🎯 Aviator Crash Game — ML Signal Analysis

Author: Abdul Hanan
Type: Data Science / Statistical Research
Dataset: 33,000+ real Aviator crash multiplier rounds

---

📌 Project Goal

Investigate whether Machine Learning can predict outcomes in a provably fair RNG-based crash game (Aviator). Followed a rigorous ML pipeline: EDA → Feature Engineering → Leakage Detection → Model Evaluation → Statistical Validation.

🔑 Key Finding

After detecting and fixing data leakage (AUC: 93% → 50%), all models performed at random baseline — confirming the game is mathematically unpredictable.

---

📊 What I Did

• Exploratory Data Analysis on 33,000+ rounds (distributions, zones, outliers)
• Statistical tests: Autocorrelation, Chi-Square, Cramér's V effect size
• Feature engineering: lag features, rolling stats, streak detection
• Detected & fixed critical data leakage in model pipeline
• Compared 3 ML models: Logistic Regression, Random Forest, Gradient Boosting
• Validated with shuffled baseline test (ground truth null model)

---

🛠️ Tech Stack

Python · Pandas · Scikit-learn · SciPy · Statsmodels · Matplotlib · Seaborn

---

📈 Results

Model	AUC-ROC
Logistic Regression	~0.50
Random Forest	~0.50
Gradient Boosting	~0.50
Shuffled Baseline	~0.50

All models = random baseline → Zero predictive signal exists.

---

💡 Key Skills Demonstrated

Skill	Applied
Exploratory Data Analysis	✅
Statistical Testing	✅ Chi-Square, Autocorrelation, Cramér's V
Outlier Detection	✅ IQR method
Feature Engineering	✅ Lag, rolling stats, streak
Data Leakage Detection	✅ 93% → 50% AUC collapse
ML Model Comparison	✅ Chronological split
Shuffled Baseline Test	✅ Null hypothesis validation
Scientific Integrity	✅ Honest null result reported

---

▶️ How to Run

pip install -r requirements.txt

Open aviator_ml_analysis.ipynb in Jupyter or Google Colab

---

⚠️ Conclusion

Any software claiming to "predict" Aviator outcomes is statistically a scam. This project proves it with data.