Housing price prediction using machine learning

Abstract

Predicting the housing price is an enduring topic since the price change of real estate has a great relationship with the economy, policy, and market. This dissertation explored the use of deep learning models to predict the resale prices of the Housing and Development Board (HDB) flats. In this dissertation, a comprehensive study of the HDB flat transaction data in Singapore has been conducted from 3 aspects: web crawling and analysis, resale price prediction, and performance comparison. Prediction methods were divided into two-phase and single-phase. For the two-phase method, the median resale price per square meter (MRP/m^2) in one month was initially predicted by the Long Short-Term Memory (LSTM) model in the first phase, based on the data from the previous 24 months. Then the second phase models, including LSTM, Multilayer Perceptrons (MLP), and Convolutional Neural Network were proposed to predict the resale prices of HDB flats. The first and the second phase were connected by inputting the MRP/m^2, along with the intrinsic and external attributes of flats, to the second phase models. On the other hand, to judge the effect of the single-phase method, only the intrinsic and external attributes of flats were fed into the second phase models. Grid search with cross-validation was applied to these models. Then, the models with the optimal combination of hyper-parameters were evaluated and compared the performance on the test set. The experiment demonstrated that the two-phase methods outperformed the single-phase ones, where the collaboration of the LSTM and the MLP model achieved the minimum error and the highest accuracy.