Success in predicting troubled projects requires an adaptable model

Abstract

Software risk management is a well established discipline that has generated continued academic interest as the complexity and nature of software projects have evolved over time. Traditional risk management techniques have been focused on identifying and codifying best practices that prevent or reduce the failure rate. Large IT organizations have assimilated many of these findings in their risk management practice. However, adopting these best practices does not guarantee that risk is eliminated or even reduced to an acceptable level new software development models driven by globalization, competition and an ever changing software landscape create new patterns of trouble. Systems that predict trouble early in the project life cycle have had significant impact on our global portfolio of thousands of projects. Managers have mitigated the risks that were identified in our system. As such, the key predictors and their importance have shifted. We have discovered that our prediction model requires frequent updates and this was not included in our initial design. We have learned that it is equally important to design and select statistical algorithms that will support automated model retraining as a way of incorporating a feedback loop of human behavior responding to our predictions. © 2011 IEEE.