Choosing the metric
Bankruptcies are the rare class. Missing one may cost more than a false alarm. I also used F3, which gives recall more weight than precision, to examine that trade-off.
Where it started
A university project starting with a question. What useful information about bankruptcy risk can I draw from company data?
Problems and decisions
Comparing strategies
I compared models and thresholds to see how detected cases and false alarms change, and studied feature importance with Random Forest.
How it works
- A Python notebook prepares and standardizes tabular data with pandas, NumPy and scikit-learn tools.
- Random Forest supports feature-importance analysis. Experiments compare RF configurations with Support Vector Machines.
- Neural-network experiments use TensorFlow/Keras to examine architecture, regularization and signs of overfitting.
- Decision thresholds and feature sets are compared through F3-score, ROC curves and charts saved in the notebook.
Inside the project
From data to interpretation: the stages of the project.
- Data and preparation
- Features and standardization
- Random Forest · SVM · neural networks
- Thresholds · F₃ · ROC · overfitting
What I learned
This project taught me to read a score alongside the errors behind it. Changing the metric or threshold changes which cases are detected. I want that trade-off to be clear whenever I compare models.
Current status
Exploratory university project. The charts document runs saved in the notebook. Testing the conclusions would require an independent evaluation on new data.
An interesting problem?
Tell me what you want to build or improve, who will use it and what constraints you have. I’ll get back to you so we can work out where to start.