📈 Machine Learning Project — Predicting CAC 40 Daily Direction

🎯 1. Project Overview

The goal of this project is to predict the next-day direction of the CAC 40 index (Up / Down ) using historical data and engineered technical indicators.

This is a supervised classification problem using tabular time-series data.

Main question:
👉 “Given today's market conditions, can we predict whether tomorrow's CAC 40 closing price will go Up or Down?”

The project follows the complete ML workflow:

• Data acquisition
• Preprocessing & feature engineering
• Model training & tuning
• Evaluation with relevant metrics
• Interpretation of results

---

📂 2. Dataset

Source: Downloaded with yfinance API

Ticker: ^FCHI Interval: Daily Period: Max

Core variables:

• Date
• Open
• High
• Low
• Close
• Volume

Engineered features:

• Daily returns
• Rolling means (5, 10, 20 days)
• Volatility
• RSI
• MACD
• Stochastic Oscillator
• Volume trends
• Date features (month, weekday, etc.)

🎯 3. Target Variable

We generate a 2-class classification label based on the next-day return:

• 1 = Up (return > threshold)
• 0 = Down (return < threshold)

Thresholds were optimized to balance the dataset.

---

🧹 4. Preprocessing Pipeline

✔ Missing values

• Checked & forward-filled where needed
• Dropped indicator warm-up rows

✔ Datetime parsing

• Converted Date to datetime
• Added calendar features (month, day of week…)

✔ Feature Engineering

The following indicators were computed:

• Rolling returns
• Rolling volatility
• RSI
• MACD + signal line
• Stochastic
• Volume-based features
• Standard scaling where required

✔ Feature Selection

• Correlation filtering
• Permutation importance

---

🤖 5. Machine Learning Models

We tested multiple algorithms:

Baseline models

• Decision Tree
• Random Forest
• KNN

Advanced models

• LightGBM
• Calibrated Classifier
• Final Stacking Classifier

Tuning

• GridSearchCV
• TimeSeriesSplit for chronological validation
• Optuna for best hyperparameters

---

📈 6. Evaluation

Metrics used

• F1-binary
• Accuracy
• Confusion Matrix
• Classification Report

🚀 7. Model Interpretation with a Streamlit Application (Full Interactive Dashboard)

This project includes a complete Streamlit dashboard that allows you to:

• Fetch real-time market data from Yahoo Finance
• Rebuild the full feature engineering pipeline
• Predict next-day UP / DOWN direction
• Visualize probabilities and historical charts
• Display live TradingView charts inside the app
• Explore SHAP explainability (global + local)

The app uses the same preprocessing functions and the same LightGBM model as the training notebook, ensuring consistency between offline and live predictions.

▶️ How to run the Streamlit app

Make sure you are in the project root folder, then run:

streamlit run streamlit_app/app.py

---

👤 8. Author

Gabriel Enthoven
BSc in Data Science for Responsible Business
EM Lyon & Centrale Lyon

---

🏁 9. Conclusion

This project shows that:

• Technical indicators give predictive signal on market direction
• Stacking improves performance compared to individual models
• Financial time-series remain very noisy and hard to predict
• The pipeline is clean, reproducible, robust, and interpretable

---

🚀 How to Use — User Guide

Welcome to the Financial Direction Predictor!
This guide explains how to install the project, reproduce the analysis, and run the interactive prediction app.

---

🔧 1. Installation — Quick Setup

Clone the repository:

git clone https://github.com/gabentho/ML-Project-Financial-Predictions.git
cd ML-Project-Financial-Predictions

Create and activate the conda environment:

conda env create -f environment.yml
conda activate ml-project-financial-predictions

📄 2. Use the Analysis Notebook

- Open the main notebook Cleaned_notebook.ipynb in JupyterLab or any compatible environment    
- Inside you will find:  
- data cleaning routines  
- feature engineering and technical indicators  
- model training and tuning  
- evaluation of model performance  
- result interpretation  

⚙️ This option lets you reproduce the full analysis, test your own hypotheses or adapt the pipeline if you like data science.

⸻

🌐 3. Run the Interactive Web App (Streamlit)

For fast and user-friendly predictions, you can launch the web dashboard.

▶️ Start the application:

From the project root, run:

streamlit run streamlit_app/app.py

This will open a web interface in your browser.

In the application, you can:

•	Enter any Yahoo Finance ticker (default: ^FCHI)   
•	Select how many years of historical data to load  
•	Get a next-day prediction: UP 📈 or DOWN 📉  
•	View prediction probabilities  
•	Explore historical charts & technical indicators  
•	Use SHAP explainability (global + local)  
•	Inspect the engineered dataset  

⸻

⚠️ Notes & Recommendations

•	An internet connection is required (Yahoo Finance live data)  
•	Ensure your conda environment is activated before launching Streamlit    
•	Predictions are probabilistic — this is an academic project, not financial advice    

Everything runs in real time — no manual data preparation needed.