Exploring Bayesian techniques for Code Technical Debt prediction

Abstract

Background. Code Technical Debt is a crucial concept in software engineering, representing the long-term costs associated with suboptimal coding practices. Accurate prediction of Code Technical Debt is essential for proactive maintenance and software quality assurance. Some implemented techniques provide promising results in predicting Code Technical Debt, but there are no consistent and robust methods capable of fully capturing the factors that influence the progression of Code Technical Debt.

Objective. The study aims to assess the effectiveness of Bayesian techniques in predicting Code Technical Debt, focusing on the SQALE index as a key metric. The goal of analyzing different Bayesian approaches is to identify models that offer reliable and accurate forecasts across periodic and non-periodic datasets.

Methods. Four Bayesian models, Damped Local Trend, Exponential Smoothing, Dynamic Linear Model, and Dynamic Generalized Linear Model, were applied to datasets derived from 31 ASF Java projects. We predicted future SQALE index values to assess the accuracy of their predictive performance.

Results. The study demonstrated that Bayesian analysis is a robust approach for predicting Code Technical Debt across both serialized and non-serialized datasets, effectively capturing temporal dynamics and delivering reliable forecasts. While Bayesian models did not outperform traditional time series models like ARIMAX and SARIMAX in overall accuracy, they addressed key limitations by being applicable to non periodic datasets. The Damped Local Trend model exhibited superior performance for periodic datasets, efficiently capturing trends and seasonality. Conversely, the Dynamic Generalized Linear Model excelled with non-periodic datasets, providing the highest accuracy in handling their complex and irregular patterns.

Conclusion. This study highlights the potential of Bayesian models for Code Technical Debt prediction, emphasizing their adaptability to diverse dataset characteristics. The findings provide practical guidance for software practitioners in selecting predictive models tailored to specific project needs.