Gabagool AI

Short-horizon prediction for Polymarket crypto markets. Predict whether the UP contract price will rise by ~$0.01–0.02 within 5 seconds, then train an ML model on real-time order book and spot data to drive a trading bot.

---

Overview

This repository contains:

Component	Stack	Role
Collector	Rust (Tokio, WebSockets)	Streams Polymarket CLOB order book + optional RTDS crypto prices → JSONL.
Training pipeline	Python (PyTorch, pandas)	Builds features from book + spot, stream-based labels (UP/DOWN), trains a small MLP.
Agent	Python	Orchestrates collect → train → evaluate in one run.

The model does not predict the 15-minute market resolution. It predicts a short-period move: UP token price (or spot) will rise by a configurable amount (e.g. 0.01–0.02) within a short horizon (e.g. 5 seconds), so you can trade on that signal.

---

Requirements

• Rust (1.70+) — for polymarket_collector
• Python 3.10+ — for polymarket_prediction (see polymarket_prediction/requirements.txt)
• Polymarket CLOB and RTDS WebSocket access (public endpoints)

---

Quick start

1. Build the collector

cd polymarket_collector
cargo build --release

2. Collect data (BTC 15m market + BTC/USD)

From the repository root:

./polymarket_prediction/scripts/collect_btc15m.sh

Run for at least 20–30 minutes, then stop with Ctrl+C. Data is written to polymarket_prediction/data/raw/market/ and polymarket_prediction/data/raw/crypto/.

3. Train

cd polymarket_prediction
pip install -r requirements.txt   # if needed
python -m src.train \
  --market-dir data/raw/market \
  --crypto-dir data/raw/crypto \
  --market-meta data/raw/processed/market_meta.csv \
  --horizon-sec 5 \
  --min-up-move 0.01

Optional: --use-up-token-price labels by UP contract mid price instead of spot. Model is saved to models/model.pt.

4. Evaluate

python -m src.evaluate \
  --model models/model.pt \
  --data data/processed/features.parquet \
  --horizon-sec 5 \
  --min-up-move 0.01 \
  --crypto-dir data/raw/crypto

Use --use-up-token-price if the model was trained with that option.

5. Agent (collect → train → evaluate in one go)

cd polymarket_prediction
python scripts/run_agent.py --collect-minutes 20 --epochs 80

Use --skip-collect to train/evaluate on existing data only.

---

Project structure

gabagool_AI/
├── polymarket_collector/          # Rust: real-time data collection
│   ├── src/main.rs
│   ├── Cargo.toml
│   └── README.md
├── polymarket_prediction/         # Python: ML pipeline
│   ├── config/default.yaml       # horizon_sec, min_up_move, model, paths
│   ├── data/
│   │   ├── raw/                  # market/*.jsonl, crypto/*.jsonl (from collector)
│   │   └── processed/            # features.parquet, labels_5s.csv, market_meta.csv
│   ├── src/
│   │   ├── data/                 # load_ws, load_crypto, build_features, get_labels
│   │   ├── model/                # MLP (PyTorch)
│   │   ├── train.py
│   │   └── evaluate.py
│   ├── scripts/
│       ├── collect_btc15m.sh     # BTC 15m + btcusdt collection
│       └── run_agent.py          # collect → train → evaluate
│    
│       
│       
│       
└── README.md                     # this file

---

Configuration

Key settings live in polymarket_prediction/config/default.yaml:

Section	Key	Description
data	horizon_sec	Label horizon in seconds (default: 5).
data	min_up_move	Minimum price rise to label UP (e.g. 0.01 or 0.02).
data	use_up_token_price	If true, label by UP token mid; if false, by spot.
data	crypto_symbol	Spot symbol for features/labels (e.g. btcusdt).
model	hidden_dims	MLP hidden layers (e.g. [128, 64, 32]).
training	epochs, batch_size, lr	Training hyperparameters.

Override via CLI (e.g. --horizon-sec, --min-up-move, --use-up-token-price) when running train or evaluate.

License

See repository license file (if present). Use at your own risk; no warranty for trading or financial decisions.