End-to-End ML for German Real Estate Prediction — live demo

This project started as my first ML notebook — a research artifact comparing 10 different models for two prediction tasks on the German Apartment Rental Offers Kaggle dataset (268k rows). The notebook used MLflow to track experiments and identified Gradient Boosting as the best elevator-presence classifier and XGBoost as the best base-rent regressor.

I then refactored the notebook into a production-shaped commercial version: two serializable sklearn.pipeline.Pipeline objects, a FastAPI backend, this static HTML form, and a 100-record holdout dataset preserved for live predicted-vs-actual demos (use the dropdown below).

After a code review I iterated to v2: added geographic features (state + city, the latter via TargetEncoder to dodge dimensionality blowup), removed PCA from the regression pipeline (it was destroying signal in tree models), log-transformed the rent target via TransformedTargetRegressor, and applied class-balanced sample weights to the elevator classifier.

Try the geographic intuition: in the BaseRent form below, set an apartment with realistic features and change only the city. Munich predicts ~€1,286/month for an 80 m² apartment; the same apartment in Magdeburg predicts ~€521. The model learned this entirely from data.

Tech stack: scikit-learn 1.6 · XGBoost 3 · FastAPI · pydantic v2 · joblib · vanilla HTML/JS (no framework). Full documentation, training scripts, and the original research notebook are in the GitHub repository — see README.md, PROJECT_EVOLUTION.md, and MODEL_IMPROVEMENTS.md.