Predicting the Volatility of Highway Construction Cost Index Using Long Short-Term Memory

Abstract

The highway construction cost index (HCCI) is a composite indicator that reflects the price trend of the highway construction industry. Most available indexes exhibit significant variation, creating a challenge for state agencies to make accurate budget estimations. Numerous researchers have attempted to forecast the index using quantitative models, but two major problems still exist. First, few models work effectively with highly volatile data. A model that is only fitted well with stable data does not validate its forecasting power. Second, a good prediction model should be able to forecast at different time horizons. Many prior research projects only predicted one index point ahead, limiting the application effectivity in practice. This research fills the gap by applying the long short-term memory (LSTM) units built in the encoder and decoder architecture to model and predict the variation of the HCCI. An illustrative example used the Texas HCCI as the raw data, and the results were compared to the seasonal autoregressive integrated moving average model. The results show that the developed LSTM model outperformed the time series models in terms of providing more accurate prediction in all three forecasting scenarios, short-term, medium-term, and long-term prediction. The main contributions of this study to the body of knowledge in cost engineering and forecasting are summarized in the following two areas: first, the paper presents a novel application of an artificial intelligence algorithm for cost index forecasting that provides more accurate prediction than the prevailing time series models, particularly for highly volatile cost indexes. Future researchers could be benefited from the explored results in this paper and use them as a referenced benchmark. Second, this is one of the first papers in construction management that shows the performance of the forecasting models when shape-change of index exists.