How I Built a Deal-Flow Signal From Public GitHub Data (219 Fundraises Backtested)

Eighteen months ago I noticed something odd while stalking a startup's GitHub org before an angel check: their commit activity had tripled about a month before they announced their round. Not after. Before. I'm a data person, so I did what data people do. I stopped looking at one company and started looking at all of them. The hypothesis Startups behave differently on GitHub in the weeks before a fundraise. They clean up repos, ship faster, onboard new engineers hired ahead of the announcement, and spin up infrastructure for the growth they're about to buy. All of that is visible in public API data if you know what to measure. So I built a tracker. It now watches 4,200+ startup GitHub orgs weekly across 20 sectors, computing three signals: Commit velocity — commits across tracked repos in a trailing 14-day window vs. the prior window Contributor growth — distinct active contributors over 30 days vs. baseline (new hires show up here first) New repo creation — fresh public repos in 30 days (infrastructure buildout) These get combined into a composite score and classified: breakout, acceleration, steady, cooling. Did it actually work? I backtested the pattern against 219 documented fundraises. Companies showing the acceleration pattern raised at roughly 3.4× the base rate, with the signal appearing 21–47 days before the public announcement in the cases where it fired. The full methodology is published as an SSRN preprint (DOI: 10.2139/ssrn.6606558) with the backtest dataset on Zenodo (10.5281/zenodo.19650920). I wanted this to be checkable, not another black-box "AI signal." What it caught this quarter A few live examples from the current Q3 2026 period (all public data, verifiable on GitHub right now): Fleetbase (logistics OS, pre-seed): commit velocity up ~16× over its prior window, 7 active contributors (+73% in 30 days). Classic engineering-hiring-burst shape. Tuva Health (healthcare analytics): 48 active contributors, 3 new repos in 30 days. Infrastructure buildout pattern. Swell (e-commerce infra): contributor count doubled in 30 days, 2 new repos. Hiring burst. Reliza (DevOps/supply chain, pre-seed): contributors up 5× from a small base, 2 new repos. Will all of these raise? No. That's the honest part. The honest limitations ~23% false-positive rate. Acceleration sometimes means a big enterprise deal, an open-source push, or a hackathon, not a round. Stealth-first companies are invisible. No public repos, no signal. AI-heavy startups are noisy. Model releases create commit spikes that look like fundraise prep. Small-base distortion. A 3-contributor org going to 6 is +100% but means little. The composite score penalizes low-base orgs, but it's not perfect. Anyone who sells you a startup signal without a false-positive number is selling you a story. The stack, briefly GitHub REST API v3 (search/repositories, stats/commit_activity, contributors), a weekly batch pipeline that processes all 4,200 orgs for under €50/month of compute, and a scoring layer. No scraping, no private data, nothing you couldn't rebuild yourself — the signal computation logic is being open-sourced (MIT). There's also a free MCP server, so if you use Claude Desktop or Cursor you can query the live signal data directly from your editor: six tools, no paywall on any of them. Where this goes I publish the weekly top movers in a free Sunday email at gitdealflow.com. The heavier stuff (full rankings, sector sweeps, dashboards) is paid, which is what funds the compute. But the core idea is free and reproducible: public engineering activity is a leading indicator of private funding events. If you're an angel, a scout, or just someone who likes watching startups through their commits, the data is sitting there in the open. Happy to answer questions about the methodology, the backtest design, or the pipeline in the comments. — The Data Nerd