Assessment of machine learning methods for mass real estate appraisal

In this thesis, the use of machine learning (ML) approaches for the purpose of real estate mass appraisal was investigated using five different methods in a large area considering the efficiency, accuracy, and transparency. The study area was located in the Mamak District of Ankara, the capital city of Turkey. The data used in the thesis were inspected and analysed thoroughly, and thus exhibit high quality and reliability. The applicability of the ML methods in the context of mass appraisal is discussed in terms of the accuracy, reliability, interpretability, and the generalization capability. The results were also compared with the conventional appraisal methods. The results obtained here have shown that the ML-based methods can appraise many real estates together at once and rapidly; and thus, they can be preferred over the conventional valuation methods. Among the methods compared here, the Random Forest (RF) provided the highest prediction performance followed by the Artificial Neural Networks (ANN), Support Vector Machines (SVM), Multiple Regression Analysis (MRA), and Adaptive Neuro-Fuzzy Inference System (ANFIS). The stepwise MRA method, which is a transparent and interpretable linear ML method, was preferred as the conventional approach. Another important outcome was that although the models built with the non-linear ML methods yielded high accuracies, their interpretability was lower and thus usability for the valuation purposes may be questionable. In this thesis, the employed methods are explained and investigated in more detail with the aim of contributing to the mass appraisal context. In addition, recommendations on the real estate valuation system were derived based on the study outcomes together with possible contributions of the methods presented in the field of mass valuation studies to Turkey, which has not yet been institutionalized in the field of real estate valuation.