Evolution Enhancing Property Price Prediction by Generating Artificial Transaction Data

Property price prediction is not only of great importance for home-buyers but also vital for economic studies. The common approach in price prediction is analysing historical market prices and relevant economic data by a predictive model, such as ARIMA model, linear regression, and artificial neural networks. Regardless of the model, a successful prediction using this approach relies on an abundant amount of data, e.g. more than one million property transactions over the last few decades. However, such a large data set might not be readily available, and also expensive to obtain. Hence we propose evolutionary generative Generative Adversarial Networks (GAN), to learn the distribution of these transaction data and generate samples to enrich the data set. This generated data set can help achieve better prediction. In our experiments, we compare not only the performance of several GAN varieties including evolutionary GAN (E-GAN) and knowledge distillation evolutionary GAN (KDE-GAN), but also the differences in prediction results before and after using GAN enriched data set. Our study shows that evolutionary GAN can improve prediction performance without requiring a large amount of data. In addition, evolution does provide extra benefits to the GAN process and enhances its value not only in methodological advancement but also in commercial potential.