Back in lesson 2 you met overprecision — the quieter, nastier face of overconfidence. Not “my guess is too high,” but “my error bars are too narrow.” You watched the famous 90% interval test catch the truth only about half the time: people draw ranges so tight that reality slips out the side. This lesson is the antidote, and it is gloriously practical. Overprecision is a habit, and habits have cures. The cure is learning to think — and to answer — in ranges.
You already learned in Thinking in Probabilities that a range beats a single guess. Here we go one level deeper: not just “use a range,” but how to calibrate its width so the range tells the honest truth about what you don’t know.
Before you read — take a guess
You're asked for next quarter's revenue. Which answer is the most honest reflection of what you actually know?
A point estimate is a confident lie
When someone asks “how long will this take?” and you say “three weeks,” you have just told a small, confident lie. Not because you’re dishonest — because the true answer was never a single number. It was a spread of possible outcomes, and you flattened it into a point to sound decisive. A point estimate is one number with no width. It looks crisp and authoritative, and it is wrong essentially every time: the project takes 19 days, or 26, or 41, but never exactly the 21 you named.
The honest alternative is a confidence interval (also called an error bar): a range together with a stated probability that the true value falls inside it. “80% sure: 3 to 5 weeks” is a confidence interval. It carries two pieces of information the point estimate threw away — how wide your uncertainty is, and how sure you are about that width.
Think of it like a weather forecast. “It will be 22 degrees tomorrow” is a point estimate, and it will be wrong. “18 to 25, and I’d be surprised to be outside that” is an error bar, and it can be right. The forecast didn’t get vaguer — it got honest.
The one-line definition
A confidence interval is a range plus a probability: “I’m C% sure the true value is between L and H.” The number C is a promise you can be graded on, exactly like the probabilities in lesson 1 — over many such intervals, a calibrated 80% interval should contain the truth about 80% of the time.
When to use it
Reach for an interval the moment a single number would be acted on as if it were certain — budgets, deadlines, forecasts, engineering loads, medical prognoses. The test: if I’m wrong about this number, does it cost something? If yes, the width of your uncertainty is part of the answer, and hiding it inside a point estimate just hides the risk. Keep the point estimate for genuinely trivial, low-stakes guesses where nobody will build on the number.
What exactly distinguishes a confidence interval from a plain point estimate?
Your first interval is always too narrow
Here is the universal bug, and naming it is half the fix: the first interval you write down is too narrow. Overprecision isn’t a flaw in unusual people; it’s the default setting of the human estimator. Your brain anchors on the number that comes to mind, sprinkles a little width around it to feel humble, and stops — far too soon.
So the cure is mechanical: draw your interval, then deliberately widen it. The single best widening move is to ask one question — “what could put the answer outside this range?” — and keep stretching the bounds until you’d be genuinely surprised to be wrong, not merely a little uncomfortable.
Let’s estimate something nobody memorized: the weight of an adult blue whale.
| Stage | Low | High | Reasoning |
|---|---|---|---|
| First instinct (too narrow) | 50 tons | 80 tons | ”Big animal, biggish number.” |
| Ask: what’s outside this? | — | — | Calves are far lighter; record bulls are far heavier; my units could be off. |
| Widened interval | 40 tons | 200 tons | Now I’d genuinely be surprised to miss. |
A typical blue whale weighs around 140 tons. The cramped first guess of 50 to 80 misses badly — the truth was never even close to inside it. The widened 40-to-200 interval contains the truth comfortably. Notice what widening cost you: almost nothing. A wider range is mildly less impressive at a dinner party and dramatically more likely to be right. That trade is one you should take nearly every time.
Widening is not cheating
It feels like giving up to stretch “50 to 80” out to “40 to 200.” It isn’t — it’s the opposite of giving up. A narrow interval that misses taught you nothing and lied to whoever trusted it. A wide interval that contains the truth is doing its only job. If widening your range feels like weakness, that feeling is the overprecision talking.
When to use it
Every single time you produce an interval, run the widening pass — it’s free and it’s where the calibration actually happens. It matters most exactly where it feels worst: on questions you feel you should know (the hard-easy effect from lesson 2 means experts get overprecise on their own turf). The more confident the first interval feels, the more it needs the “what’s outside this?” interrogation.
Your honest 90% interval for a quantity feels embarrassingly wide. What does that most likely mean?
The 90% interval discipline
You can’t feel your way to calibration; you have to measure it. Here’s the drill. For a batch of estimates, set every interval to your honest 90% confidence level. The promise baked into “90%” is simple and testable: across the batch, you should miss about 1 in 10. Get ten of them, expect to miss one. If you’re missing four or five out of ten, your ranges are systematically too tight — that’s overprecision, quantified, staring back at you.
A clean trick for building each interval is the absolute-pairs technique: don’t think about the range as a whole. Think about the two bounds separately.
- Pick a low bound you are 95% sure is too low — the truth is almost certainly above it.
