phbench

PHBench is a research benchmark for testing whether machine learning models can predict startup fundraising outcomes from publicly visible launch data. Specifically, it asks: given a product's launch on Product Hunt (a popular website where startups debut new products), can you predict whether that startup will raise a Series A funding round within the next 18 months?

The dataset covers 67,292 featured Product Hunt launches from 2019 to 2025, matched against Crunchbase records to identify 528 confirmed Series A raises within the 18-month window — about 1 in every 127 launches. Each launch is described by 61 engineered features derived from the post itself, such as upvotes, comments, daily and weekly rankings, and website metadata. The test set labels are kept private; you submit your probability predictions to a leaderboard at phbench.com and receive scores back.

The benchmark reports five evaluation metrics, with a precision-weighted score (F0.5) as the primary leaderboard metric. The current best-performing entry is an ensemble of three models that achieves 4.7 times the precision of a random baseline on the held-out test set. A logistic regression baseline and a large language model zero-shot baseline are included for comparison.

Researchers or builders interested in predicting startup traction, building investor tools, or studying what public launch signals correlate with funding would find this benchmark useful. The dataset requires access request via Hugging Face. The code and paper (submitted to NeurIPS 2026) are written in Python and released under CC BY 4.0.