Every round is a prediction
Venture is hard to predict. It spans every market and stage, outcomes turn on timing, people and luck, and most of what you can see about a startup is noise. We don’t claim to know who will raise. We’re trying to find the signal in the noise.
No single signal decides anything. Each one is an indicator: a factor that can contribute to a company’s success, positively or negatively. On its own it only tilts the odds. Combined with others and tested against what actually happened, signals become predictive, and we publish how often our predictions come true.
VC Signals turns public facts about startups into a single number: the odds a company closes its next round within 12 months. The numbers are only worth something if they come true, so we use this data to test our probabilities against what actually happened and tune the prediction models every month.
All of it feeds the Venture Decision Model, the venture intelligence model we’re building. Our goal is to give AI the tools to help investors make better venture decisions: signals it can check, evidence it can cite, and a track record that shows when to trust it.
- Create a signal
Anyone can describe something they think predicts a raise, like “net revenue retention above 120%”.
- Agents find the answers
AI agents search the web for each company’s answer and save the evidence and the date they found it. For a private signal, they search the fund’s own data instead.
- Test the factor
We measure whether companies with the signal went on to raise more often than those without, how strongly, and whether the difference is significant.
- Run it back in time
Every answer is dated, so each experiment runs as of a point in time: only what was knowable then, checked against what happened next.
- Record the odds
Every day we save each startup’s odds and the signals behind them. Old records are never edited.
- Wait for the outcome
Twelve months later we check filings, announcements and databases: did the company close a round or not?
- Grade the prediction
We compare predicted odds with what happened. If 70% predictions come true about 70% of the time, the model is calibrated.
- Re-fit the weights
Each signal’s weight is re-estimated from the new outcomes. Signals that stop predicting anything are retired.
No two startups leave the same trail. A stealth pre-seed company and a Series B with a public changelog can’t be measured the same way, so we don’t pretend they can.
Every signal is true, false or unknown. If we couldn’t check something, it counts as unknown, never as a no. Signals that don’t fit a company, like GitHub stars for a biotech, don’t count at all.
Each company starts at the typical odds for its stage and sector. Only what we actually know moves it from there, so a company we know little about stays close to typical instead of looking falsely good or bad.
Odds come with a range and a count of how many applicable signals we know. Fewer known signals means a wider range and lower confidence, and each profile lists which unknowns would tighten it most.
Startups grouped by the odds we gave them a year earlier. Each bar is the share that went on to raise. On the diagonal means the odds meant what they said. Tested on 11,526 predictions from May 2025 to October 2025, a period the model never trained or tuned on.
Weights after the October 2026 re-fit. Weight is the change in log-odds when a signal is true.
- Speaking at a major conference0.00▼ -0.05
- 18+ months since last round-0.46▼ -0.50
- New Form D filed0.55▼ 0.51
- Site outages or broken checkout-0.19▼ -0.22
- Founder went to UC Berkeley-0.09▲ -0.07
- Technical co-founder-0.12▲ -0.09
Ideas that sounded right and did not predict anything. Retired signals stay listed so nobody re-adds them.
Company uses a .ai domain×1.11No difference once sector is accounted for.
Founder has a blue check on X×0.97True for almost everyone; told us nothing.
Office in a co-working space×1.08Too noisy to measure from public data.