Fraud Prediction Model
A small web app that scores a credit-card transaction for fraud, using a logistic regression model trained on a public Kaggle dataset.
Overview
The model is trained on the Kaggle credit-card fraud dataset, where fraudulent transactions are a tiny fraction of the total. A FastAPI app wraps it with a simple page for checking a transaction.
Worth noting
Built for imbalance
Class weights are balanced and the split is stratified, so the rare fraud cases are represented in both training and test data.
The threshold is a choice
The training script reports precision and recall at several thresholds instead of accepting the default of 0.5. The chosen value lives in configuration.
Same preprocessing in training and serving
The fitted scaler is saved beside the model and reused by the app, so a transaction is scaled the same way it was during training.
Limits
Logistic regression is a baseline. It was chosen to learn the problem end to end, not to compete with production fraud systems.