Machine Learning

With the increasing risk of a recessionary market environment, many investors are asking themselves what impact economic cycles have on the private equity industry.

We assemble a proprietary dataset of 395 private equity (PE) fund prospectuses to analyze fund performance and fundraising success. We analyze both quantitative and qualitative information contained in these documents using econometric methods and machine learning techniques. PE fund performance is unrelated to quantitative information, such as prior performance, and measures of document readability. Measures of fundraising success, in contrast, are correlated to most fund characteristics but are not related to future performance. Meanwhile, machine learning tools can use qualitative information to predict future fund performance: the performance spread between the funds within the top and bottom terciles of predicted probability of success is about 25%. Our findings support the view that in opaque and non-standardized markets, investors fail to incorporate qualitative information in their asset manager selection process, but do incorporate salient quantitative information.

We use machine learning to study how venture capitalists (VCs) make investment decisions. Using a large administrative data set on French entrepreneurs that contains VC-backed as well as non-VC-backed firms, we use algorithmic predictions of new ventures’ performance to identify the most promising ventures. We find that VCs invest in some firms that perform predictably poorly and pass on others that perform predictably well. Consistent with models of stereotypical thinking, we show that VCs select entrepreneurs whose characteristics are representative of the most successful entrepreneurs ( i.e., characteristics that occur more frequently among the best performing entrepreneurs relative to the other ones). Although VCs rely on accurate stereotypes, they make prediction errors as they exaggerate some representative features of success in their selection of entrepreneurs (e.g., male, highly educated, Paris-based, and high-tech entrepreneurs). Overall, algorithmic decision aids show promise to broaden the scope of VCs’ investments and founder diversity.

I conduct an investment screening performance benchmarking between 111 venture capital (VC) investment professionals and a supervised gradient boosted tree (or “XGBoost”) classification algorithm to create trust in machine learning (ML) -based screening approaches, accelerate the adoption thereof and ultimately enable the traditional VC model to scale. Using a comprehensive dataset of 77,279 European early-stage companies, I train a variety of ML algorithms to predict the success/failure outcome in a 3- to 5-year simulation window. XGBoost algorithms show particularly excellent performance in terms of accuracy and recall, which denote the most important metrics in my setup. I benchmark the performance of the selected algorithm against that of the VC investment professionals by providing equal information in the form of 10 company one-pagers via an online survey and requesting respondents to select the five most promising companies for further evaluation. In addition to finding characteristic- specific performance dependencies for VCs, I find that the XGBoost algorithm outperforms the median VC by 25% and the average VC by 29%. Although I do not suggest replacing humans with ML-based approaches, I recommend an augmented solution where intelligent algorithms narrow down the upper part of the deal-flow funnel, allowing VC investment professionals to focus their manual efforts on the lower part of the funnel. Using this approach, they can rely on a scalable but objective pre-selection and focus their manual resources on evaluating the most promising opportunities and putting themselves into the best position to secure these deals.