Exogenous Time-Series Modeling with Structural Graph Learning for Agile Project Development

Modern Agile pipelines create artefact networks whose evolving structure and exogenous signals fundamentally shape sprint-level effort, yet remain underutilized in conventional forecasting pipelines. This study introduces an integrated estimation approach that combines seasonal time-series modeling with exogenous regressors and a PSI-GNN to explicitly encode semantic dependencies, co-change links, and ownership ties among user stories, bugs, and issues. Using the TAWOS dataset, which aggregates over 31000 issues across 26 diverse projects, we benchmark a baseline regressor, SARIMA and SARIMAX models, and the PSI-GNN to evaluate how temporal, exogenous, and structural signals jointly affect effort prediction. The results indicate that enriching ARIMA-family models with exogenous variables improves accuracy, while the PSI-GNN achieves the lowest error rates by capturing latent relational effects that standard time-series models overlook. This demonstrates that integrating structural graph learning with time-series forecasting offers a robust pipeline for sprint-level story point estimation in realistic Agile/DevOps environments.