---
language: en
tags:
- pokemon
- competitive-pokemon
- pokemon-showdown
- team-builder
- win-prediction
- xgboost
- pytorch
- transformer
license: mit
---

# FutureSightML

**ML-powered Pokemon team builder that predicts win rates from team composition alone.**

Pre-trained models for **61 competitive formats** across Generations 1-9.

## Downloads

- **Windows Desktop App (CUDA):** [FutureSightML-win-x64.zip](https://huggingface.co/HotHams/FutureSightML/resolve/main/FutureSightML-win-x64.zip) (~3 GB) — Extract and run. Models download automatically on first launch.
- **Model Data:** [model-data.tar.gz](https://huggingface.co/HotHams/FutureSightML/resolve/main/model-data.tar.gz) (~300 MB) — For running from source.

## Architecture

Neural transformer encoder with pairwise matchup matrix branch, ensembled with XGBoost (640+ hand-engineered features). Per-format calibrated weights. Generation-aware mechanics (Mega Evolution, Z-Moves, Dynamax, Tera types).

## Performance

Team-only AUC on held-out test sets (ratings equalized, no Elo leakage):

| Format | Neural AUC | XGB AUC |
|---|---|---|
| Gen 3 OU | 0.859 | 0.828 |
| Gen 4 OU | 0.853 | 0.841 |
| Gen 9 AG | 0.845 | 0.772 |
| Gen 8 Ubers | 0.830 | 0.805 |
| Gen 9 Ubers | 0.812 | 0.752 |
| Gen 9 OU | 0.760 | 0.688 |

For context: Dota 2 draft prediction AUC is 0.66-0.71, Hearthstone deck prediction AUC is 0.65-0.68.

## Source Code

[github.com/HotHams/FutureSightML](https://github.com/HotHams/FutureSightML)