Selection of optimal window size and ensemble methods for rental price estimation with XGBoost: Evidence from Buenos Aires

Abstract

In the context of forecasting rental prices of residential properties, the fact that the relationship between rents and property characteristics may be subject to significant changes over time can hinder the predictive performance of models that do not adjust to these changes. Previous studies have tested the accuracy of different statistical models in predicting rent prices, but none has focused on how the definition of the estimation window of observations used in the training stage can affect predictive performance. This paper explores the impact of selecting different window sizes in a forecasting experiment of rent prices in the City of Buenos Aires from 2020 to 2022 using the machine learning algorithm XGBoost. The results obtained for out-of-sample one-month-ahead forecasts indicate that the decision to use an expanding window or rolling windows is not trivial in terms of the achieved predictive performance. Among a set of rolling window sizes, the model with a size of 6 months yields the lowest error in terms of root mean square error and mean absolute error. Additionally, three methods that dynamically combine the predictions of models with different window sizes are tested. An ensemble method that outputs a weighted average of the predictions using time-varying weights based on the inverse of previous forecasting errors emerges as the superior strategy, producing a mean absolute percentage error of 19.7%. Overall, these results underscore the need to take into account the temporal dimension when selecting observations for training if the goal is to maximize out-of-sample predictive performance.