Multidimensional house price prediction with SOTA RNNs

Abstract

This paper introduces insights into the Turkish real estate market, which can be generalized globally. It primarily aims to find the best forecasting algorithms for the housing price index and compare their prediction performance over three, six, nine, and twelve months ahead by using recurrent neural networks (RNN) with a comparison of out-of-sample predicting power of econometrical models. For these purposes, we employ three RNN architectures in twenty-four settings, revealing that certain RNN architectures are the best predictors in forecasting the Turkish real housing price index. The RNN architectures outperform traditional econometric models; however, the more months forecasted, the lower the prediction power. The lagged values of the price-to-rent ratio, real rents, and the lagged USDTRY values contribute more than the other predictors in forecasting the real housing price index. The outcomes suggest that stocks, real estate investment trusts, and gold are neither complementary nor competing financial instruments since housing is an illiquid asset.

Keywords

• housing price index prediction
• recurrent neural networks
• deep learning