ML-powered Pokemon team builder that predicts win rates from team composition alone.
FutureSightML uses a neural transformer encoder with a pairwise matchup matrix branch, combined with an XGBoost ensemble, to evaluate Pokemon teams before a single move is made. It ships pre-trained models for 61 competitive formats across Generations 1-9 and includes a retro-styled desktop GUI, a FastAPI backend, and a genetic algorithm team builder.
Pre-game win prediction from team composition alone (held-out test set, no rating leakage):
Select formats shown (61 total):
| Format | Neural AUC | XGB AUC |
|---|---|---|
| Gen 1 OU | 0.8034 | 0.8023 |
| Gen 3 OU | 0.8590 | 0.8276 |
| Gen 4 OU | 0.8531 | 0.8413 |
| Gen 5 OU | 0.7577 | 0.7429 |
| Gen 6 OU | 0.7640 | 0.7649 |
| Gen 7 OU | 0.7959 | 0.7449 |
| Gen 7 Ubers | 0.8212 | 0.7797 |
| Gen 8 OU | 0.7611 | 0.7223 |
| Gen 8 Ubers | 0.8302 | 0.8047 |
| Gen 8 VGC 2020 | 0.8008 | 0.7807 |
| Gen 9 OU | 0.7598 | 0.6876 |
| Gen 9 Ubers | 0.8119 | 0.7517 |
| Gen 9 AG | 0.8448 | 0.7717 |
| Gen 9 Monotype | 0.8231 | 0.7905 |
| Gen 9 VGC 2026 | 0.7351 | 0.7117 |
All numbers are team-only AUC on a held-out test set with equalized ratings (no Elo leakage). For context: Dota 2 draft prediction AUC is 0.66-0.71, Hearthstone deck prediction AUC is 0.65-0.68.
Team A (6 Pokemon) ──► PokemonEncoder ──► TeamEncoder (3-layer Transformer)
├──► CrossAttention
Team B (6 Pokemon) ──► PokemonEncoder ──► TeamEncoder ──┤
└──► MatchupMatrixBranch (6x6 pairwise)
│
▼
MatchupHead ──► P(Team A wins)
XGBoost (618+ engineered features) ──────────────────────► Calibrated Ensemble
- Neural model: Species/move/item/ability embeddings, continuous stat features, 3-layer transformer encoder with cross-attention and a blade-chest matchup matrix decomposition
- XGBoost: 618+ hand-engineered features including type coverage, stat distributions, threat matrices, item/ability effects, and dual-rating features
- Ensemble: Per-format calibrated weights, saved to JSON
- Gen-aware: Type charts, stat formulas, move categories, and learnsets are all generation-correct
- OS: Windows 10+ (desktop app), or any OS with Python 3.10+ (from source)
- GPU: NVIDIA GPU with CUDA recommended for fastest inference. CPU-only works but is slower.
- RAM: 4 GB minimum, 8 GB recommended
- Disk: ~3.5 GB for desktop app, ~500 MB for source + models
Download from HuggingFace (~3 GB). Extract, run FutureSightML.exe. Models download automatically on first launch.
# Clone and install
git clone https://github.com/HotHams/FutureSightML.git
cd FutureSightML
pip install -e .
# Download pre-trained models (~310 MB)
python scripts/download_models.py
# Start the server (opens browser automatically)
python scripts/run_server.pyEverything runs locally. Nothing leaves your machine.
- Team generator — A genetic algorithm evolves teams to maximize predicted win rate against the metagame
- Win rate prediction — Paste or build a team, see its predicted win rate and matchup spread
- Deep analysis — Archetype detection, team roles, win conditions, game plan, speed tiers, type coverage, Tera suggestions, and top threats
- Head-to-head — Pit any two teams against each other for a win probability breakdown
- Damage calculator — Calculate damage between any two Pokemon with full move/item/ability context
- Metagame overview — Usage statistics, top Pokemon, common sets, items, and abilities per format
- Cross-format tier list — Compare Pokemon viability across all supported formats at a glance
- Battle simulator — Monte Carlo battle simulation with turn-by-turn replay
- Import/Export — Paste Showdown teams in, export generated teams out
- Full API — FastAPI backend with auto-generated docs at
/docs
61 formats with pre-trained models across Generations 1-9, including OU, UU, RU, NU, Ubers, VGC, Doubles OU, LC, Monotype, National Dex, 1v1, AG, and more. See config.yaml for the full list.
python scripts/build_exe.pyThis builds the PyInstaller backend + Electron frontend into a standalone desktop app.
The FastAPI server provides auto-generated docs at http://localhost:8000/docs.
Key endpoints:
POST /api/team/generate- Generate an optimized team for a formatPOST /api/team/analyze/full- Analyze a team's predicted win rate and matchupsGET /api/meta/{format}- Get metagame usage statisticsPOST /api/battle/simulate- Monte Carlo battle simulation
FutureSightML/
├── showdown/ # Main Python package
│ ├── models/ # Neural network architectures
│ ├── data/ # Feature extraction, damage calc, data loading
│ ├── scraper/ # Replay scraping from Showdown API
│ ├── teambuilder/ # Genetic algorithm, evaluation, analysis
│ ├── simulator/ # Battle simulator with Monte Carlo estimation
│ ├── api/ # FastAPI server and app state
│ └── utils/ # Constants, logging, type system
├── gui/ # Electron + React frontend
│ └── static/index.html # Single-file React app with retro UI
├── scripts/ # CLI entry points
│ ├── run_server.py # Start the API server
│ ├── download_models.py # Download pre-trained models
│ ├── build_exe.py # Build desktop app
│ └── train_model.py # Train models (for development)
├── data/ # Model checkpoints, pool data
├── tests/ # Test suite (90 tests)
├── config.yaml # Format and training configuration
└── FutureSightML.spec # PyInstaller build spec
Contributions are welcome! Please:
- Fork the repository
- Create a feature branch (
git checkout -b feature/my-feature) - Run existing tests (
pytest tests/ -v) - Submit a pull request
- HolidayOugi for the 30.5M replay dataset on HuggingFace
- Smogon and Pokemon Showdown for usage statistics, replay data, and the battle simulator
- pkmn.cc for the Showdown data files (pokedex, moves, items, abilities, learnsets)
MIT