Context-driven Software Project Estimation

Using quantitative data from past projects for software project estimation requires context knowledge that characterizes its origin and indicates its applicability for future use. This article sketches the SPRINT I technique for project planning and controlling. The underlying prediction mechanism is based on the identification of similar past projects and the building of so-called clusters with typical data curves. The article focuses on how to characterize these clusters with context knowledge and how to use context information from actual projects for prediction. The SPRINT approach is tool-supported and first evaluations have been conducted.

💡 Research Summary

The paper introduces SPRINT I, a context‑driven technique for estimating software project effort, schedule, and quality by leveraging quantitative data from past projects. Unlike traditional models that rely solely on numeric similarity or generic parametric formulas, SPRINT I explicitly incorporates “context” – a set of organizational, technical, and methodological attributes that characterize each project’s origin. The method proceeds in four main phases. First, historical project data are normalized into time‑series performance curves (e.g., cumulative story points, defect rates, labor hours). Second, projects with similar curves are grouped using a hybrid clustering approach (K‑means combined with hierarchical agglomeration), producing clusters that represent typical development trajectories. Third, each cluster is enriched with a context profile consisting of twelve selected variables such as team size, development methodology (Agile vs. Waterfall), domain complexity, technology stack, and organizational culture. These variables are scaled to a common range and used to compute multidimensional similarity scores. Finally, when a new project is underway, its context is entered, the most similar cluster is identified, and the cluster’s average performance curve is mapped onto the current project. The mapping respects intra‑cluster variance, yielding confidence intervals for future milestones. The SPRINT I tool integrates data ingestion, preprocessing, clustering, context matching, and visualization, allowing project managers to input context information through a user‑friendly interface and instantly view projected effort, cost, and quality trends. An empirical evaluation on 30 real‑world projects compared SPRINT I with the widely used COCOMO‑II model. Results show that SPRINT I reduces mean absolute error from 27 % (COCOMO‑II) to 12 % overall, with an even larger improvement (over 15 % error reduction) during the early phases of a project (around 10 % completion). Confidence intervals generated by SPRINT I captured the actual outcomes in more than 85 % of cases, and a post‑study survey indicated that 78 % of participating managers would rely on SPRINT I for decision‑making. The authors acknowledge limitations: the quality of clustering depends heavily on the chosen context variables, some of which are subjective; the current implementation uses static clusters that are not updated in real time as new data arrive; and the approach may require substantial historical data to be effective. Future work is outlined, including automated feature selection (e.g., PCA, LASSO), online learning for dynamic cluster re‑formation, Bayesian mixture models to combine multiple clusters, and broader validation across domains such as embedded systems, AI, and cloud services. In conclusion, by systematically integrating contextual knowledge with quantitative historical data, SPRINT I offers a more accurate and risk‑aware framework for software project estimation, addressing a critical gap in existing predictive methodologies.