Evaluating the performance of combining neural networks and genetic algorithms to forecast construction demand: the case of the Singapore residential sector

Abstract

In recent years, forecasting demand for residential construction in Singapore has become more vital, since it is widely perceived that the next trough of the real estate cycle is approaching. This paper evaluates the use of a combination of neural networks (NNs) and genetic algorithms (GAs) to forecast residential construction demand in Singapore. Successful applications of NNs, especially in solving complex non-linear problems, have since stimulated interest in exploring the capabilities of other biological-based methods such as GAs, and in exploiting the synergy of these two techniques to create more problem-solving power. In the study, a basic NN model is used as a benchmark to gauge the performance of the combined NN-GA model. A relative measure of forecasting accuracy, known as the mean absolute percentage error (MAPE), is used for the comparison. The models are checked also for internal validity by allowing each to be trained twice and having a set of forecasts generated after each training. Both models are found to produce accurate forecasts, because their MAPE values consistently fall within the acceptable limit of 10%. However, the combined model out-performs the basis model remarkably by reducing the average MAPE from about 6% to a mere 1%. For each model, the marginal difference in the MAPE values (i.e., 0.5% for the NN model and 0.06% for the NN-GA model) of its two forecasts indicates consistency in performance, hence establishing internal validity as well. The findings reinforce the reliability of using NNs to model construction demand and reveal the benefit of combining NNs and GAs to produce more accurate models.