Automation of Sprint Cost Estimation in Agile Projects Using a Random Forest-Based Decision Support System

Abstract

Purpose: The manual construction of a Work Breakdown Structure (WBS) is inherently susceptible to human cognitive biases; hence, the objective of this research shifts toward Intelligent Automation, leveraging a Decision Support System (DSS) to replicate expert financial logic.

Methods: The dataset comprises 30 validated sprint records from 23 internal projects; other WBS documents were excluded due to incomplete sprint details. Pre-processing involved completeness checks and One-Hot Encoding of team role compositions, without feature scaling. A Decision Tree Regressor, Random Forest Regressor, and Hybrid Model (Voting Regressor) were evaluated using 3-fold cross-validation based on MAPE, RMSE, and R² metrics.

Result: The Random Forest Regressor demonstrated the best relative performance among the three models, achieving a MAPE of 23.55%, an RMSE of IDR 35.27 million, and an R² of -0.535. Across the 30 sprint predictions, the Random Forest model showed a mean deviation of 23.55% ± 31.69%, a median of 9.98%, and a range of 0.78%–125.33%, highlighting performance variation across sprints and supporting the model's application within a human-in-the-loop framework. Feature importance analysis identified `sprint_duration_days` as the dominant feature, followed by `role_programmer_jr` and `role_qa_sr`.

Novelty: These findings demonstrate the system's practical contribution, particularly in streamlining budgeting bureaucracy by accelerating the process from days of manual preparation to just a few minutes through automated execution.