Final Machine Learning project — ESILV, 4th year (2025–2026)

Compare classical and machine-learning-based portfolio allocation strategies on historical S&P 500 data and evaluate their risk-adjusted performance.

This project was developed as part of the Machine Learning course at ESILV (École Supérieure d'Ingénieurs Léonard de Vinci). The goal is to implement and benchmark several portfolio allocation strategies — from naive baselines to supervised machine learning models — on a selection of S&P 500 stocks over the 2013–2018 period.

• Equal-Weight Baseline — naive 1/N allocation across all assets, buy-and-hold
• Markowitz Minimum Variance Portfolio — classical closed-form mean-variance optimization
• Random Forest Strategy — per-ticker binary classifiers with probability-weighted allocation and GridSearchCV tuning
• Logistic Regression Strategy — similar supervised ML approach with L2 regularization
• Technical Feature Engineering — moving averages (20d, 60d) and rolling volatility (20d) per ticker
• ANOVA Feature Selection — top-k features selected per model
• Comprehensive Evaluation — total return, annualized Sharpe ratio, max drawdown, Calmar ratio

# Clone the repository
git clone https://github.com/MaximeFARRE/Projet_final_ML.git
cd Projet_final_ML

# Create a virtual environment
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Note: A Kaggle API key is required to download the dataset automatically. Set up ~/.kaggle/kaggle.json before running. See the Kaggle API docs.

Run the full pipeline with a single command:

python run_all.py

Or execute each step individually:

python scripts/run_prepare.py              # Download data & generate EDA plots
python scripts/run_baselines.py            # Evaluate Equal-Weight and Markowitz
python scripts/run_random_forest.py        # Train and backtest Random Forest
python scripts/run_logistic_regression.py  # Train and backtest Logistic Regression

All outputs (metrics CSV files and figures) are saved under reports/.

Performance metrics on the test set (approx. 2016–2018):

Generated figures are available in reports/figures/.

Equity Curves	Correlation Heatmap

Normalized Prices	Returns Distribution

Projet_final_ML/
├── run_all.py                         # Main pipeline entry point
├── requirements.txt
├── data/
│   ├── raw/                           # Raw downloaded price data
│   └── processed/                     # Processed feature matrix (CSV)
├── models/                            # Saved model checkpoints
├── reports/
│   ├── figures/                       # Generated plots (PNG)
│   └── tables/                        # Performance metrics (CSV)
├── scripts/                           # Runnable pipeline steps
│   ├── run_prepare.py
│   ├── run_baselines.py
│   ├── run_random_forest.py
│   └── run_logistic_regression.py
└── src/                               # Core library
    ├── config.py                      # Tickers, time range, paths
    ├── data/                          # Loading & preprocessing
    ├── features/                      # Technical indicators & ANOVA selection
    ├── baselines/                     # Equal-weight & Markowitz
    ├── models/                        # Random Forest & Logistic Regression
    └── evaluation/                    # Performance metrics

• Only 7 S&P 500 tickers (AAPL, MSFT, AMZN, GOOGL, FB, T, GS) over a single market regime
• Transaction costs and slippage are not modeled
• ML models are retrained from scratch on each pipeline run
• The PPO reinforcement learning component (models/ppo_portfolio.zip) is experimental and not yet integrated into the main pipeline
• No live trading or real-time data feed

This project is licensed under the MIT License — see LICENSE for details.