A neural network model to forecast construction output in the united kingdom

A major problem in the construction industry is the high numbers of insolvencies of construction firms. Cyclical fluctuations in construction demand are identified as one of the important causes and, as a result, a number of demand forecasts have been produced to assist firms with their planning processes. However, the forecasts employing traditional approaches have significant drawbacks; qualitative methods are time-consuming and costly to conduct, and forecasts using statistical techniques are often inaccurate.

This study aims to improve the efficiency and reliability of the forecasting process. As there are a large number of factors that relate to construction demand and the knowledge of the interrelationships between them is limited, it is very difficult to precisely model the system using traditional techniques. Implementing a more advanced forecasting technique may be a solution for developing effective forecasts. From the problem characteristics, it is obvious that any new, improved technique must be able to cope with a large number of variables that are highly interrelated to each other.

Based on this assessment of the problem, neural networks were selected for this study, due to their ability to model implicit functions underlying complicated relationships. Provided an adequate set of historical data can be made available, they can learn and map the relations between various variables without prior knowledge and generalise outputs with relatively high accuracy. The technique was thus adopted, aiming to forecast construction outputs one year ahead in three different construction sectors: housing, non-housing, and repair and maintenance. In addition, two input-filtering techniques were adopted to identify the sets of relevant variables for neural networks: regression analysis' stepwise technique and principal component analysis. These techniques reduced the number of input variables significantly, whilst still permitting the neural network models to perform relatively well in terms of accuracy. From the initial total of 43 economic indicators, the reductions were considerable, with 75% reduction in the case of the stepwise procedure and as much as 86% reduction in the other. The models using principal component analysis, however, performed slightly better than those using regression inputs.

When compared to two existing published forecasts performed by experts in the field and using conventional qualitative methods, the results were close in terms of overall accuracy. However, the neural network models outperformed the regression models, which were also produced along side them as the representatives for quantitative methods in the present study.