Survivor Winner Prediction Model

A state-of-the-art machine learning system that predicts Survivor season winners using an ensemble approach trained on 48 seasons of historical data. The model achieves 91.7% accuracy on winner predictions across Seasons 1-48.

🏆 Performance Results

The model was validated on all past seasons (S1-S48) with impressive accuracy:

Metric	Accuracy	Seasons Correct
Winner	91.7%	44/48 seasons
Top 2	91.7%	44/48 seasons
Top 3	79.2%	38/48 seasons
Top 4	81.2%	39/48 seasons
Top 5	89.6%	43/48 seasons

🎯 Season 49 Final Predictions

Rank	Name	Probability	Confidence
🥇 1	Savannah	35.89%	High
🥈 2	Rizo	22.85%	Moderate
🥉 3	Sage	18.40%	Moderate
4	Sophi	16.77%	Moderate
5	Kristina	6.10%	Low

✅ Actual Season 49 Results

The Season 49 finale aired on December 17, 2025. Here are the actual results compared to model predictions:

Actual Place	Contestant	How They Finished	Predicted Rank	Jury Votes
🥇 1st (Winner)	Savannah Louie	Won 5-2-1 jury vote	1st ✅	5
🥈 2nd	Sophi Balerdi	Runner-up; won final immunity	4th	2
🥉 3rd	Sage Ahrens-Nichols	Taken to FTC by Sophi	3rd	1
4th	Rizo Velovic	Lost fire-making to Savannah	2nd	—
5th	Kristina Mills	Voted out 4-1 at F5	5th ✅	—

Key Finale Events

• Savannah won her record-tying 4th individual immunity at F5, sending Kristina to the jury
• Sophi won the final immunity challenge and chose to take Sage to the Final 3
• Savannah beat Rizo in the fire-making challenge to earn the last Final 3 spot
• The season featured the first all-female Final 3 since San Juan del Sur (S29)

Model Accuracy Breakdown

• ✅ Winner prediction: CORRECT — Savannah was the model's #1 pick at 35.89%
• ✅ 5th place: CORRECT — Kristina correctly predicted as the weakest finalist
• ✅ Top 3 contestants: 2 of 3 correct — Savannah and Sage were in both predicted and actual top 3
• ❌ Rizo vs. Sophi swap — Model ranked Rizo 2nd and Sophi 4th; actual placements were reversed (Sophi 2nd, Rizo 4th)

The model correctly predicted Savannah as the Sole Survivor of Season 49, extending its winner accuracy to 93.9% (45/49 seasons).

🧠 Technical Overview

Model Architecture

• 7-Seed Ensemble for maximum stability and robustness
• 28 Total Models: 4 models per seed (2 Neural Networks + 2 XGBoost classifiers)
• Dual Training Strategy:
  • Model 1 (40% weight): Winner-focused training with emphasis on winner vs. non-winner distinctions
  • Model 2 (60% weight): Full ranking training capturing all placement relationships

Neural Network Specifications

• Architecture: 512-256-256-128 hidden layers
• Activation: ReLU
• Solver: Adam optimizer
• Iterations: 3,000 per model
• Regularization: L2 regularization (alpha=0.0001)

XGBoost Specifications

• Trees: 8,000 estimators per model
• Depth: 12 maximum depth
• Learning Rate: 0.005
• Regularization: L1=0.2, L2=0.2
• Sampling: 90% subsample, 90% column sampling

📊 Feature Engineering

The model uses 8 carefully engineered features derived from contestant data:

Feature	Description	Source
conf_zscore	Confessional count z-score within season	Confessional data
conf_percentile	Confessional percentile rank	Confessional data
tribal_sigmoid	Tribal council participation score	Tribal attendance
vote_sigmoid	Vote safety/threat assessment	Vote history
survival_rate	Votes received vs. tribals attended ratio	Vote/tribal data
winner_profile_1	Composite winner signal	Multi-feature blend
jury_appeal	Estimated jury likability	Derived metric
consistent_edit	Stable confessional trajectory	Temporal analysis

📁 Project Structure

survivor49_git/
├── predict_winner.py          # Main model implementation
├── submission.csv            # Season 49 predictions output
├── requirements.txt          # Python dependencies
├── README.md                # This file
├── CURRENT_MODEL_DETAILS.md  # Detailed technical specifications
└── .gitignore               # Git ignore rules

🚀 Quick Start

Prerequisites

• Python 3.8 or higher
• pip package manager

Installation & Usage

# Clone the repository
git clone <repository-url>
cd survivor49_git

# Create virtual environment
python3 -m venv .venv
source .venv/bin/activate  # On Windows: .venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Run the model
python3 predict_winner.py

Expected Output

The script will:

1. Load historical Survivor data (S1-S48)
2. Engineer features and train the ensemble
3. Validate on historical data
4. Generate Season 49 predictions
5. Save results to submission.csv

🔬 Methodology

Data Sources

Utilizes the survivorpy package for:

• Contestant information and placement data
• Confessional count tracking
• Tribal council attendance
• Vote history and patterns

Training Strategy

1. Pairwise Comparison Training: Models learn to compare contestants directly
2. Temporal Weighting: Recent seasons weighted more heavily
3. Importance Weighting: Winner predictions weighted more than placement predictions
4. Ensemble Averaging: 7 different random seeds ensure stability

Validation Approach

• Leave-One-Season-Out: Each season predicted using models trained on all other seasons
• Comprehensive Metrics: Tracks accuracy across top-1 through top-5 predictions
• Conservative Evaluation: Only exact matches counted as correct

📈 Model Performance Analysis

The model demonstrates exceptional performance with:

• Consistency: 91.7% winner accuracy maintained across diverse seasons
• Robustness: Strong performance on both old-school and modern seasons
• Reliability: Ensemble approach minimizes variance in predictions

🔧 Dependencies

• pandas>=2.0.0 - Data manipulation and analysis
• numpy>=1.24.0 - Numerical computing
• scikit-learn>=1.3.0 - Machine learning algorithms
• xgboost>=2.0.0 - Gradient boosting framework
• survivorpy>=0.1.0 - Survivor data access

📄 License

This project is licensed under the MIT License.

---

Note: This model is for entertainment and analytical purposes. Survivor outcomes depend on many factors beyond statistical modeling, including social dynamics, strategic gameplay, and unpredictable events.