Scholarlink

<p>Identifying who is likely to engage in entrepreneurship - both in Intention (EI) and in realized Entrepreneurial Outcome Behaviour (EOB) - is essential for designing training, incubation, and policy interventions. By implementing eight waves of data from the Global Entrepreneurship Monitor (GEM) (2014–2021) generated from 43 countries (<i>n</i> &gt; 1.2&#xa0;million), this paper identifies a compact and generalisable predictive framework for EI and EOB using an explicit predictive model based on interpretable tree-based machine learning (Classification and Regression Trees (CART) and Stochastic Gradient Boosting (TreeNet)). The aim of the paper is to test whether a small, theory-driven subset can predict entrepreneurial behaviour across countries and years. This is achieved through the use of a construct-level framework that embeds intention models (TPB/EEM) within micro-foundations of management (RBV: human and social capital; dynamic capabilities: perceiving-seizing-transforming) and the institutional/economic context. The results show that the resulting five-variable EI model achieves a sensitivity of 75% and a baseline logistic regression of 12%, respectively. In contrast, the six-variable model for EOB achieves a sensitivity of 82% and a baseline logistic regression of 15%. Furthermore, performance does not change across pre-pandemic and COVID-era data. Finally, the model highlights actionable levers - strengthening self-efficacy, reducing institutional frictions that slow the intention-to-action transition, and reinforcing ties with role models - supporting scalable, low-cost interventions. The approach illustrates the ability of parsimonious, transparent, and self-optimising algorithms to uncover and maintain predictive structures from immensely scaled, class-biased behavioural data, while still adhering to the underlying theory of the chosen constructs and algorithms.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A new predicting model of an entrepreneurial behaviour: interpretable machine learning on GEM data for multi-year and multi country analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月7日

查看全文
  • Ray Duplock,
  • Mirko Perano,
  • Gian Luca Casali,
  • Claudio Del Regno

摘要

Identifying who is likely to engage in entrepreneurship - both in Intention (EI) and in realized Entrepreneurial Outcome Behaviour (EOB) - is essential for designing training, incubation, and policy interventions. By implementing eight waves of data from the Global Entrepreneurship Monitor (GEM) (2014–2021) generated from 43 countries (n > 1.2 million), this paper identifies a compact and generalisable predictive framework for EI and EOB using an explicit predictive model based on interpretable tree-based machine learning (Classification and Regression Trees (CART) and Stochastic Gradient Boosting (TreeNet)). The aim of the paper is to test whether a small, theory-driven subset can predict entrepreneurial behaviour across countries and years. This is achieved through the use of a construct-level framework that embeds intention models (TPB/EEM) within micro-foundations of management (RBV: human and social capital; dynamic capabilities: perceiving-seizing-transforming) and the institutional/economic context. The results show that the resulting five-variable EI model achieves a sensitivity of 75% and a baseline logistic regression of 12%, respectively. In contrast, the six-variable model for EOB achieves a sensitivity of 82% and a baseline logistic regression of 15%. Furthermore, performance does not change across pre-pandemic and COVID-era data. Finally, the model highlights actionable levers - strengthening self-efficacy, reducing institutional frictions that slow the intention-to-action transition, and reinforcing ties with role models - supporting scalable, low-cost interventions. The approach illustrates the ability of parsimonious, transparent, and self-optimising algorithms to uncover and maintain predictive structures from immensely scaled, class-biased behavioural data, while still adhering to the underlying theory of the chosen constructs and algorithms.