A hybrid forecasting model to predict the duration and cost performance of projects with Bayesian Networks

Abstract

This paper presents a new hybrid forecasting model to predict the final time and cost of a project using input parameters from the project scheduling and risk analysis literature. The hybrid method integrates two well-known risk models. A Structural Equation Modeling constructs and validates a theoretical risk model to represent known relations between project indicators and the project performance. A Bayesian Networks is used to train the theoretical model using artificial project data from the literature. These two integrated models are then used to predict the final duration and cost of a new unseen project.

The accuracy of this integrated model is compared with other well-known forecasting methods from the literature. The computational experiments on a set of 33 empirical projects show that risk models demonstrate a noteworthy advantage for time and cost forecasting. To show the usefulness of this method, it is compared with a set of known machine learning forecasting algorithms.

These static predictions of risk models are also compared with some well-known dynamic forecasting methods that continuously update the time/cost predictions along the project progress. These dynamic models make use of predictors from the earned value management and earned duration management literature. The results show that the static risk models offer more precise forecasts than the dynamic methods in the first half of the project progress for time forecasting, but then loose their power in favor of the dynamic forecasts.