Predicting Real Estate Funds: A Comparative Study of Machine Learning and Time Series Methods

This work investigates different strategies for predicting the price variation of Real Estate Investment Funds (FIIs) using machine learning models, in comparison with a traditional time series method. We analyze the performance of the models in terms of fund categories (i.e., paper, brick, or hybrid), model settings (i.e., one model per fund or a general model), and forecast time window (i.e., 6 months or 1 month). An analysis was also carried out considering an enrichment of the data with characteristics of the properties belonging to the funds, which is a pioneering contribution of this work. The results reveal, among other conclusions, that machine learning models outperform the time series technique only for the medium term, and that the information on the properties belonging to the funds was important for improving forecasts.