MSc thesis · Data Science & Society · Tilburg University · Jan 2024

Comparative Analysis of ML Algorithms for Time on Market (TOM)

Prior academic work on how long Dutch homes sit on the market before selling leaned almost entirely on Ordinary Least Squares (OLS). Using the Funda-Sold dataset of Dutch house sales from 2020–2023, I benchmarked OLS against Random Forest, XGBoost, LightGBM and CatBoost — combined with Recursive Feature Elimination with Cross-Validation (RFECV-DT) for feature selection and randomized-search hyperparameter tuning.

Dataset: Funda-Sold dataset — Dutch house sales from 2020 to 2023, sourced from Funda.nl, the leading real-estate platform in the Netherlands.

Headline results

R² (CatBoost)0.632
MAE11.36 days
RMSE26.18 days
Lift over OLS+45%

Method

From Funda.nl listings to a predictive TOM model.

  1. 01

    Sourced the Funda-Sold dataset (2020–2023) of Dutch residential listings and cleaned it in Python.

  2. 02

    Engineered features and reduced dimensionality with RFECV-DT, converging on 18 optimal variables.

  3. 03

    Trained OLS as a baseline plus Random Forest, XGBoost, CatBoost and LightGBM with randomized-search tuning.

  4. 04

    Compared performance on full and RFECV-reduced datasets using R², MAE and RMSE on a held-out test set.

Model comparison — test set (full dataset)

Sort
  • CatBoostbest
    0.632
    MAE
    11.36
    RMSE
    26.18
  • Random Forest
    0.609
    MAE
    11.67
    RMSE
    26.97
  • LightGBM
    0.609
    MAE
    11.84
    RMSE
    26.95
  • XGBoost
    0.587
    MAE
    12.19
    RMSE
    27.71
  • OLS (baseline)
    0.435
    MAE
    14.96
    RMSE
    32.41

Source: Table 5, test-set performance on the full Funda-Sold dataset. Tap a column header (or a sort chip on mobile) to reorder.

What actually drives Time on Market

The variables that mattered — across every model.

  • 01Construction status (existing vs. new build)
  • 02Roof type
  • 03Volume (m³)
  • 04Plot area (m²)
  • 05Building-related outdoor area (m²)
  • 06Asking price per m² & last asking price
  • 07Living space (m²)
  • 08Type of residence

Takeaway

CatBoost outperformed OLS on every metric and remained the strongest model on both the full and RFECV-reduced datasets. Crucially, the top drivers — construction status, roof type, volume, plot area, outdoor area — stayed stable across every feature-selection method, giving the finding methodological weight beyond the raw metrics.