Autonomous soccer livestream analysis and prediction market trading system.

All compute runs on cloud infrastructure — OVH H100 for training, Lightning AI for inference. Nothing runs locally.

GitHub Actions → OVH H100 (training) → Lightning AI (inference) → Polymarket/Kalshi

1. Frame Extraction: FFmpeg captures 1fps from HLS/RTMP stream
2. OCR: PaddleOCR extracts score, clock, team names (99.6% accuracy)
3. Event Detection: CLIP ViT-L/14 classifies goals, red cards, VAR reviews
4. Player Detection: YOLOv10-X computes pressure zone signal
5. Prediction: XGBoost + LightGBM ensemble → calibrated win probabilities
6. Trading: Edge detection → Quarter-Kelly sizing → Order placement

Trigger train_model.yml from GitHub Actions:

gh workflow run train_model.yml \
  -f epochs=100 \
  -f grid_search=false

This will:

• Provision OVH H100 instance (~$3.60/hr)
• Build dataset from 8 sources (~2.1M snapshots)
• Train XGBoost + LightGBM ensemble
• Fine-tune YOLOv10-X (100 epochs)
• Fine-tune CLIP ViT-L/14
• Upload model artifacts to Lightning AI Drive
• Terminate OVH instance

Estimated cost: ~$31.50 for ~8.75 hours

Before live trading, run in paper mode:

gh workflow run paper_trade.yml \
  -f stream_url="https://example.com/match.m3u8" \
  -f match_id="match_001"

All signals are logged to SQLite. No orders are placed.

Once paper trading validates:

gh workflow run deploy_stream_worker.yml \
  -f stream_url="https://example.com/match.m3u8" \
  -f match_id="match_001" \
  -f dry_run=false

All configuration via environment variables:

• Live state: score_diff, clock, red_cards, pressure, xG, shots
• Interaction: score_diff × time_remaining (most predictive)
• Pre-match: ELO, form, H2H, squad value, injuries
• Tactical: pressing intensity, xG form, xG conceded
• Context: competition tier, importance, fatigue
• Momentum: recent goals, cards, xG delta

• XGBoost: 1500 estimators, max_depth=6, lr=0.05
• LightGBM: 1500 estimators, num_leaves=63
• Ensemble: Weighted average (50/50) + isotonic calibration
• Validation: GroupKFold by match_id (no data leakage)

Trading halts automatically if:

1. Stream lag > 8 seconds
2. OCR confidence < 0.70 for 5 consecutive reads
3. API errors > 3 in last 60 seconds
4. Bankroll drawdown > 20%
5. Unhandled exception in signal loop

soccer-trade-bot/
├── .github/workflows/     # CI/CD orchestration
├── vision/                # Frame extraction, OCR, CLIP, YOLO
├── model/                 # XGBoost, LightGBM, calibration
├── market/                # Polymarket, Kalshi integration
├── trading/               # Edge calc, Kelly sizing, signal engine
├── data/                  # Dataset building, logging
├── infra/                 # OVH, Lightning AI provisioning
├── config.py              # All configuration
└── main.py                # Entrypoint

MIT