PHBench

What is PHBench?

PHBench is a reproducible, research-grade benchmark developed by Vela Partners and the University of Oxford that identifies Series A funding potential directly from Product Hunt launch signals. Trained on 67,292 featured launches from 2019 to 2025 and cross-referenced with Crunchbase funding records, it surfaces statistically significant predictors of early-stage startup success. Its best-performing ensemble model achieves a 4.7× lift over random baseline, and every new Product Hunt launch automatically receives a real-time prediction score. All code, baseline models, and anonymized dataset splits are publicly available, making it a fully open platform for researchers and investors alike.

Key Features

• Large-Scale Launch Dataset

  Built on 67,292 featured Product Hunt posts from 2019–2025, linked to Crunchbase records via deterministic domain matching, with 528 verified Series A raises identified as ground-truth labels.

• 61 Engineered Predictive Features

  Each launch is represented by 61 structured features covering engagement signals (votes, comments, reviews), rank signals (daily, weekly, monthly), maker attributes, temporal patterns, topic flags, and interaction terms.

• Proven Predictive Power

  The top ensemble model (ENS_avg, ENS_ISO, XGB) achieves a 4.7× lift over random baseline on the held-out test set, turning a 0.78% base rate into actionable investment signals.

• Live Real-Time Scoring

  Every new Product Hunt launch automatically receives a prediction score from top-performing models, giving investors and founders an immediate signal without manual analysis.

• Public Leaderboard & Open Submission

  A public leaderboard ranks submitted ML and LLM models on a five-metric evaluation harness, with a held-out blind test set for fair, standardized comparison.

• Research-Grade Reproducibility

  Provides public training, validation, and blind test splits alongside all code and baseline models, following peer-reviewed methodology co-developed with the University of Oxford.

Use Cases

• Venture Capital Screening : VCs and angel investors can filter thousands of Product Hunt launches down to top-ranked prospects, prioritizing due diligence on startups with the highest predicted Series A probability.
• Startup Self-Benchmarking : Founders can compare their own Product Hunt launch performance against historical Series A winners to understand where they stand and which signals to strengthen.
• ML Model Development & Competition : Data scientists and researchers can submit their own models to compete on the public leaderboard, using the standardized dataset and evaluation harness to measure genuine predictive performance.
• Academic Research : Researchers studying startup success, early-stage funding dynamics, or venture signal extraction have access to a clean, citable, documented dataset with peer-reviewed methodology.
• Market Structure Analysis : Analysts can use the temporal performance decay patterns in the dataset to study how funding markets shift over time, from the 2020–2021 boom to the subsequent contraction.

FAQs