network using a Sequential architecture is employed, reaching approximately 92% accuracy, though hyper‑parameter tuning is needed to avoid over‑fitting.

The final step simulates champion prediction by selecting eight of the most frequent 16‑team participants from 2002‑2018, merging their statistics, and applying the deep learning model. The results are presented as reference only, acknowledging limitations such as small sample size, unpredictable group draws, and the exclusion of knockout‑stage dynamics.