Effort Estimation of Software Projects Using Deep Learning Based on Proprietary Dataset

Abstract

This research proposes a novel approach to software effort estimation in the Fintech domain by leveraging deep learning and hybrid machine learning techniques. Traditional estimation methods such as COCOMO and Function Point Analysis struggle with the complexity and dynamic nature of modern software projects. To address these limitations, this study develops a deep learning - based model that compares traditional machine learning algorithms (e.g. Random Forest, XGBoost) with deep learning architectures (e.g. LSTM, MLP) using a proprietary dataset. The methodology involves feature-level fusion and model stacking techniques to enhance predictive accuracy. The dataset undergoes data augmentation, normalization, and encoding. The models are trained and optimized through hyperparameter tuning and evaluated based on key metrics such as RMSE, R-squared, and MAPE. The system is designed to improve estimation accuracy, scalability, and reliability while providing a user-friendly interface for project managers and pre-sales teams. The results demonstrate that the hybrid model significantly outperforms standalone models. The feature-level fusion approach achieves the highest accuracy, reducing the MSE and MAPE compared to traditional models. The findings validate the effectiveness of combining deep learning with machine learning for effort estimation, making it a promising tool for project planning, resource allocation, and decision-making in FinTech software development.