Machine learning-based cost predictive model for better operating expenditure estimations of US light rail transit projects

Inaccurate forecasts of operating expenditures (OPEX) during the planning phase for new Light Rail Transit (LRT) projects in the United States underestimated future costs by up to 45% (Harmatuck, 2008). When OPEX exceeded projected levels, local transit agencies often reduced public transit services to operate within their respective annual budgets. Therefore, it is imperative for transit agencies to produce reasonably accurate planning estimates to secure sufficient funding to support future operations, maintenance, and service delivery associated with LRT systems. The research aimed to develop a more accurate LRT OPEX predictive model to be used during the planning stage. Traditional statistical analysis and various machine learning-based algorithms were utilized with input from 22 LRT systems in the United States spanning between 2008 and 2018 from various U.S. governmental public databases. This study extended the current state of practice that relied pri-marily on sum of unit-cost estimates (also known as the unit-cost method) which generally failed to produce accurate forecasts due to lack of engineering details at the planning stage. Existing research attempted to develop regression-based methodologies using system-based attributes but did not substantially increase prediction ac-curacy from using the unit-cost method. The research improved current practices and research by having de-veloped a more accurate and replicable machine learning-based predictive model using available geographic, socio-economic and LRT system-related variables at the planning stage.

Automated summary

Inaccurate forecasts of operating expenditures during the planning phase for new LRT projects in the United States underestimated future costs. It is important for transit agencies to produce accurate planning estimates to secure funding for future operations. This study developed a more accurate predictive model using traditional statistical analysis and machine learning algorithms.