farrelo25/lol-esports-predictor

LoL Esports Predictor for Polymarket

A complete League of Legends esports match prediction system integrated with Polymarket for edge detection. Built on Oracle's Elixir match data (2014–2026) with a dual ELO rating system and XGBoost probability model.

🏗 Architecture

┌────────────────────────────────────────────────────────────┐
│                    DATA LAYER                              │
│  Oracle's Elixir CSV (165 columns, 1.1M+ rows)           │
│  Google Drive auto-sync (daily updates)                    │
└──────────────┬─────────────────────────────────────────────┘
               │
┌──────────────▼─────────────────────────────────────────────┐
│                 FEATURE ENGINEERING                         │
│  • Rolling team stats (15-game window, exp. weighted)      │
│  • Gold/XP/CS diffs at 10, 15, 20, 25 min                │
│  • Objective control (dragons, barons, towers, heralds)    │
│  • Vision metrics (wards, vision score)                    │
│  • Win/loss streaks                                        │
│  Total: 115 pre-game features per matchup                  │
└──────────────┬─────────────────────────────────────────────┘
               │
┌──────────────▼─────────────────────────────────────────────┐
│              DUAL ELO RATING SYSTEM                        │
│  Regional Rating: Updated from domestic league games       │
│  Meta Rating: Updated from international events only       │
│  μ_total = μ_regional + μ_meta                            │
│  → Solves cross-region comparison (LCK vs LPL vs LEC)     │
│  • Margin of victory scaling (gold diff / game length)     │
│  • Time-based σ inflation (off-season uncertainty)         │
│  • Roster change detection → σ increase                    │
└──────────────┬─────────────────────────────────────────────┘
               │
┌──────────────▼─────────────────────────────────────────────┐
│              PREDICTION MODEL                              │
│  XGBoost + Isotonic Regression Calibration                 │
│  • 63% accuracy, 0.67 AUC-ROC, 0.225 Brier Score         │
│  • Time-series cross-validation (no data leakage)          │
│  • Bo3/Bo5 series probability extension                    │
│  • Ensemble: 70% model + 30% ELO                          │
└──────────────┬─────────────────────────────────────────────┘
               │
┌──────────────▼─────────────────────────────────────────────┐
│             POLYMARKET INTEGRATION                          │
│  Gamma API → Market discovery (tag_slug: esports)          │
│  CLOB API → Live bid/ask/mid prices                        │
│  Edge = Model_prob - Market_mid                            │
│  Kelly criterion for position sizing                       │
│  248+ active LoL markets detected                          │
└────────────────────────────────────────────────────────────┘

📊 Model Performance

Validation Results (7,763 train / 1,941 val)

Metric	Value
Accuracy	63.0%
AUC-ROC	0.672
Brier Score	0.224 (calibrated)
Log Loss	0.639 (calibrated)

5-Fold Time-Series Backtest (2022–2026)

League	Accuracy	Brier Score	Games
LCK	65.8%	0.227	1,853
LCP	63.8%	0.227	376
LTA N	67.6%	0.225	213
MSI	62.4%	0.234	234
LEC	61.8%	0.257	1,077
LPL	59.9%	0.250	2,796
Worlds	59.2%	0.251	520
Overall	61.5%	0.245	8,085

Top Features (by XGBoost importance)

1. elo_diff — ELO rating difference (4.7%)
2. diff_opp_towers_mean — Opponent tower diff (3.7%)
3. diff_earned_gpm_mean — Gold per minute diff (2.2%)
4. diff_towers_mean — Tower control diff (1.9%)
5. diff_result_mean — Win rate diff (1.7%)

🏆 League Strength Rankings (auto-updated from international results)

Rank	League	Strength	Method
1	LCK	2.621	MSI/Worlds results
2	LPL	1.827	MSI/Worlds results
3	LEC	0.960	MSI/Worlds results
4	LTA	0.850	Prior (limited intl)
5	CBLOL	0.319	MSI/Worlds results
6	LCP	0.283	MSI/Worlds results

How it works: League strengths start from priors based on historical international performance. They are then updated via a Bayesian ELO-style system every time teams from different leagues meet at MSI, Worlds, FST, or EWC. The system uses K=0.02 for conservative updates.

