VC Signals: Separating Signal from Noise in Startup Data
How to evaluate VC signals: the three properties of a useful signal (observable early, costly to fake, sector-normalized), the four signal families ranked by lead time, and the three tests that keep noise off your scorecard.
Key Takeaway
Not every moving number is a signal. This guide defines what makes a VC signal useful, lead time, costly-to-fake, and sector-normalized movement, ranks the four signal families from engineering to market data, and gives three tests (dated receipts, base rates, fakeability cost) that separate investable signal from dashboard decoration.
Every investor has signals; few have a filter. The difference decides returns. A signal is any observable that correlates with an outcome you care about, and noise is everything else. In early-stage venture, the outcome is usually fundraise quality or company survival, and the discipline is separating observables that carry information from ones that merely move together with attention.
Signals vs noise: the definitions that matter#
A useful signal has three properties: it is observable before the outcome, it is costly to fake, and it survives sector normalization. Press coverage fails all three, it arrives after the fact, is bought, and means different things in fintech versus dev tools. Engineering activity passes: commit, contributor, and repository trends are public, updated continuously, and expensive to sustainably fake. The leading vs lagging VC signals breakdown formalizes the taxonomy.
The four families of VC signal#
Practically, signals cluster into four families. Engineering signals: GitHub activity, the longest lead time at three to six weeks pre-announcement and the hardest to game. Hiring signals: job postings and team expansion, two to eight weeks. Product signals: launches, changelogs, dependency adoption. Market signals: press, funding-database entries, social spikes, all trailing. The deal flow signal guide ranks the families by lead time and gameability, and the alternative data overview situates them in the broader data landscape.
How to test whether a signal is real#
Three tests. Lead-time test: does the observable move before the outcome, dated receipts, not retrospective stories. Base-rate test: what fraction of the time does the signal fire without the outcome following; a signal that fires without a round half the time still narrows the funnel enormously, but you must know the base rate. Costly-to-fake test: could a founder induce the observable for under 1,000 dollars; if yes, discount it. Signals passing all three earn a place in the scoring framework; the rest are dashboard decoration.
Normalization: why raw numbers mislead#
Raw activity levels are sector-confounded. AI startups commit more than enterprise SaaS at every stage; comparing them unnormalized rewards the wrong companies. Normalize within sector and stage, read the slope not the level, and weight breadth, multiple metrics moving together, over any single series. The engineering velocity benchmarks by stage and sector taxonomy post provide the reference distributions.
Key takeaways#
Signal quality is defined by lead time, fakeability cost, and sector-normalized movement. Engineering signals dominate on all three axes, which is why they anchor most systematic sourcing stacks. Test every candidate signal against dated receipts and base rates before it enters your scorecard, and normalize within sector and stage before comparing any two companies.
A worked example of signal decay#
Signals decay as they get adopted. Job-posting data was a genuine edge for sourcing in the early 2010s; by the time every fund scraped the same boards, the lead time compressed toward zero and the signal became table stakes. The pattern repeats: press mentions, Product Hunt rankings, app-store rank tracking, each was an edge, then a feature, then noise. The implication is not cynicism, it is portfolio thinking about signals themselves: hold several, measure each one's base rate continuously, and expect today's proprietary observable to be next year's commodity.
Engineering signals are earlier in this lifecycle than hiring or web-traffic data, partly because reading them requires more interpretation. Interpretation cost is a moat: the observables are free, but the sector-normalized reading of them is work. That gap between raw access and usable signal is where the current edge lives, and the how-to-read guide exists to close it for you faster than the market closes it for everyone.
Building a personal signal register#
The durable practice is a register: a written list of every observable you use, with for each its lead time, its base rate from your own logged history, its fakeability cost, and its sector caveats. The register turns signal evaluation from vibes into maintenance: when a signal's base rate degrades, you see it in the numbers and demote it. Most investors never write this down, which is why most investors run the same decaying signals for a decade. The scoring framework gives the structure; the register is the habit that keeps it honest.
The base-rate discipline#
When a signal fires, ask: of the last hundred times something like this fired, how many rounds followed within eight weeks. If you have not been logging, start now, because the register's value accrues from your own data, not from industry averages. Six months of honest logging beats any borrowed benchmark, and the companies you logged and passed on are the control group that makes the whole exercise scientific rather than anecdotal.