A Supervised Approach to Predict Company Acquisition with Factual and Topic Features Using Profiles and News Articles on TechCrunch

Guang Xiang, Zeyu Zheng, Miaomiao Wen, Jason Hong, Carolyn Penstein Rosé, Chao Liu

Abstract

Merger and Acquisition (M&A) prediction has been an interesting and challenging research topic in the past a few decades. However, past work has only adopted numerical features in building models, and yet the valuable textual information from the great variety of social media sites has not been touched at all. To fully explore this information, we used the profiles and news articles for companies and people on TechCrunch, the leading and largest public database for the tech world, which anybody can edit. Specifically, we explored topic features via topic modeling techniques, as well as a set of other novel features of our design within a machine learning framework. We conducted experiments of the largest scale in the literature, and achieved a high true positive rate (TP) between 60% to 79.8% with a false positive rate (FP) mostly between 0% and 8.3% over company categories with a small number of missing attributes in the CrunchBase profiles.

1. Introduction

Predicting Mergers and Acquisitions (M&A) is a high-stakes endeavor for investors, competitors, and employees alike. Traditionally, researchers have relied on numerical financial data—such as cash flow or market-to-book ratios—to build predictive models. However, this approach faces significant hurdles in the startup ecosystem. Financial details for private companies are often a "black box," and traditional models ignore the massive amount of textual information available in tech journalism, which often captures the momentum and "signals" of a company's health better than a balance sheet.

Researchers from Carnegie Mellon and Microsoft Research have developed a supervised learning approach to bridge this gap. By combining traditional factual data with "topic features" extracted from tech news using Latent Dirichlet Allocation (LDA), the team achieved true positive rates as high as 79.8%, proving that the "buzz" in tech journalism is a powerful predictor of a startup's eventual exit.

2. Methodology and Data

The study utilized a massive crawl of the TechCrunch ecosystem from December 2011, encompassing 59,631 company profiles, 105,795 person profiles, and 38,617 news articles covering 5,075 specific companies. The researchers used a Bayesian Network framework to classify companies as "successful" (acquired) or "unsuccessful."

2.1 Feature Engineering

The feature set was divided into two main categories:

  • Factual Features (22 total): Pulled from CrunchBase profiles, these included basic metrics (employee count, company age), financial data (funding rounds, VC/PE presence), and managerial "pedigree" (founder exit history).
  • Topic Features: The core innovation involved using LDA to distill news articles into five distinct topics. This allowed the model to interpret the semantic context of the news coverage surrounding a company.
Table 1: Top 10 words for each of the five latent topics extracted from TechCrunch news articles using LDA.

3. Key Insights and Results

3.1 The Impact of Topic Modeling

Adding topic features significantly boosted the True Positive (TP) rate for categories with heavy news coverage. For example, in the "Web" category, adding topic features increased the TP rate from 58.3% to 78.4%. Similarly, the "Advertising" category saw a jump from 56.8% to 68.0% TP when topics were included. This confirms that the model successfully captured industry-specific signals from the text that factual profiles alone missed.

Table 2: Performance metrics across various company categories, comparing models with 0 topics versus 5 topics. Note the significant TP improvements in sectors like Web and Advertising.

3.2 Predictive Signals

  • The "Buzz" Factor: One of the strongest predictors across almost all categories was the frequency of revisions on a company’s CrunchBase profile. This suggests that community engagement and frequent updates serve as a proxy for a company’s relevance.
  • Founder Pedigree: Managerial features typically outperformed basic factual features like the number of products, reinforcing the investment philosophy of "betting on the jockey, not the horse."
  • Performance Metrics: The model achieved TP rates between 60% and 79.8% for most categories, while keeping False Positive (FP) rates remarkably low (mostly between 0% and 8.3%).

4. Conclusion and Future Work

This research demonstrates that textual context from tech journalism is a potent predictor of company acquisitions. While the model is highly effective, the authors identified areas for improvement, such as the "IPO Problem." Currently, companies going through an IPO are labeled as "unsuccessful" in this binary classification; future models could treat IPOs and acquisitions as a combined "positive exit" class. Additionally, expanding the data scope to include social signals from platforms like Twitter or Quora could further refine the predictive power of the "buzz" factor.

Summary updated

Abstract

Merger and Acquisition (M&A) prediction has been an interesting and challenging research topic in the past a few decades. However, past work has only adopted numerical features in building models, and yet the valuable textual information from the great variety of social media sites has not been touched at all. To fully explore this information, we used the profiles and news articles for companies and people on TechCrunch, the leading and largest public database for the tech world, which anybody can edit. Specifically, we explored topic features via topic modeling techniques, as well as a set of other novel features of our design within a machine learning framework. We conducted experiments of the largest scale in the literature, and achieved a high true positive rate (TP) between 60% to 79.8% with a false positive rate (FP) mostly between 0% and 8.3% over company categories with a small number of missing attributes in the CrunchBase profiles.
BibTeX
@inproceedings{xiang2021supervised,
  author = {Guang Xiang and Zeyu Zheng and Miaomiao Wen and Jason Hong and Carolyn Penstein Rosé and Chao Liu},
  title = {{A Supervised Approach to Predict Company Acquisition with Factual and Topic Features Using Profiles and News Articles on TechCrunch}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/icwsm.v6i1.14306},
}
Figure