Econ Papers Daily / 期刊论文
基于自助随机森林的企业创新预测方法
A bootstrap random forest approach to predicting business innovation
摘要
Despite a vast literature on business innovations, there remains limited global evidence on how internal capabilities, external exposure, and institutional conditions jointly predict innovation across diverse contexts. This study addresses that gap by providing a comprehensive assessment of innovation predictors using machine learning approaches. Drawing on World Bank Enterprise Survey data covering more than 100,000 firms in 122 countries, we apply a Bootstrap Forest model with test-set validation to examine predictors of product and process innovation. The results show that capital investment and formal worker training consistently emerge as the strongest predictors, accounting for at least one-fourth of the out-of-sample predictive power. At the same time, several widely discussed innovation barriers, such as ownership composition, gender of top management, and perceived institutional obstacles, display negligible predictive power. These results suggest that innovation is more closely associated with intentional investments in skills and capital than with static firm attributes or generalized institutional conditions. Subsample analysis reveals meaningful heterogeneities of second-tier-innovation predictors across regions and sectors: export intensity predicts innovation in Europe and Central Asia, while foreign-licensed technology predicts product innovation in Latin America and the Caribbean, and process and product innovations in service sectors. These differences highlight the need for policy approaches that combine broad capability-building priorities with context-specific complementary measures.
中文摘要
本文在商业创新情境下，对吸收能力框架进行了首批大规模样本外检验之一。我们使用世界银行企业调查中覆盖122个国家、超过101，000家企业的数据，应用Bootstrap随机森林模型对产品创新和工艺创新的预测变量进行排序，并在留出的测试集上加以验证。在置换重要性下，研发投资成为两类结果最强的预测变量，其次是资本投资和正式员工培训。这些能力的突出地位在随机种子、不同训练/测试划分、梯度提升模型以及固定效应逻辑设定下均保持稳健。相比之下，若干在参数研究中具有统计显著性的协变量，包括所有权构成和感知到的制度障碍，其预测重要性 consistently 较低。研发投资、资本投资和员工培训不仅在给定样本内与创新相关，而且能够在一个高度异质的国家与制度集合中作为预测变量加以推广——这是以往样本内研究未能提供的证据。最后，结果表明，一些中间层因素，如企业网站和出口强度，对产品创新的预测强于对工艺创新的预测，而企业年龄和管理者经验对工艺创新的预测强于对产品创新的预测。