This study investigates apartment price formation through a comparative assessment of econometric and Machine Learning (ML) models applied to a micro-level dataset of residential properties in Tirana, Albania. The research addresses the challenge of modelling housing prices in emerging real estate markets characterized by heterogeneous property attributes and spatial variation. To ensure comparability across dwellings of different sizes, the dependent variable is defined as price per square meter. The methodological framework combines a hedonic Ordinary Least Squares (OLS) model and a nonlinear Random Forest (RF) model, allowing the evaluation of both model interpretability and predictive performance. Elastic Net and Extreme Gradient Boosting (XGBoost) models are additionally employed as robustness benchmarks. Model performance is assessed using standard prediction accuracy measures, while variable effects and importance metrics are analysed to identify the main determinants of housing prices. The results reveal that location-related factors and structural housing characteristics constitute the dominant drivers of apartment values. Apartment size is negatively associated with price per square meter, whereas the number of bathrooms is positively and statistically significantly associated. The number of rooms becomes insignificant after controlling for other explanatory variables. Strong neighbourhood effects confirm substantial spatial heterogeneity in the housing market. The comparative analysis demonstrates that RF achieves superior predictive accuracy relative to the alternative models, highlighting the ability of nonlinear methods to capture complex relationships in housing price data. The findings contribute to the application of statistical and ML techniques in real estate valuation and provide evidence on the relative strengths of interpretable and predictive modelling approaches.