Prediction of SMEs’ R&D performances by machine learning for project selection

Abstract

To improve the efficiency of government-funded research and development (R&D) programs for small and medium enterprises, it is necessary to make the process of selecting beneficiary firm objective. We aimed to develop machine learning models to predict the performances of individual R&D projects in advance, and to present an objective method that can be utilized in the project selection. We trained our models on data from 1771 R&D projects conducted in South Korea between 2011 and 2015. The models predict the likelihood of R&D success, commercialization, and patent applications within 5 years of project completion. Key factors for predicting the performances include the research period and area, the ratio of subsidy to research budget, the firm’s region and venture certification, and the average debt ratio of the industry. Our models’ precisions were superior to qualitative expert evaluation, and the machine learning rules could be explained theoretically. We presented a methodology for objectively scoring new R&D projects based on their propensity scores of achieving the performances and balancing them with expert evaluation scores. Our methodology is expected to contribute to improving the efficiency of R&D investment by supplementing qualitative expert evaluation and selecting projects with a high probability of success.