Conceptual cost estimating (CCE) is a challenging task for highway agencies due to the limited design information available at early stages of project development. As a result, agencies frequently experience large variance from the initial construction estimate to the final cost. Despite the initial estimate’s low level of confidence, it is required for all highway projects as an input to feasibility studies and to establish the project’s budget.

Many authors have explored the use of artificial intelligence and multiple-regression analysis with promising findings to aide CCE. Unfortunately, at this writing, no highway agencies are known to have implemented these data-driven techniques in practice. One of many reasons for this situation is related to a belief that accurate quantities of work are required to produce an accurate estimate. This approach is termed ‘bottom-up’ estimating and is clearly impossible at the initial stage of project development. A second reason relates to the investment necessary to create a reliable database structure that permits high-level statistical analysis. Therefore, this thesis seeks to investigate improvements to data-driven, ‘top-down’ CCE methods to enable practical application.

Firstly, a method to rationally select data used in the model is investigated. The analysis reported in this thesis found that random sampling does not test the true performance of a model for its future application. Secondly, a method to select input variables that have the largest impact on predicting the construction cost but require the least amount of effort is proposed. The models reached a point whereby expending additional effort to include more input variables did not yield an increased performance and debunked the notion that ‘bottom-up’ estimating approaches are intuitively more accurate. This finding is significant for practitioners as resources expended to collect and store additional data points than required is wasted at the conceptual stage.

Finally, a method to express the conceptual estimate stochastically is proposed. The traditional deterministic approach of relying on a specific number communicates false precision. This thesis proposes combining artificial neural networks with bootstrap sampling to create an empirical distribution of the construction costs and better communicate a likely range of project costs.