Skip to content

Changelog

All notable changes to this project are documented here. The format follows Keep a Changelog, and the project uses Semantic Versioning.

Unreleased

[0.1.1] - 2026-11-01

Fixed

  • test_notebook_parity.py asserted the regression figures with an absolute tolerance of one cent, which only held on Windows with Python 3.11-3.13. The reference values are unchanged; they are now asserted with a 6% relative tolerance, covering the 4.3% spread observed across the CI matrix. See docs/DECISIONS.md D29.

  • test_cli.py::test_plots_command_wiring failed wherever the temporary directory was long enough for Rich to wrap the filename mid-word. The test now widens the console for its own duration.

  • CI no longer fails on ubuntu-latest with Python 3.9: Ubuntu 26.04 runners have no Python 3.9 build, so that one combination is skipped. Python 3.9 is still tested on macOS and Windows.

Changed

  • Releases are distributed through GitHub: the release workflow now runs when a GitHub release is published and attaches the wheel and source distribution to it. Install with pip install "amesml @ git+https://github.com/Ahsan-muhd-444/house-price-prediction@v0.1.1". Publishing to PyPI is deferred.
  • Documentation no longer claims the defaults reproduce the notebook "exactly" or "to the cent" without qualification.

0.1.0 - 2026-11-01

First packaged release of the house-price-prediction coursework as amesml.

Added

  • amesml Python package with a typed public API: load_ames, build_preprocessor, train_regressor, train_classifier, train, evaluate_regressor, evaluate_classifier, predict, save_model and load_model.
  • Defaults that reproduce the original notebook's published results closely (linear regression MAE $21,139.99, R² 0.868; logistic regression accuracy 94.5%, F1 93.9%). The classification figures reproduce exactly; the regression figures reproduce to the cent on Windows with Python 3.11-3.13 and to within about 5% elsewhere, because they come from an iterative solver on an ill-conditioned matrix. See docs/DECISIONS.md D1.
  • Ridge, lasso and histogram gradient-boosting models for both regression and above/below-median classification, with cross-validated default regularisation.
  • Optional engineered features: total area, house and remodel age, total bathrooms and porch area.
  • TrainConfig, loadable from YAML or JSON, with validation of every option.
  • Dataset loading with schema validation and a checksummed, atomic download pinned to an immutable URL.
  • Model bundles with metadata.json: configuration, metrics, feature defaults, training-data fingerprint, library versions and a checksum of the model file.
  • amesml command-line interface: data download, train, evaluate (including --split test to reproduce held-out metrics), predict, plots and serve.
  • FastAPI service with /predict, /predict/class, /health and /model/info, partial-record support and reporting of unseen categories.
  • Multi-stage Docker image and Docker Compose file.
  • Figures drawn with matplotlib: price distribution, missing values, residuals, confusion matrix and permutation importance.
  • Test suite with 99% coverage, run on Linux, macOS and Windows for Python 3.9-3.13.
  • CI, release (PyPI trusted publishing), documentation, CodeQL, labeler and stale-issue workflows, plus Dependabot.
  • MkDocs Material documentation: user guide, model card, benchmarks, architecture, design decisions and API reference.
  • Contributing guide, Code of Conduct, security policy, governance, support page, citation metadata, and issue and pull request templates.

Changed

  • Repository reorganised: the report moved to docs/, the notebook's pickled models to models/notebook/, and the notebook is committed with its outputs stripped (the original is kept in notebooks/archive/).
  • The author name and the licence's copyright holder corrected to Muhammad Ahsan.