Prediction of Sprint Delivery Capability in Agile Software Development Using Machine Learning

Abstract

Scrum is one of the most prominent agile development frameworks as it delivers the project in many sprints in an incremental manner. Sprint delivery capability is a crucial concept in the scrum projects as it assesses whether sprint can capable of delivering its issues on its length. It is difficult to consider forecasting sprint delivery capability in large scrum projects as multiple sprints are processed simultaneously and agile dynamics. Machine Learning (ML) models were applied in the existing studies to forecast sprint delivery but, had limitations in finding non-linear and hidden patterns. To address these problems, this research exploited the untapped ensemble learning approach to bridge the gap to improve the prediction of sprint delivery capability in different project progression levels. The study gathered the data from one of the large open-source scrum projects, JIRA which included 1873 iterations and 10853 issues in table format spanning at 30%, 50%, and 80% project progression levels. Data preprocessing enabled to refinement of the data for analysis. The author enhanced the feature aggregation by using both statistical and automatic feature learning to derive the new features from the issue table and combine them with the iteration table. The key features affecting sprint delivery capability were identified using hybrid feature selection approach. The study developed the untapped ensemble models by choosing the different combinations of ML, Deep Learning (DL), and ensemble algorithms using appropriate ensemble methods and combination rules. The author examined the models and concluded that Categorical Boost (CatBoost) gained the highest accuracy of 96.8% at the 80% project progression level. This prediction assists in improving sprint planning, delivering the software on time, and getting more insights about sprint performance.