Decoding Innovation Potential of Japanese Firms with A Machine Learning Approach

This study proposes an efficiency technique for predicting firm-level innovation capabilities utilizing machine learning models for improving the forecast accuracy. The study employed boosted trees and neural boosting models and compared them with traditional statistical regression methods in anticipating a firm’s innovation potential represented by patent applications. The 8 financial internal resources were used as a predictor that emerged from 1991 to 2019. Two key findings of validation are merged to predict an internal indicator of innovation capability. First, firm size is the most important predictor, contributing over half of the predictive power according to the model, followed by R&D intensity and efficiency. Second, as the most effective machine learning model for prediction results, the boosted tree model outperformed the neural boosting model and fixed effect regression, as evidenced by higher R-squared values and lower RASE, demonstrating its ability to capture the complexity of invention activity. These findings provide a solid foundation for future research on firm-level innovation prediction, demonstrating the effectiveness of machine learning algorithms in recognizing complex innovation patterns. The findings also show that selected internal resources could strategically invest in innovation to promote the firm's innovation development.