🎯 Current Top Teams (ELO Ratings)

Rank	Team	League	μ	Games
1	Gen.G	LCK	70.4	906
2	T1	LCK	64.4	910
3	Bilibili Gaming	LPL	55.5	899
4	Dplus Kia	LCK	52.5	903
5	Top Esports	LPL	52.4	896
6	JD Gaming	LPL	50.8	845
7	G2 Esports	LEC	49.4	698

⚡ Quick Start

# Install dependencies
pip install -r requirements.txt

# Download data from Google Drive
python -c "
import gdown
gdown.download_folder('https://drive.google.com/drive/folders/1gLSw0RLjBbtaNy0dgnGQDAZOHIgCe-HH', 
                       output='data/raw', quiet=False)
"

# Train the model (uses 2022-2026 data)
python -m lol_predictor.main train --backtest

# Predict a specific match
python -m lol_predictor.main predict --team-a "T1" --team-b "Gen.G" --league LCK --format bo5

# Compare against market odds
python -m lol_predictor.main edge --team-a "T1" --team-b "Gen.G" --market-odds 0.40 --format bo5

# Scan Polymarket for live LoL markets
python -m lol_predictor.main scan

# Show team ratings
python -m lol_predictor.main ratings --top 30

# Run full scheduler (daily retrain + 5-min scanning)
python -m lol_predictor.main run

📐 How League Strength is Computed

The system uses a dual rating architecture (inspired by PandaSkill):

1. Regional Rating (μ_regional)

• Updated after every domestic league game
• Uses Glicko-style update with margin of victory scaling
• Decays with inactivity (σ inflates 0.8 per month)
• Inflates σ on roster changes (1.5 per position changed)

2. Meta Rating (μ_meta)

• Updated only from international events (MSI, Worlds, FST, EWC)
• Higher K-factor (48 vs 32) because international games are rarer and more informative
• Cross-region games update both team meta-ratings AND league strength scores

3. Combined Rating

μ_total = μ_regional + μ_meta
σ_total = sqrt(σ_regional² + σ_meta²)
P(A wins) = Φ((μ_A - μ_B) / sqrt(σ_A² + σ_B²))

4. Temporal Weighting

• Current season: Full weight, 0.95 per-week exponential decay
• Previous season: 30% of current weight
• 2+ seasons ago: 9% weight (0.3²)

🔄 Scheduling Architecture

Schedule 1: Daily Model Retraining (06:00 UTC)

1. Download latest CSVs from Google Drive (updated daily)
2. Rebuild ELO ratings from scratch (warm-up: 2019-2021, train: 2022-2026)
3. Recompute all rolling features
4. Retrain XGBoost with early stopping
5. Calibrate with isotonic regression
6. Save model artifacts

Schedule 2: Market Scanning (every 5 minutes)

1. Query Polymarket Gamma API for active LoL esports events
2. Parse market questions to extract team names + match format
3. Fetch live CLOB prices (mid, bid, ask, spread)
4. Generate model predictions for each detected matchup
5. Compute edges: model_prob - market_mid
6. Flag edges > 5% with Kelly criterion sizing
7. Log results to data/processed/scan_logs/

📁 Project Structure

lol_predictor/
├── __init__.py
├── config.py                  # All configuration (leagues, params, API URLs)
├── main.py                    # CLI entry point
├── train.py                   # Training pipeline orchestrator
├── predict.py                 # Match prediction + market edge analysis
├── data/
│   └── loader.py              # CSV loading, cleaning, game pairing
├── features/
│   ├── engineering.py         # Rolling stats, H2H, league strength
│   └── elo_system.py          # Dual ELO (regional + meta) rating system
├── model/
│   └── predictor.py           # XGBoost model + calibration + backtesting
├── polymarket/
│   └── scanner.py             # Gamma API + CLOB price fetcher + edge detection
└── scheduler/
    └── runner.py              # Cron-like scheduler for retrain + scan

🔬 Methodology & References

Papers

• PandaSkill (2025): Player-level XGBoost + OpenSkill dual rating. 90.7% accuracy. (arxiv:2501.10049)
• OpenSkill (2024): Bayesian Plackett-Luce rating model. (arxiv:2401.05451)
• Augmented Glicko (2025): Margin of victory scaling. (arxiv:2603.02574)
• Balance Score (2023): Better evaluation metric for esports probability models. (arxiv:2309.06248)

Data Sources

• Oracle's Elixir — Primary data: 165 columns per game (oracleselixir.com)
• Google Drive — Daily updated CSVs (link)
• Polymarket — Live market odds (polymarket.com)

⚠️ Limitations

1. Champion draft not modeled — Adding DraftRec-style features could add +2-5% accuracy
2. Player-level ratings not used — We rate teams, not players. Sub-ins not immediately reflected
3. Patch sensitivity — Meta shifts take ~2 weeks to propagate through rolling windows
4. Minor region calibration — CBLOL, LCP have fewer international games → high uncertainty
5. Market efficiency — Major match odds (LCK, LPL) are generally efficient. Best edges are in minor leagues

📜 License

MIT