Effort Estimation in Agile Development
MetadataShow full metadata
Context: Organizations that develop software deploy considerable resources on projects in order to drive innovation and transformation. However, enduring evidence indicates that many software projects fail to deliver within time or budget and do not achieve business objectives. One of the reasons underlying such failures has been an inability to provide accurate forecasts of the time and resources needed for project completion. This may be due to the use of inherently poor estimation techniques, the poor use of good estimation techniques, the lack of use of any such techniques, or some combination of the three. Numerous estimation techniques have been proposed for use in the software industry to forecast project or artefact size, the human effort required, the likely cost of development and the project schedule, with highly varying degrees of success. Some have proved to be effective in traditional software development (TSD) but their effectiveness in agile software development (ASD) is yet to be determined. ASD is currently the most popular methodology for the development of software because of its emphasis on fast release cycles, frequent delivery of value-added features to customers, and close collaboration between the product owner, the customer and the development team. The prevailing approach to forecast size, human effort and schedule for a software project in ASD is simple, which is a benefit, but still there are gaps and limitations in that approach. The current study is proposed to address the most important of these. Purpose: This research study is conducted to understand and improve the utility of forecasts of size, effort, cost and schedule for ASD projects through variants of approaches based on a range of relevant objective and subjective indicators, including the size of stories, the velocity/capacity of the team, sprint and system complexity, the domain knowledge, expertise and experience of the team, stakeholder coordination, team relationships, project/sprint risks, and organizational culture, type, and work environment. Methods: The Design Science Research Methodology encompasses three phases, being problem identification, solution design and evaluation. These phases have been enacted in this study through reviews of the state-of-the-art and the state-of-practice, quasi-experimentation, and detailed validations. The core empirical work in the study is conducted via quasi-experimental methods and face-to-face interviews with 21 IT experts. COEEMO – the proposed model of estimation along with two other suggested techniques (Bottom-Up and variants of Program Evaluation and Review Technique (PERT)) are applied to three real-industry ASD projects of AUSIT, and their performance is assessed using Mean Absolute Error (MAE), Balanced Relative Error (BRE), BREbias, and Standardized Accuracy (SA). To determine the statistical significance of the above suggested methods one-way ANOVA (Parametric), and Kruskal-Wallis (Non-Parametric) tests are conducted, while the effect size (Glass delta ─ ∆) is calculated to determine the practical efficacy of COEEMO. Further, COEEMO is corroborated through empirical and laboratory validations: the model is implemented by using Mamdani Fuzzy Inference (MFIS) in the MATLAB toolbox and applied to the data of 85 sprints from ten previously completed ASD projects. Contributions: There are three primary contributions deriving from this research in the area of effort estimation with the intent to improve accuracy in effort estimates for ASD projects: design and development of the novel COEEMO model, which delivers a pooled effect size Glass delta ∆ of 0.23, outperforming the existing method of estimation by exhibiting 56% and 43% overall improvements in accuracy during the field-trials and empirical evaluations, respectively; a tailored approach of PERT, referred to as PERT’s Variant-1, which shows 25% enhancement in accuracy during field-trials; and the use of Expert Judgement based on a Bottom-Up approach, demonstrating a 47% improvement in accuracy for effort estimation as compared to AUSIT’s existing method during the field-trials.