Comparing Univariate and Multivariate Time Series Models for Technical Debt Forecasting

Technical debt (TD) is a successful and widely used metaphor that expresses the quality compromises that can yield short-term benefits but may negatively affect the overall quality of a software product in the long run. There is a vast variety of techniques and methodologies that have been proposed over the past years to enable the identification and estimation of TD during the software development cycle. However, it is only until recently that researchers have turned towards the investigation of methods that focus on forecasting its future evolution. Getting insights on the future evolution of TD can enable on-time decision-making and allow stakeholders to plan preventive strategies regarding TD repayment. In our previous studies, we have investigated time series analysis and Machine Learning techniques in order to produce reliable TD forecasts. In our current attempt, we aim to explore the capabilities of a statistical ARIMA model both in a univariate and a multivariate fashion. More specifically, the present paper investigates whether the adoption of an ARIMA model that takes into account, in addition to the TD value itself, various TD-related indicators may lead to more accurate TD predictions than its univariate alternative. For this purpose, dedicated models are constructed, evaluated, and compared on a dataset of five long-lived, open-source software applications.