The Founder-Background Signal Reference: What a Resume Line Predicts

You are looking at a founder's public background and need to know what each line actually predicts before you weight it in a decision. This reference is for early-stage investors, platform and talent partners at funds, and angels. It ties each observable signal - prior exit, serial status, big-company tenure, industry depth, age, education prestige - to a measured base rate from a named study, and names the specific way that signal lies. Jump to the row you need and leave.

The point is not to tell you what to value. It is to stop you over-weighting pedigree and under-weighting the signals that actually move the odds, by putting a number and a caveat on each line.

The signal table: what each background line predicts

Each observable signal carries a base rate and a documented failure mode. Read the row, then read the failure before you weight it. The two strongest numeric signals - prior exit and industry depth - are also the ones most often misread, in opposite directions: exits are over-weighted, depth is under-weighted.

The sources below use different success definitions, so a number means nothing without its definition attached. The VCBench-derived work defines success as an acquisition or IPO above $500M, or raising more than $500M; companies that raised $100K to $4M and did not exit within eight years count as unsuccessful. Gompers and co-authors define success as starting a company that went public. These are not the same event, and their rates are not poolable.

The structural insight behind this table: the strongest resume signal is also the rarest, which is exactly why it is over-weighted. Two exits predict 60% success, but that rule fires on roughly 106 founders in the training data. Because 84% of winners show zero exits, an exit-first screen maximises precision while discarding most of the base of actual successes.

Prior-venture outcome: the single strongest lift, and its two traps

Prior-venture outcome is the strongest single background signal, but only when you resolve the outcome rather than counting ventures. Two independent datasets agree on direction and disagree on magnitude, so read them side by side and never average them.

The columns use different success definitions - a $500M outlier bar versus an IPO bar - so they are not directly poolable. What survives both: prior success predicts, and it predicts more than any other single line on the page. The feature documentation puts the lift concretely: a prior IPO carries 4.7x lift, a prior acquisition 2.7x.

The first trap is reading count instead of outcome. Gompers' failed-prior founders succeed at 20%, barely ahead of first-timers at 18%. The predictive weight sits almost entirely in prior success (30%), not in repetition. A three-time founder with three flameouts is not "experienced" in any sense the data rewards. Tag the outcome of each prior venture, not the number of them.

The second trap is reading the exit flag without reading the exit. A founder whose prior "exit" was a sub-$10M acqui-hire scores the same boolean as a founder who took a company public. Verify exit size and the founder's actual role before you let the flag move your weighting.

Founder age: a batting-average story, not a talent story

Older founders outperform at the top of the distribution, but this is an aggregate effect, not a per-candidate rule. The mean founding age for the one-in-1,000 fastest-growing ventures is 45.0, against 41.9 for all ventures. A 50-year-old founder is 1.8 times more likely than a 30-year-old to achieve upper-tail growth.

Age rises about 3.1 years from the all-ventures mean to the top 0.1%. The study drew on U.S. Census data covering 2.7 million founders between 2007 and 2014, so this is a population fact, not a sample artefact.

The failure mode is reading this gradient as causation. The 45.0 mean reflects base rates - more middle-aged people both start companies and convert better - not a youth penalty applied to any individual. Auto-discounting a 28-year-old domain expert because of this number destroys signal that only exists in aggregate. Use age as a population prior when you model deal flow, never as a filter on a single founder in front of you.

Age is a batting-average story. A per-founder age filter destroys signal that only exists in the aggregate.

Industry depth: the cheapest high-yield signal

Three or more years in the exact industry of the startup roughly doubles the odds of a one-in-1,000 fastest-growing company. This is the best use of your diligence minutes: it is easy to verify from a public timeline and hard to fake in substance.

Depth outranks pedigree on cost-adjusted yield. Parsing which school a founder attended tells you little; counting relevant years tells you a lot, and both take about the same time to check. Prior experience in the specific industry predicts much greater rates of entrepreneurial success, and the threshold is concrete at three years.

The trap is that titles lie about relevance. "Ten years in fintech" can mean a decade in fintech sales, which is adjacent, not product-domain depth. A founder solving a payments-infrastructure problem needs years near that problem, not years in the sector's org chart. Confirm the three-plus years sit in roles relevant to the startup's actual problem, not merely in the same industry.

