Cross-Checking Top-Down Against Bottom-Up

You're asked to estimate how many coffee shops there are in Manhattan. A bottom-up chain, population, coffee drinkers per person, cups per day, customers per shop, gives you one number. But how do you know it's not off by 10×? A single chain of estimates, however careful, has no built-in way to catch a bad assumption baked into step three. The fix isn't to make the chain longer or more careful, it's to build a completely different chain from a different starting point and see if the two answers agree.

The idea: two independent derivations, one sanity check

Bottom-up starts from small, personal-scale numbers and multiplies up: how many cups you drink, times people like you, times shops needed to serve them. Top-down starts from a big aggregate you might already know or can estimate independently, total retail square footage, total coffee industry revenue, commercial rent per block, and divides down to the same target quantity. The two approaches use almost entirely different input numbers, so if they agree within roughly a factor of 2–3, that's real evidence neither chain has a badly wrong assumption buried in it. If they disagree by 10× or more, at least one of the input assumptions is off, and figuring out which one is often more informative than either estimate alone.

Worked example 1: coffee shops in Manhattan

Bottom-up: Manhattan population $\approx 1.6$ million; say 40% are regular coffee-shop buyers, buying roughly twice a week $\Rightarrow$ about $1.6\text{M}\times0.4\times2\approx1.28$ million shop-visits per week. A shop serving, say, 300 customers a day (2,100/week) implies roughly $1{,}280{,}000/2{,}100 \approx 610$ shops.

Top-down: Manhattan has roughly 23 square miles of land, a large fraction of it dense retail corridors; if there's on average one coffee shop per 2–3 city blocks across the borough's walkable grid (a rough density estimate), and Manhattan has on the order of 3,000 blocks, that's roughly $3{,}000/2.5 \approx 1{,}200$ shops.

Both land somewhere in the hundreds-to-low-thousands range, 610 vs 1,200 is within a factor of 2, which for a Fermi estimate counts as strong agreement. (The real figure, from public directories, is in the 1,500–2,000 range including chains and independents, both estimates were in the right order of magnitude, with the top-down one closer.)

Worked example 2: catching a bad assumption

Estimating a food-delivery app's daily order volume in a city bottom-up (population × adoption rate × orders/week) gives 200,000 orders/day. A top-down check, restaurant count in the city × average delivery orders per restaurant per day, a number an interviewer might expect you to guess at ~15–20, gives only 30,000. A 6–7× gap is too large to shrug off: re-examining the bottom-up chain, the adoption rate assumption ("40% of adults order delivery at least weekly") is probably too high for anywhere outside a handful of dense markets. The top-down number, anchored in something more physically constrained (restaurants can only fulfill so many orders), is likely closer to right, and the discrepancy itself is the useful finding, not just the final number.