Ensemble Effort Estimation for Novice Agile Teams

CONTEXT: To establish a reliable development plan, developers should investigate the software being developed. One main challenge for developers is estimating the effort required to develop the software. Agile teams deliver the software in a set of iterations, with each iteration containing user stories. Therefore, unlike traditional development, software development effort estimation (SDEE) in agile should focus on the user stories level. An inaccurate estimation has detrimental consequences for software development such as poor resource allocation or the delivery of low-quality software. However, limited works have developed new estimation methods for agile projects compared to traditional ones. OBJECTIVES: This study introduces an ensemble model for estimating efforts in agile user stories development. It also creates a new dataset with 140 user stories, aiming for future research use. METHODS: This research followed the Design Science Research methodology (DSR). Six individual models were examined to build the ensemble model. The top three models - Extra Trees, K-Nearest Neighbors, and MultiLayer Perceptron - were employed. The model's performance was assessed through Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). Additionally, an experiment tested the model's efficacy on real software projects by novice teams. RESULTS: The results show that the ensemble model outperformed individual models, as it scored 0.78 in MAE, 1.62 in MSE, and 1.15 in RMSE. The experiment results showed that the model outperformed human estimation and proved its effectiveness in improving the accuracy of human estimation. CONCLUSION: The findings demonstrate the model's success in refining effort estimates for novice Agile teams, leading to fewer errors. Practically, it means enhanced project planning and resource management. Additionally, developers' estimation confidence improved, indicating a positive impact on team dynamics and decisionmaking.