This study examines realized Days on Market (DOM) for residential listings in Türkiye using a nationwide dataset of more than 650,000 online listings from 2024–2025. The study develops an interpretable machine-learning framework to explain observed listing duration in a heterogeneous housing market and compares a regularized linear baseline with tree-based boosting models, namely XGBoost and Histogram-based Gradient Boosting, using harmonized features related to pricing behavior, physical property characteristics, and multi-level spatial information. Model performance is evaluated using MAE, RMSE, R², and relative error on held-out data. The best-performing specification, based on XGBoost, achieves held-out performance consistent with the inherent difficulty of DOM prediction (R² ≈ 0.39, MAE ≈ 43 days against a mean DOM of 147 days), indicating that substantial unexplained variance remains even after accounting for pricing behavior, property attributes, and spatial heterogeneity. To move beyond aggregate model fit, the study employs gain and permutation importance, SHAP values, and partial dependence plots to interpret model behavior. The results indicate that pricing-related variables and location-based factors play a substantially greater role than most physical housing attributes in explaining realized DOM, and the divergence between gain-based and SHAP-based rankings reveals that price variables drive model fit while spatial variables dominate individual-level predictions. Overall, the findings provide a scalable and interpretable retrospective framework for diagnosing listing-side housing-market liquidity, with implications for pricing strategy, investment benchmarking, and spatial market monitoring in the Turkish housing market.

Keywords

Days on Market, Housing Market, Machine Learning, Interpretable Modeling, Pricing Strategy