How Errors Compound Down an Estimation Chain

A candidate builds a Fermi chain with four factors, each one they're honestly only confident in to within a factor of 1.5 either way, and confidently states the final product as a single number. That's the most common structural mistake in estimation interviews: treating a chain of rough guesses as if the final multiplication were exact, when in fact each factor's uncertainty compounds with every one after it, and a chain of four "reasonable" 50%-uncertain factors can leave the true answer anywhere from roughly 5x too low to 5x too high.

The idea: multiplicative errors compound multiplicatively

If a quantity is estimated as a product of independent factors, $X = a\times b\times c\times d$, and each factor is uncertain by a multiplicative factor of $(1\pm r)$ — meaning your true value for $a$ could plausibly be anywhere from $a/(1+r)$ to $a\times(1+r)$ — then the combined uncertainty factor across all four multiplies together, roughly:

$$\text{combined factor} \approx (1+r)^n ,$$

for $n$ roughly-equal-uncertainty factors multiplied together. In plain English: uncertainties don't add, they compound, the same way compound interest outpaces simple interest — a chain of several "not too bad, maybe 30-50% off" guesses can multiply into an overall estimate that's honestly only good to within several multiples of the true value, even though no single input felt wildly uncertain on its own.

Worked example 1: four factors at 50% uncertainty each

Suppose you're 50% confident each factor in a four-step chain is within a range of $\times0.67$ to $\times1.5$ of the true value (roughly a 50% swing either way). Multiplying the upper bounds of all four factors together compounds to roughly $1.5^4 \approx 5.1\times$ too high, and the lower bounds compound to roughly $0.67^4\approx 0.2\times$, i.e. 5x too low. So a point estimate built from four honestly-50%-uncertain factors should really be reported as "somewhere in a roughly 25x range" — from a fifth of the estimate to five times it — not as a single confident number. This is why a good Fermi answer states a range, not a point, and why the range should widen visibly with each additional multiplicative step in the chain.

Worked example 2: why cross-checks matter more as chains get longer

Compare a 2-factor chain (market cap $=$ revenue $\times$ multiple) against a 5-factor chain (market cap $=$ users $\times$ revenue/user $\times$ margin $\times$ multiple $\times$ a growth adjustment). Even if each individual factor in the 5-factor chain is estimated more carefully than the factors in the 2-factor chain, the compounding of five uncertain multiplications typically produces a wider total error band than two. This is precisely why cross-checking a long chain against an independent, shorter chain (or a top-down aggregate) is disproportionately valuable the longer your derivation gets — the longer chain is structurally more fragile, not because any one step is worse, but because there are simply more places for error to compound.