Counter-Strike Performance Forecasting

Time series forecasting for professional Counter-Strike player performance using deep learning.

📊 Overview

This project introduces a rigorous approach to performance forecasting in professional Counter-Strike using time series analysis. Leveraging Rating 1.0 data from HLTV.org, multiple forecasting models are evaluated — including classical statistical methods, machine learning baselines, and deep learning architectures — to predict player performance over horizons of 30, 60, 120, and 150 days.

Key Results:

• All learning-based models significantly outperform the Random Walk baseline across all horizons.
• Deep learning models (GRU, LSTM, Transformer) achieve the strongest overall performance, particularly at short and medium horizons.
• Performance differences between architectures narrow as the forecast horizon increases, indicating a shared modeling ceiling on long-term uncertainty.
• Tree-based methods (Random Forest) perform competitively at short horizons but remain less robust for long-term extrapolation.
• Improvements over Random Walk range from ~50–55% at 30–60 days to ~33–37% at 120–150 days, all at very high statistical confidence.

⚠️ Disclaimer : Data Collection & Terms of Service

This repository includes the scraping logic solely for the sake of methodological transparency and reproducibility.

I strongly discourage using this codebase to scrape HLTV.org actively. HLTV employs strict anti-bot measures (Cloudflare) to protect their infrastructure. The scraper included here is designed for low-volume, compliant data gathering, but using it may still result in IP bans or violate HLTV's Terms of Service.

For Researchers:
The goal of this repository is to allow replication of the results presented in my paper. To avoid burdening HLTV's servers or dealing with anti-bot blocking, I am willing to share the dataset privately for academic replication purposes.

Please contact me at yvann.vincent@gmail.com to request access to the dataset used in the paper.

🚀 Quick Start

Installation with uv

1. Install [uv]:

curl -LsSf https://astral.sh/uv/install.sh | sh

2. Clone the repository:

git clone https://github.com/IUseAMouse/csgo-performance-forecasting.git
cd csgo-performance-forecasting

3. Create and activate virtual environment:

uv venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

4. Install dependencies:

uv pip install -e ".[dev]"

Alternative: pip installation

python -m venv venv
source venv/bin/activate
pip install -e ".[dev]"

📁 Project Structure

csgo-performance-forecasting/
├── data/
│   ├── metadata/
│   ├── raw/
│   ├── processed/
│   └── models/
├── src/csgo_forecasting/
│   ├── data/
│   ├── models/
│   ├── training/
│   └── evaluation/
└── scripts/

🎯 Usage

1. Download Data

make download-data
# or
python scripts/download_data.py

2. Preprocess Data

make preprocess
# or
python scripts/preprocess_data.py --input data/raw/player_data_all.json \
                                   --output data/processed/player_data_all_cleaned.json

3. Train Models

make train-all
# or
python scripts/train_models.py --model lstm --seq-length 90 --out-length 30

Available models: lstm, gru, transformer, ridge, random_forest

4. Evaluate Models

make evaluate
# or
python scripts/evaluate_models.py --models-dir data/models

📊 Dataset

The dataset includes 832 professional Counter-Strike players meeting the following criteria:

• Minimum 200 maps played at top-tier competition
• Team ranked within the global top 50 during active competition

Statistics:

• Average time series length: 1,742 days (~4.8 years)
• Total rating observations: 1,449,752
• Players span the full competitive spectrum, from elite stars to lower-rated professionals

🧠 Models

Implemented Models

1. Baseline

   • Random Walk

2. Statistical Methods

   • AutoARIMA
   • ETS

3. Classical ML

   • Random Forest Regressor
   • Ridge Regression

4. Deep Learning

   • LSTM
   • GRU
   • Transformer (RoPE + RevIN)

Comprehensive Performance Evaluation

Updated model performance across all forecasting horizons. Statistical significance is computed against the Random Walk baseline using paired tests.

Horizon	Model	RMSE	MAE	MAPE	$R^2$	Improv. vs RW	Win Rate	Cohen's $d$	$p<0.05$	$p<0.01$	$p<0.001$
30 Days	GRU	0.0096	0.0057	0.58	0.99	+54.7%	81.0%	0.69	✔️	✔️	✔️
	LSTM	0.0095	0.0061	0.63	0.99	+55.2%	73.8%	0.64	✔️	✔️	✔️
	Transformer	0.0102	0.0067	0.69	0.99	+51.9%	76.2%	0.61	✔️	✔️	✔️
	Random Forest	0.0103	0.0067	0.70	0.99	+51.6%	73.8%	0.62	✔️	✔️	✔️
	Ridge	0.0120	0.0077	0.80	0.98	+43.5%	76.2%	0.61	✔️	✔️	✔️
	Random Walk	0.0212	0.0132	1.36	0.95	–	–	–	–	–	–
60 Days	LSTM	0.0206	0.0131	1.35	0.95	+54.9%	77.4%	0.62	✔️	✔️	✔️
	Transformer	0.0221	0.0141	1.44	0.95	+51.8%	71.4%	0.59	✔️	✔️	✔️
	GRU	0.0225	0.0140	1.44	0.95	+50.9%	75.0%	0.57	✔️	✔️	✔️
	Random Walk	0.0458	0.0284	2.89	0.77	–	–	–	–	–	–
120 Days	GRU	0.0444	0.0306	3.12	0.77	+42.5%	75.0%	0.63	✔️	✔️	✔️
	Transformer	0.0511	0.0355	3.60	0.70	+33.8%	70.2%	0.48	✔️	✔️	✔️
	Random Walk	0.0773	0.0543	5.56	0.31	–	–	–	–	–	–
150 Days	GRU	0.0536	0.0367	3.74	0.66	+37.2%	67.9%	0.63	✔️	✔️	✔️
	Transformer	0.0559	0.0392	3.97	0.63	+34.5%	66.7%	0.59	✔️	✔️	✔️
	Random Walk	0.0853	0.0603	6.07	0.14	–	–	–	–	–	–

Metrics: RMSE, MAE, MAPE (lower is better). All reported improvements are relative to the Random Walk baseline.

🛠️ Development

make format
make lint
make typecheck

make check-all

📝 Citation

🚀 Preprint submitted to HAL/arXiv
Time Series Forecasting for Professional Counter-Strike Player Performance

📄 License

MIT License — see the LICENSE file.

📧 Contact

Yvann VINCENT — yvann.vincent@gmail.com
Project link: https://github.com/IUseAMouse/CSForecast