Conducting the Best House Price Prediction Model for Turkish Housing Market with Machine Learning
The objective of this paper is to empirically conduct the best machine learning regression model for Turkish Housing Market by comparing accuracy scores and absolute deviations of test results. This paper uses the locational listing data of Turkish leading property portal Zingat.com, for the four-and-a-half-year period from January 2015 to June 2019. In our study, fifteen explanatory building and dwelling variables are used for each prediction model; Location (latitude& longitude), Dwelling Size, Terrace Size, Number of Room, Number of Bathroom, Building Age, Total Floor, Floor Number, Property Type, View, View Direction, Lift, Heating System, Security, and Parking. Data preprocessing (cleaning contradictory data, completing missing data etc.) was done on dataset. Dataset was separated as rental listings and for sale listings. Training the data model; data was apportioned into training and test sets, with an 80-20 split. Three different data models were created by using Support vector machine, Feedforward Neural Networks, Generalized Regression Neural Networks algorithms of forecasting models with supervised learning. Python programming language and Keras library were used. Mean Absolute Percentage Error (MAPE) was selected as the error metric. When the success of the methods was examined according to the results of the analysis, it has been found that the Feedforward Neural Network model gives better results than the other supervising learning models. Smallest MAPE observed as 9.3% for rental listings and 14.9% for on sale listings. The most important variable in the model was building location when we consider the importance of the 15 variables in network development. Dwelling size was the 2nd the most important variable and terrace size had the least importance.
Keywords - Turkey, Housing Market, Machine Learning, Zingat.com, House Price Prediction.