This is where a timeline-first search beats a keyword screen. You cannot find "three relevant years in payments before founding" with a title filter, because the relevant roles are described a dozen different ways.

Refolk reads the career timeline the way this section asks you to, so you can ask for earned industry depth in plain English and get back founders whose relevant years are already counted.

Education prestige: the signal that inverts under diligence

Education prestige read alone is noise, and worse than noise when it masquerades as signal. In the feature-engineering analysis, misclassified failures carry a higher average education prestige tier (3.10) than true positives (2.78). Elite degrees cluster in the funded-but-failing.

The mechanism is clean: pedigree opens the door to capital, which raises the population of prestigious founders who get funded and then fail. Prestige correlates with access, not with building. It predicts only in combination with role relevance, which is why the feature taxonomy pairs them in a single tier called "Education x relevance."

The operational rule is to force prestige and relevance into separate columns and never fuse them. The false positive this prevents is scoring a Stanford MBA with no domain role above a ten-year industry operator. Prestige is allowed to raise your weight only when relevance is already present.

Personality inference: a weak prior, never evidence

Any trait you infer from a founder's public footprint is a hypothesis, not evidence, and it should never be load-bearing in a memo. The honest ceiling is the meta-analytic correlation range: digital footprints correlate with Big Five traits between 0.29 for agreeableness and 0.40 for extraversion.

You will see higher headline numbers. A 2024 Scientific Reports study reached about 85% classification accuracy from activity data and 82.2% for predicting Big Five traits from Facebook profile characteristics alone. Those are not the number to quote in diligence. Classification accuracy on a balanced task is not the same as a correlation strong enough to carry a per-founder judgement, and the meta-analytic correlation is the conservative, defensible limit.

The failure mode is labelling a founder "low conscientiousness" from sparse public posts and letting that label move the decision. Correlations capping at 0.40 cannot support that. Require behavioural evidence - shipped products, resolved conflicts, documented decisions - before any trait claim earns weight.

The serial-founder signal is nearly unobservable from self-report

Absence of the phrase "serial entrepreneur" on a profile proves nothing, because almost nobody writes it. In Refolk's index, only 1,536 of 725,093 US Founder and Co-Founder profiles self-tag as "serial entrepreneur."

That is roughly one in 470 US founder profiles. A keyword screen for serial status therefore misses almost every serial founder who exists. The status has to be reconstructed from the career timeline - two or more prior founding roles with dated tenures - not read off a label.

This is the same lesson as the industry-depth section, applied to a different field: the signals that matter are reconstructed from dated history, not lifted from self-description. Rebuild serial status from founding roles, and resolve each prior venture's outcome, because only prior success carries the lift.

The read procedure: from observable record to weighted memo

Work the signals in a fixed order so no line gets weighted before its failure mode is applied. The procedure below moves from an analyst assembling the raw timeline to a partner converting each tag into a base rate discounted for how it lies. One order dispute is worth naming: some practitioners weight industry fit before exit history, but the data supports exit history as the stronger single lift, so resolve outcome first and let depth confirm.

How this goes wrong: the failure modes

Most background-read errors are the same handful of mistakes, and each has a concrete check. These are the most valuable rows in this reference, because each one is a false positive that looks like diligence done well.

The thread running through all seven: each error substitutes something easy to observe for the thing that actually predicts. An exit flag for exit quality. A degree for domain work. A keyword for a resolved career history. A sample statistic for a per-founder judgement. The check in every row forces you back to the substance the easy signal is standing in for.

Keeping this reference current

The base rates here are stable because they come from large, published datasets, but two things drift and are worth re-checking locally. First, the success definitions: when you read a new study claiming a founder-success rate, find its definition of success before you compare it to anything on this page, because a $500M-exit bar and an IPO bar produce different numbers for the same founder type. Second, your own screens: audit whether your pipeline is quietly selecting on exits or prestige, since both feel like rigor while costing you recall.

The durable move is to build the read from the timeline, not from labels. Prior-venture outcome, industry depth, and serial status are all reconstructable from dated history that is hard to fake, and all three are nearly invisible in self-description. When a new signal appears in a study, slot it into the table only once you can state its base rate, its source's success definition, and the specific way it misleads. A row without all three does not belong in a reference a practitioner keeps open.