Legendary Pokémon Classification

Predicting whether a Pokémon is legendary using gameplay and stat features.

Overview

This project started as a fun question I had while looking through a full Pokémon dataset I found online:

Can you tell whether a Pokémon is legendary just from a small set of numeric gameplay/stat features?

I trained and compared a couple of classic ML models on a cleaned version of the dataset and ended up with two pretty interesting takeaways:

• even a simple logistic regression model performs extremely well
• a random forest model is almost perfect on this task and stays that way under cross-validation

Additonally, game-design metadata like capture rate, base egg steps, and experience growth turned out to be some of the strongest signals, which surprised me.

Dataset

• Source: Kaggle Pokémon dataset scraped from Serebii
• Rows: 801 Pokémon
• Target: is_legendary
• Final feature subset used in the main models:
  • attack
  • base_egg_steps
  • base_happiness
  • capture_rate
  • defense
  • experience_growth
  • hp
  • sp_attack
  • sp_defense
  • speed
  • weight_kg

What I built

I trained and evaluated:

• Logistic Regression
• Random Forest

To make the comparison more reliable, I used:

• median imputation for missing numeric values
• feature scaling for logistic regression
• stratified 5-fold cross-validation

Results

Cross-validated model performance

Model	Accuracy	Precision	Recall	F1	ROC-AUC
Logistic Regression	0.9651	0.7466	0.9286	0.8260	0.9939
Random Forest	0.9963	0.9600	1.0000	0.9793	0.9996

On a held-out test split, logistic regression reached 98.1% accuracy and the random forest reached 100% accuracy (which I was a little skeptical about). I was a little more careful about the random forest result in this writeup, since one perfect split matters less than performance across multiple folds. The cross-validation numbers ended up telling the more useful story anyway.

Logistic regression coefficients

Random forest feature importances

(Importance based on rf.feature_importances_ in scikitlearn, which uses gini impurity).

Both models pointed to the same small set of standout predictors:

• base_egg_steps
• capture_rate
• experience_growth

That was probably the most interesting part of the project for me. I didn't know that a niche stat like base egg steps (how long the Pokémon takes to hatch from an egg) would be more informative than something like total health or attack stats.

Summary / Key takeaways

• A small set of numeric features is enough to separate legendary from non-legendary Pokémon extremely well.
• Random forest clearly outperformed a strong logistic regression baseline, which suggests some nonlinear structure in the data.
• The most informative features were tied to rarity and progression mechanics, not just battle strength.