Six ways estimates go wrong

Most bad estimates fail in a small number of predictable ways. Learn to recognise the six, and your worst misses stop being mysterious.

Zoggle · a guide for people who like numbers

When an estimate is badly wrong, it is rarely because the arithmetic slipped. It is almost always because one assumption was wrong by a lot, and it was wrong for a reason that has a name.

1. Anchoring

The first number you encounter contaminates every number that follows. If someone asks whether a city has more or fewer than 5,000 dentists before asking you to estimate the count, your answer will land closer to 5,000 than it otherwise would — even if the anchor was arbitrary, and even if you know it was arbitrary.

The fix: build your estimate from its components before you let yourself look at any suggested figure. If you’ve already seen an anchor, deliberately construct the calculation from the other end — start from the thing being served rather than the thing serving it.

2. Scale blindness

Human intuition treats million, billion and trillion as roughly the same word, distinguished only by vague largeness. They differ by factors of a thousand.

This is the single most common source of catastrophic misses. Someone estimating a government budget line or a global commodity flow will frequently land three orders of magnitude away, not because their reasoning failed but because their sense of “big” topped out well below the answer.

The fix: convert to something physical. A billion dollars at $50,000 a year is 20,000 salaries. A billion seconds is 32 years. Anchoring an abstract magnitude to a concrete quantity restores the sense of scale your intuition doesn’t supply.

3. Availability

We estimate frequency by how easily examples come to mind, which means vivid, recent and heavily-reported things get systematically overestimated, and quiet common ones get underestimated. People reliably overestimate deaths from dramatic causes and underestimate deaths from ordinary illness, because the news covers one and not the other.

The fix: when estimating how often something happens, ask whether you’re sampling from reality or from coverage. If every instance you can recall came from a headline, your sample is not a sample.

4. False precision

Carrying 3,847,291 through a calculation where every input was a guess produces an answer that looks authoritative and isn’t. Worse, precision is contagious: a tidy-looking output invites confidence the inputs never justified.

The fix: round to one significant figure and stay there. If your inputs are good to a factor of two, your output is good to a factor of two, and writing more digits is a claim you can’t back. Where it matters, state a range instead of a point.

5. Correlated errors

Estimation works because independent errors cancel. When your errors all lean the same way, they compound instead.

This is what happens when you already know what answer you want. Every judgement call goes the favourable direction, and five 1.3× optimisms multiply into a 3.7× fantasy. It also happens innocently: if all your assumptions come from one mental model of the situation, and the model is wrong, everything downstream is wrong together.

The fix: estimate the same quantity by a completely different route and compare. If a bottom-up count and a top-down division disagree by 10×, one of your models is broken — and you’ve just learned something you’d never have discovered by refining a single approach.

6. Unit confusion

The least glamorous and most costly error. Kilometres versus miles. Per month versus per year. Kilowatts versus kilowatt-hours. Per person versus per household. Every one of these is a silent factor between 1.6 and 12, and unit errors don’t announce themselves — the arithmetic works perfectly, it just answers a different question.

The fix: write the units next to every number and cancel them like algebra. If the units of your answer don’t come out as the units of the question, you have a bug regardless of how good the number looks.

The meta-fix

Every one of these is easier to catch in someone else’s work than your own, which suggests the most effective habit: after producing an estimate, spend thirty seconds arguing against it. Where would a critic attack? Which single assumption, if wrong, would move the answer most?

That assumption is where your attention belongs. In most estimates, one input dominates the uncertainty and the rest barely matter — and finding it is more valuable than refining everything else combined.