Time Series Visualization and Forecasting from Australian Building and Construction Statistics

Abstract

The Australian Bureau of Statistics (ABS) regularly releases statistical information, for the whole of Australia, for public access. Building-and construction-related statistics are important to reflect the status of this pillar industry of Australia and help researchers, practitioners, and investors with decision-making. Due to complex retrieval hierarchy of ABS’s website and irregular update frequency, it is usually time-consuming to find relevant information. Moreover, browsing the raw data from ABS’s webpages could not provide the insights to the future. In this work, we applied techniques from computer science to help users in the building and construction domain to better explore the ABS statistics and forecast the future trends. Specifically, we built an integrated Web application that could help collect, sort, and visualize the ABS statistics in a user-friendly and customized way. Our Web application is publicly accessible. We further injected our insights into the Web application, based on the existing data by providing online forecasting on user’s interested information. To achieve this, we identified a series of related economic factors as features and adjusted a multi-variant, LSTM-based time series forecasting model by considering the most informative factors. We also compared our approach with the most widely used SARIMA-based forecasting model to show the effectiveness of the deep learning-based models. The forecast values are depicted at the end of the time series plots, selected by the users.