A founder cites a widely repeated statistic — something close to "90% of startups fail" — as a general reference point for thinking about their own venture's odds, without examining what specific population of companies, measured over what time period, using what definition of failure, actually produced that particular number. Different studies measuring different populations under different definitions produce meaningfully different failure rate figures, and applying a generic, memorable statistic to a specific situation without checking whether the underlying reference class actually resembles that situation can produce a genuinely misleading risk assessment in either direction.
Why the specific population studied changes the number considerably
A failure rate calculated specifically from venture-capital-backed technology startups, a small and specific subset of all new businesses generally, describes a genuinely different population — with different funding dynamics, different growth expectations, and often a different risk profile by design — than a failure rate calculated across all new small businesses of every type, including traditional local service businesses that never sought outside investment or aimed for the kind of high-growth trajectory venture-backed startups specifically pursue. These different underlying populations produce different failure statistics for reasons that have nothing to do with either figure being more or less accurate — they're simply describing different things.
Why the definition of failure itself varies meaningfully between studies
Some studies define failure narrowly as complete business closure and liquidation, while others define it more broadly to include any outcome short of the original founders' intended trajectory, including an acquisition at a valuation the founders consider disappointing, or a pivot into an entirely different business — these different definitions produce different failure rates even when applied to comparable underlying populations of companies, which means comparing statistics across studies without checking their specific failure definition risks comparing genuinely incompatible measurements.
Why the specific reference class matters more than the general concept of a base rate
A founder pursuing a bootstrapped, cash-flow-focused local service business is operating under a genuinely different risk profile than a founder pursuing a venture-backed technology startup aiming for rapid scale, and applying either group's specific failure statistics to the other's very different situation produces a base rate that may not usefully describe the actual risk being faced, regardless of how frequently the borrowed statistic gets cited in general startup discourse.
What actually produces a more useful base rate for a specific situation
Identifying the most genuinely comparable reference class available — similar industry sector, similar funding approach, similar growth ambitions and stage — and seeking out failure or success statistics specific to that closer reference class, rather than defaulting to whichever generic statistic happens to be most widely repeated in general startup commentary, produces a meaningfully more useful and more accurate risk assessment for the specific situation actually being evaluated.
What this means for founders trying to reason about their own venture's realistic odds
- Identify the specific population, time period, and failure definition behind any startup failure statistic before applying it to a specific situation
- Seek out failure and success data from the most genuinely comparable reference class available — similar sector, similar funding approach, similar stage — rather than a generic, widely repeated figure
- Be specifically skeptical of a single, memorable statistic used to describe startup risk broadly across genuinely different types of ventures
- Recognize that a well-chosen, genuinely comparable reference class produces a more useful base rate than a technically larger but less relevant general population
A startup failure statistic is only as useful as the match between the population it was actually calculated from and the specific situation it's being applied to — and much of the widely repeated startup failure commentary skips this matching step entirely, treating a specific reference class's number as if it applied universally.