Artificial neural networks for pre-tender engineering consultancy cost estimation with missing-data representation and robustness analysis

Estimating engineering consultancy service costs during the tender preparation stage remains a challenging task in public procurement. Although artificial intelligence has been widely applied to construction cost estimation, empirical studies specifically addressing pre-tender engineering consultancy services remain limited. This study analyzes 205 engineering consultancy contracts awarded between 2022 and 2025 in Türkiye and identifies a substantial discrepancy between estimated costs and awarded contract prices, with the total awarded contract value being approximately 31% lower than the total estimated cost. To investigate whether tender-stage engineering consultancy contract prices can be predicted using data-driven techniques, four multilayer perceptron artificial neural network (MLP-ANN) configurations were developed using variables commonly available in tender documentation and trained with the Levenberg-Marquardt algorithm. A flag-based representation strategy was proposed to incorporate incomplete technical personnel information without excluding incomplete observations from model development. In addition, supplementary robustness analyses based on repeated predefined train-test partitions were conducted to evaluate the stability of the proposed modelling framework under alternative partitioning strategies. The results demonstrate that ANN-based models can effectively estimate tender-stage engineering consultancy service costs and that the proposed flag-based representation enables practical use of incomplete procurement data while maintaining strong predictive performance. The robustness analyses further provide complementary evidence regarding the stability of the proposed modelling framework under different train-test partitioning conditions.