Modelling Residential Property Price Indices Within Kampala City Using Convolutional Neural Networks and Fuzzy Inference Systems

Abstract

A population increase of approximately 2.4 billion people is expected in the world’s urban areas by the year 2050 instantiating a need for property metrics that can be used as a gauge for urban vitality and to better understand the property trade. The housing price index is one of those metrics. It is a tool that investors use to interprete the property price phenomenon across borders and financial institutions to assess risk exposure. The researcher specifically aimed to compile all possible variables/factors that affect residential property price index from a comprehensive review of literature and ultimately develop a robust temporal convolution network (TCN) algorithm and a trained adaptive-neuro fuzzy inference system (ANFIS) both of which would be in position to predict house price indices. Given the proposed models were calibrated based on headline housing price indices, the scope of the study was a high-level assessment of the five divisions of Kampala namely; Rubaga, central, Kawempe, Nakawa, and Makindye. The TCN and ANFIS models show training RMSE values of 0.1086 and 1.835 x 10-6 respectively. Concurrently with the aforementioned results, the models recorded training R2 values of 0.92 for the TCN and 1.0000 for the ANFIS. The test RMSE results of modelling the TCN and ANFIS were 0.54 and 0.6736 respectively. The network models had satisfactory results in forecasting price indices based on the following 7 input variables: producer price index (PPI), normal effective exchange rate index (NEER), Brent oil price, S&P 500 index, CBOE’s volatility index, nominal broad US dollar index and the volume of residential mortgage loans. Overall, the fuzzy model outperformed the convolution algorithm.