# PHBench

> PHBench is the first public benchmark for predicting whether a Product Hunt launch will raise Series A funding within 18 months. It maps 67,292 featured Product Hunt posts (2019-2025) to 528 verified Series A outcomes. The champion model scores every new launch weekly. A launch simulator helps makers predict their daily rank before they launch.

## What is PHBench?

PHBench is a machine learning benchmark, weekly prediction service, and launch simulator for Product Hunt. It provides:
- 67,292 featured Product Hunt posts with 68 engineered features
- 528 verified Series A raises within 18 months of launch
- Public train/validation splits and blind test set
- Official evaluation metrics: F0.5, Average Precision, P@k
- A public leaderboard at phbench.com
- Weekly predictions ranking new launches by Series A probability
- A launch simulator to predict daily rank before launching
- A REST API for programmatic access to predictions

## Weekly Predictions

Every week, PHBench scores all featured Product Hunt launches using the champion ML model and publishes ranked predictions:
- URL: https://phbench.com/predictions
- Updated every Sunday
- Enriched with company data: founding year, location, total funding, revenue, valuation
- Companies already funded (Series A+) are filtered out using Gemini with Google Search grounding
- Weekly newsletter sent to 100+ subscribers
- Historical archive with weekly, monthly, and yearly views

## Launch Simulator

Makers planning a Product Hunt launch can simulate their expected daily rank:
- URL: https://phbench.com/simulate
- Input: product name, tagline, description, topics, makers, hunter, images, video
- Output: feature readiness score, predicted rank, launch day playbook
- Auto-lookup: enter PH usernames and we fetch follower counts
- Import: paste a public PH launch URL to auto-fill all fields
- Markdown import: paste a launch brief to auto-fill fields

## API

Predictions are available via REST API:
- URL: https://phbench.com/api/predictions
- Authentication: API key via x-api-key header
- Parameters: ?week=2026-06-08, ?month=2026-06, ?year=2026, ?days=7
- Returns: scored posts with rank, category, founded year, location, funding, revenue, valuation
- Documentation: https://phbench.com/api-docs
- Request API key: benchmark@vela.partners

## Key Findings

- Champion ensemble model achieves 4.7x lift over random on held-out test set
- Team size x engagement interaction is the strongest predictor
- B2B categories (API, Payments, Fintech) convert at 3x baseline
- Daily rank #1 launches raise Series A at 2.2x the rate of unranked launches
- Vanilla LLMs (Gemini 2.5 Flash, Gemini 3 Flash) achieve AP=0.034 at best
- The 2020-2021 VC funding boom and 2022-2023 contraction are visible in launch conversion rates

## Dataset

- 288,692 total Product Hunt posts collected (2019-2025)
- 67,292 featured posts (benchmark universe)
- 528 positive labels (Series A within 18 months)
- Class imbalance: 1:126 (0.78% positive rate)
- Matched to Crunchbase via domain resolution
- Available on HuggingFace: https://huggingface.co/datasets/ihlamury/phbench

## Participation

To participate in the benchmark:
1. Download dataset from HuggingFace: https://huggingface.co/datasets/ihlamury/phbench
2. Train a model on phbench_train.csv
3. Generate predictions on phbench_test_features.csv
4. Submit predictions CSV to benchmark@vela.partners
5. Results posted to leaderboard at phbench.com

## Research

Paper: PHBench: A Benchmark for Predicting Startup Series A Funding from Product Hunt Launch Signals
Authors: Yagiz Ihlamur, Ben Griffin (Oxford), Rick Chen (Oxford)
arXiv: https://arxiv.org/abs/2605.02974
Submitted to: NeurIPS 2026 Evaluations & Datasets Track

## Built By

- Yagiz Ihlamur — Senior Technical Product Manager at Amazon, AI researcher
- Ben Griffin — Researcher at University of Oxford / Vela Partners
- Rick Chen — Researcher at University of Oxford / Vela Partners
- In collaboration with Vela Partners (https://www.vela.partners/)

## Links

- Website: https://phbench.com
- Predictions: https://phbench.com/predictions
- Simulator: https://phbench.com/simulate
- API Docs: https://phbench.com/api-docs
- GitHub: https://github.com/ihlamury/phbench
- Dataset: https://huggingface.co/datasets/ihlamury/phbench
- Paper: https://arxiv.org/abs/2605.02974
- Vela Partners: https://www.vela.partners/
- Contact: benchmark@vela.partners