INTEGRATED NEURAL NETWORK MODELLING OF CONSTRUCTION DEMAND IN SINGAPORE

Abstract

In every country, there are always fluctuations in the total workload of the construction industry, as well as that of the individual firms. However, the changes in construction activity have been observed to be more pronounced in their extent and effect, than those in the levels of production in most of the other sectors of the economy. As the construction output has fulfilled the demands of the rapidly growing economy, Singapore's construction industry has also been transformed. The econometric models developed for, and applied in, Singapore for estimating construction demand have attempted to explain the construction and real estate cycles. To date, none of the methods has offered a straightforward, easy to apply approach to the forecasting of the total construction demand. The critical weaknesses of the existing models are the need to re-validate the entire model on every occasion when there are change(s) in the economic variables under investigation and that they capture a stationary representation of the relationship between the independent variables and the dependent variable which cannot easily explain the non-linearity and the dynamism of economic behaviour. The present study has the focus of developing a suitable construction demand model using neural networks that would help to overcome the problems associated with the traditional econometric models using conventional algorithms. The first stage of the III study involved a review of works on econometric models and identification of the indicators that are often used in the prediction of real estate and construction demand levels and patterns. In the second stage, past quarterly data of economic variables were abstracted from national statistics published by a number of public authorities and obtainable from TRENDS Electronic Database, Department of Statistics, Singapore for modeling the prediction of construction demand. The third stage comprised the training and testing of the neural network models with the selected indicators. The quarterly data between 1981 and 1996 were used for training and developing the model, while "ex post" forecasts were being made over a historical period between the fourth quarter, 1994 and third quarter, 1995. The last stage of the study involved an attempt to forecast the future volume of construction demand, that is, the Value of Contracts Awarded, by changing the connection attributable to each input variable. The neural network solutions placed emphasis on three explanatory variables, Gross Fixed Capital Formation, Building Cost and Manufacturing Output to explain the movement of construction demand in Singapore. The Reinforcement Neural Network with Directed Random Search Engine appears to be the best trainable model and to predict the total construction demand with a greater degree of accuracy via error term estimations as compared to the Fast BackPropagation Neural Network and Modular Neural Network. The forecast result is simulated under dynamic conditions which is more realistic in an uncertain business environment. The potential users can be government agencies and large construction industry enterprises who can apply this technique to predict the future level of total construction demand thereby allowing sufficient lead time to enable organisations involved in the construction industry to improve their decision-making on possible (re)positioning of their entities in the local construction industry or, where relevant, to diversify into regional market(s) or to expand their business operations.