Pokedex — 800 Pokemon Under the Microscope

Exploratory analysis and battle-outcome prediction across the complete Pokemon stat universe.

Feature engineering first
Interpretable baselines
Playful without being sloppy

Population: 800+ Pokemon
Features: 6 base stats + typing
Model: Naive Bayes classifier
Method: EDA-driven features

Overview

The Pokemon universe is a closed statistical system: hundreds of creatures, each defined by the same six base stats, two type slots, and a web of type-effectiveness rules. That makes it an unusually clean sandbox for real analysis technique. This notebook takes the full Kaggle Pokedex and treats it seriously — distribution analysis, type-matchup structure, correlation between stats, and a Naive Bayes classifier predicting battle outcomes.

Problem

Toy datasets teach nothing because nothing is at stake and no structure exists to find. Pokemon data is the opposite: deliberately designed balance, legendary outliers, and generation-by-generation power creep are all real patterns waiting in the numbers — if the analysis is careful enough to separate them.

Goals

Map the stat distributions and their outliers — what actually makes a legendary legendary
Quantify type advantages beyond the folklore of the type chart
Train and honestly evaluate a battle-outcome classifier on engineered stat differences

Solution

EDA first: distribution plots for each base stat, correlation heatmaps across Attack, Defense, HP, Speed and the special stats, and side-by-side comparisons across primary and secondary types — surfacing how the designers trade bulk against speed and where legendaries break the curve.

The predictive layer frames battles as a classification problem: given two Pokemon, engineered features (stat differentials, type relationships) feed a Naive Bayes model, chosen deliberately as an interpretable baseline whose assumptions can be checked against the data rather than a black box.

Tooling

Tech stack

Python
pandas
NumPy
seaborn
matplotlib
scikit-learn
Jupyter

Friction

Challenges

Feature engineering beats model choice

Raw stats predict poorly; differences and ratios between the two combatants carry the signal. The lesson generalises to every matchup-prediction problem — encode the comparison, not the entities.

Class balance and the Speed problem

Faster Pokemon win disproportionately, so a model can look smart by learning one feature. Evaluation had to check what the model uses, not just its accuracy.

Takeaways

Lessons learned

An interpretable baseline (Naive Bayes) that you can interrogate is worth more early on than an opaque model with two extra points of accuracy
Game-design data rewards analysis unusually well because its structure was put there on purpose
The EDA-to-model handoff is where most notebooks go wrong; features should come from the exploration, not despite it

Evidence

Dataset

The Complete Pokemon Dataset

Kaggle — Rounak Banik

Base stats, types, abilities, and metadata for 800+ Pokemon across seven generations.

View dataset

Roadmap

Future improvements

Upgrade the battle model with type-effectiveness interaction terms
Generation-over-generation power-creep analysis
An interactive matchup explorer in the browser

Continue

Data Science

Game of Thrones — Survival Analytics

UX Engineering

UX Portfolio & Dashboard Suite