On Forecasting Project Activity Durations with Neural Networks

Peter Zachares, Vahan Hovhannisyan, Carlos Ledezma, Joao Gante, Alan Mosca

Abstract:

Accurately forecasting project end dates is an incredibly valuable and equally challenging task. In recent years it has gained added
attention from the machine learning community. However, state of the art methods both in academia and in industry still rely on expert opinions and Monte-Carlo simulations. In this paper, we formulate the problem of activity duration forecasting as a classification task using a domain specific binning strategy. Our experiments on a data set of real construction projects suggest that our proposed method offers several orders of magnitude improvement over more traditional approaches where activity duration forecasting is treated as a regression task. Our results suggest that posing the forecasting problem as a classification task with carefully designed classes is crucial for high quality forecasts both at an activity and a project levels.