The Student Forecaster Model
As part of the Le Wagon Bootcamp, this was my final group project:
🧠 Proven Hypothesis
The academic performance of a student depends not only on their academic capabilities but also their socio-economic status.
🏫 Imagine this…
You’re a headteacher managing a school of thousands of students. How do you create an ideal environment where every student can thrive?
🎯 Project Goal
To build a machine learning model that forecasts academic outcomes using a dataset of Portuguese students.
The objective was to uncover socio-economic drivers influencing final grades (G3) and offer a predictive dashboard for education leaders.
🔗 Try the app: The Student Forecaster Model
🔍 Data Exploration
We started with a raw dataset and trimmed irrelevant or misleading variables like:
- Home address
- Parent jobs (too many “other” values)
- Nursery attendance
- G1/G2 (used G3 only as target)
We then grouped related features:
- Alcohol intake (weekday/weekend)
- Family dynamics
- Time management
- Educational support
- Reason for school choice
- Parental education
Tools used:
- Heatmaps
- Boxplots
- Group binning
- Value counts
- Correlation analysis
⚙️ Model Training
After feature engineering and encoding:
- Split data into train/test sets
- Used Gradient Boosting Classifier
- Preprocessing: scaling + one-hot encoding
We tested multiple models for comparison:
| Model | Precision | Test vs Train | Overfitting |
|---|---|---|---|
| Logistic Regression | 0.75 | 0.72 vs 0.72 | ❌ |
| KNN | 0.76 | 0.67 vs 0.83 | ✅ |
| Random Forest | 0.81 | 0.70 vs 0.98 | ✅ |
| XGBoost | 0.82 | 0.68 vs 0.95 | ✅ |
| Gradient Boosting | 0.76 | 0.70 vs 0.78 | ✅ Slight |
🧪 Final Metrics (on test data)
- 🎯 Accuracy:
0.96 - 📊 Precision:
0.95 - 🔍 Recall:
0.99 - 🧮 F1 Score:
0.97
These metrics show strong performance on the binary classification task: predicting whether a student would pass or fail based on inputs.
🎓 Insights
- Students who had failed before are more likely to receive support and improve.
- Mother’s education had stronger correlation than father’s.
- Study time and school choice were significant drivers.
- Most students want higher education, causing class imbalance.
🧠 Tools Used
- Python:
pandas,seaborn,scikit-learn,joblib - Web app: Streamlit
- Deployment: Streamlit Cloud
🌐 App Link
👉 student-forecaster.streamlit.app
Additional Resources 📎
For a detailed explanation:
- 📊 Data Analysis & Insights — View the presentation slides
- 🧠 Model Training & Evaluation — Explore the Jupyter Notebook
- 💻 Codebase & App — Visit the GitHub repository
💬 Final Thoughts
This project brought together all the fundamentals of data science — from data wrangling and EDA to machine learning and user deployment — to solve a real-world education problem.
Let’s build data tools that truly support educators and learners alike. 🌍📚