CS2 Major Swiss System Predictor

中文 | English

This project predicts CS2 Major Swiss stage results using ELO + Monte Carlo simulation + Valve's official Buchholz pairing rules.

Basically runs 100k Swiss round simulations, then brute-force searches through 10 million Pick'Em combinations to find the optimal prediction. Calculates team ratings based on historical match data.

v2.0 supports GPU acceleration: Dozens of times faster than v1.0's 16-core CPU (thanks to Tenzray's PR).

Data Format (Must follow this format when modifying match and team data)

data/cs2_cleaned_matches.csv

CSV file without header, 7 columns:

date,team1,score1,score2,team2,tournament,format
2025-11-21,Team A,2,1,Team B,Example Tournament,bo3
2025-11-20,Team C,16,14,Team D,Example League,bo1

• date: YYYY-MM-DD format
• team1, team2: Team names (must match your SEEDED_TEAMS list in code)
• score1, score2: Match scores
• tournament: Tournament name (just for reference, write whatever)
• format: bo1, bo3, or bo5

data/hltv_ratings.txt (optional)

CSV with header, HLTV ratings:

team,Maps,KD_diff,KD,Rating
Team A,120,+250,1.08,1.09
Team B,95,+180,1.06,1.07

Only team and Rating columns matter. Grab latest ratings from HLTV.org if you want more accurate initial values.

Usage

1. Install dependencies

pip install -r requirements.txt

Note: For GPU acceleration, ensure you have CUDA installed. PyTorch will automatically use GPU if available.

2. Configure simulation settings

Edit batchsize.yaml to adjust performance:

simulation:
  num_simulations: 500  # Number of Monte Carlo simulations

performance:
  eval_batch_size: 5000  # Adjust based on GPU memory
  save_every: 1000000    # Checkpoint frequency

device:
  use_gpu: true          # true for GPU, false for CPU
  gpu_id: 0              # GPU device ID (usually 0)

3. Edit seed rankings

Edit cs2_gen_preresult.py (current team names are just examples, change based on actual situation):

SEEDED_TEAMS = [
    "FURIA",          # Seed 1
    "Vitality",       # Seed 2
    # ... 16 teams total, in seed order
]

Just fill in seed rankings. Round 1 matchups auto-generate based on Valve rules (1v9, 2v10, 3v11, ...).

Program will show generated schedule for confirmation before running.

4. Run two-step workflow

Step 1: Generate simulation data

python cs2_gen_preresult.py

This creates output/intermediate_sim_data.json with 100k Swiss round simulations.

Step 2: GPU-accelerated Pick'Em optimization

python cs2_gen_final.py

Core Logic

ELO Rating System

Standard ELO with some improvements:

• Time decay (50-day half-life) - old matches weighted less
• Format weights: BO1=1.0, BO3=1.2, BO5=1.5
• Adaptive K-factor: starts at K=50 for quick adjustment, drops to K=30 after 30 matches
• Blends HLTV ratings with historical data (more history = less HLTV weight)
• Form variance: BO1 ±60, BO3 ±35, BO5 ±20 ELO random fluctuation per match (simulates upsets)

Win probability: P = 1 / (1 + 10^((rating2-rating1)/400))

Buchholz Swiss Pairing

Based on Valve's official rules

Pick'Em Search

Brute-force enumerates all valid combinations:

• 2 teams with 3-0
• 6 teams with 3-1 or 3-2
• 2 teams with 0-3

Total: C(16,6) × C(10,2) × C(8,2) = 10,090,080 combinations

Selects the combo with highest success rate (at least 5 hits) from 100k simulations.

v2.0 GPU Optimization: Uses PyTorch tensors and matrix multiplication for batch evaluation, achieving significant speedup on NVIDIA GPUs.

Final Output

Generated files:

• output/final_prediction.json - Full results with best Pick'Em combination
• output/optimized_report.txt - Human-readable recommendations
• output/intermediate_sim_data.json - Cached simulation data (from step 1)
• gpu_checkpoint.json - Auto-saved progress (can resume if interrupted)

Notes:

• Each team needs at least 10 historical matches for reliable predictions
• Delete gpu_checkpoint.json if you want to restart optimization from scratch
• The two-step design allows you to rerun step 2 with different settings without regenerating simulations
• CPU version cs2_swiss_predictor_cpu.py available for users without GPU

Finally, this project is inspired by claabs/cs-buchholz-simulator