In this study, several machine learning-based experimental methods are used to analyse firms' research and development (R&D)-related activities and predict their technological innovation performance. Using unbalanced panel data from the CSMAR database for all listed firms in China from 2008 to 2018, we analyse the firms' basic information, R&D investment, patent application and authorization activity, financial status, and human capital. We use a logistic regression model, decision tree model, three weak classifiers random forest model, XGBoost model, and two weak classifiers gradient boosting decision tree (GBDT) model to integrate strong classifiers separately. A comparison of the results produced using the different models shows that the performance of the XGBoost model is better than that of the other models in terms of net profit, total sales revenue, and the number of invention patent applications as a proportion of the total number of patent applications. However, the performance of the GBDT model is significantly better than that of the other models in terms of the number of patent applications per 100,000 yuan of R&D expenditure. The results of this study can help scholars to accurately predict the innovation performance of firms and help business managers to make better decisions to improve the innovation performance of their firms in the current era of rapid technological change. (C) 2020 Elsevier B.V. All rights reserved.