Predicting low-performance government-sponsored R&D project: comprehensive framework and machine learning approach

In response to the Korean R&D paradox, which has intensified during the post-global financial crisis, it would be critical to seek ways to improve the economic outcomes of R&D projects. In this context, this study aims to propose a framework and methodology to predict low-performance R&D projects. We considered various factors affecting the project's success, including funding, technological, environmental, organisational, and initial valuation factors. Our framework was applied to 797 R&D projects, introducing a novel BU ratio as a performance evaluation metric together with an F1 score. Then, the results of three machine learning algorithms were compared, and the XGBoost algorithm showed the best performance for predicting low-performance R&D projects. This study gained insight into the behaviour of the selected models through the SHAP value. Government funding, in terms of ratio and amount, was revealed as the most dominant feature in predicting performance groups. Organisational factors were crucial for high-performance projects, whereas funding factors dominantly affected low-performance projects. Our analytical framework and findings will contribute to academic knowledge, as no research has yet explored this. Moreover, this study will provide practical implications for funding agencies to proactively manage low-performance R&D projects and improve their economic success.