Research

Similarity among Startups in Accelerators

Job market paper

Start-up accelerators are key players in the entrepreneurial ecosystem, offering cohort-based programs to accelerate the growth of early-stage start-ups. Despite a growing literature on the benefits of accelerator participation, little is known about whether start-ups gain from exposure to peers within accelerator cohorts. In this paper, I investigate the role of business similarity among start-ups within accelerator cohorts and argue that it may lead to improved post-acceleration performance via greater exposure to relevant knowledge and peers. I test this argument using a rich, unique dataset comprising 2,505 start-ups accelerated in 128 cohorts run by eight U.S.-based accelerators between 2005 and 2018. I find strong evidence that business similarity is positively associated with post-acceleration performance: start-ups accelerated with more similar peers raise funding more quickly and exhibit improved long-term outcomes in terms of survival, employee growth and exit via acquisition. In preliminary analyses, I also show that the effect of similarity appears to exhibit decreasing marginal returns, which may signal the presence of factors that reduce the benefit of similarity at higher levels. These results offer actionable insights for accelerator managers seeking to design effective programs and for start-ups evaluating participation in an accelerator.

User Attention and Targeted Advertising in Social Media Platforms: When Does Attention Brokerage Benefit Users?

with Claudio Panico (Bocconi) and Carmelo Cennamo (CBS) - Revise & Resubmit

Social media platforms, as attention brokers, face a fundamental tension between the users they serve and the advertisers they depend on. We develop a formal model to investigate how a platform’s ability to broker user attention impacts pricing, advertising strategies, and the welfare of all participants. Our findings reveal that the welfare implications of attention brokerage are critically contingent on the platform’s model. On platforms offering high standalone benefits, such as Facebook, we find a fundamental misalignment of interests. While platform and advertiser profits grow with enhanced attention-brokering, user welfare follows an inverted-U shape, declining as users are exposed to a greater volume of lower-quality and irrelevant ads. This suggests a potential for user overexploitation. In contrast, on platforms with negligible standalone benefits, like Pinterest, market forces align the incentives of all parties. Here, the welfare of users, advertisers, and the platform all increase monotonically with the platform’s brokering ability. This research offers a novel explanation for the persistence of irrelevant ads, rooting it in this platform-mediated incentive misalignment. Furthermore, we derive clear policy implications, showing that the optimal level of attention-brokering a regulator would choose depends on the specific welfare standard being maximized. Our results demonstrate that regulatory approaches must be nuanced, as interventions appropriate for platforms with high standalone benefits may be counterproductive for others

Investigating the Effect of General Partners in Venture Capital using Machine Learning

with Corrado Botta (Bocconi) and Mario D. Amore (Bocconi)

Despite extensive research on the determinants of venture capital (VC) performance, the assumption that general partners (GPs) meaningfully influence fund performance has not been systematically validated using rigorous predictive methods. In this paper, we address this gap by employing machine learning algorithms and out-of-sample prediction to test whether individual GPs actually contribute to VC fund performance. Using a comprehensive dataset of 29,021 quarterly observations spanning 722 funds managed by 811 GPs between 1997-2022 from PitchBook, we compare predictive accuracy across three model specifications: a baseline model (i.e., with fund characteristics only), a GP-enhanced model (i.e., with GP dummies) and a VC firm-enhanced model (i.e., with firm dummies). We employ multiple machine learning algorithms including regularized linear methods, tree-based ensembles and kernel methods. Our findings demonstrate that GPs matter for explaining VC fund performance. The best-performing algorithm (i.e., the random forest model) shows that including GP dummies improves out-of-sample R-squared from 74.2% to 77.8%. Importantly, the GP effect consistently exceeds the VC firm effect across all algorithms, suggesting that individual-level analyses provide greater insight than firm-level studies, which were the standard in the nascent days of research on VC. Finally, these results validate the assumption underlying theories of VC performance persistence and highlight the importance of human capital in investment management.