- Pick a high bound you are 95% sure is too high — the truth is almost certainly below it.
Stitch those two together and you get an interval that’s roughly 90% likely to contain the answer (about 5% of the probability leaks out each end). Treating the bounds one at a time defeats the brain’s instinct to hug the anchor, because you’re now actively hunting for a number low enough — and a number high enough — to surprise yourself.
Here’s a worked batch. Set each 90% interval, then check it against reality:
| Question | [Low, High] (90%) | Truth | Hit? |
|---|---|---|---|
| Boiling point of water on Everest’s summit (°C) | [60, 90] | 71 | Hit |
| Year the printing press was invented | [1400, 1500] | 1440 | Hit |
| Wingspan of a Boeing 747 (meters) | [40, 80] | 68 | Hit |
| Number of bones in the adult human body | [180, 230] | 206 | Hit |
| Length of the Nile (km) | [5,000, 7,500] | 6,650 | Hit |
| Population of Australia, 2023 (millions) | [15, 22] | 26 | Miss |
Five hits, one miss out of six — a hit rate of about 83%. For a 90% interval that’s a touch low but respectable: on a longer batch you’d want to nudge the ranges a hair wider so the long-run hit rate climbs toward 90%. The Australia miss is the tell — the high bound (22) was anchored too low; a 95%-too-high bound would have sat nearer 30. One miss diagnosed; widen and re-aim.
Because an interval you never miss is uselessly wide. If you set every range to “between zero and infinity,” you’d be right 100% of the time and have said nothing. Calibration is a balancing act: the interval should be as narrow as possible while still catching the truth at the rate you claimed. A 90% interval that genuinely misses 1 in 10 is carrying the maximum useful information consistent with being honest. Missing none means you over-widened; missing four means you under-widened. “About 1 in 10” is the sweet spot the discipline is tuning you toward.
When to use it
Run the full batch-and-score drill periodically as training — ten or twenty estimable facts, scored honestly — to find out which way your ranges are biased and by how much. Use the absolute-pairs technique on every individual interval you build, all the time; it costs ten seconds and it’s the most reliable single defense against overprecision you have.
Ranges don’t have to be symmetric
A quick but important refinement: an honest interval is not always centered on your best guess. Reality is often lopsided, and your range should be too. The downside and the upside can simply differ.
The cleanest example is project timelines. Tasks can finish a little early — but they can run enormously late, because everything that goes wrong adds delay and almost nothing that goes right subtracts it. So a realistic interval for “three weeks” might be 2.5 to 8 weeks: barely any room below, a long tail above. That long right tail is the planning fallacy and fat tails wearing work clothes — estimates skew long for the same structural reason extreme events cluster in the tail. Don’t force a symmetric range onto an asymmetric world; put the width where the surprises actually live.
Why is an honest interval for a project deadline usually lopsided, with a long tail on the high side?
Ranges are the input to margin of safety
Now the payoff — the reason this skill is worth the trouble. Your honest error bar is exactly what tells you how big a buffer to leave. This is the margin of safety model (the subject of its own course), made quantitative. Engineers, investors, and planners don’t size their buffer to the midpoint of an estimate — they size it to the width of the uncertainty around it. A calibrated interval is the number that makes margin of safety a calculation instead of a vibe.
The logic is direct: margin of safety is the gap between what you can withstand and what you expect to face. But “what you expect to face” isn’t a point — it’s a range. The bigger the range, the bigger the buffer you need to stay safe across all of it. A point estimate can’t size a margin of safety at all, because it has no width; only an interval can.
Worked example — a pedestrian bridge. Suppose your point estimate of peak load is 8 tons. Tempting to build for, say, 10 and call it a comfortable 25% margin. But the point estimate hid the real question. Your honest 90% interval for peak load, after widening and thinking about crowd surges, festivals, and the occasional small truck, is 6 to 14 tons.
| Approach | Design load | Margin over… | Problem |
|---|---|---|---|
| Buffer the point estimate | 10 tons | the 8-ton guess | A 13-ton festival crowd collapses it. |
| Buffer the interval’s top | 14 tons, then add margin → ~18 tons | the 14-ton high bound | Sized to the real worst case you’ll plausibly see. |
The first design looks prudent and is secretly fragile: it bought margin against a number you already knew was wrong. The second sizes the buffer to the top of the honest range — it survives the whole interval, not just the midpoint. Same idea sizes a budget: if your costs are “80% sure: 3 to 6 million,” you don’t reserve a buffer on top of 4 — you fund toward the 6, because that’s where your own honest range told you the danger lives. The interval did the hard part; margin of safety just reads the answer off the top end.
The chain in one line
Calibrated probability → honest interval → its width → the margin of safety you need. Get the interval too narrow (overprecision) and every buffer you build on top of it is too small. Overprecision isn’t just an estimation error — it’s a safety error, because it shrinks the very number your buffer is supposed to be sized against.
When to use it
Whenever a decision needs a buffer — a structural load, a cash reserve, a schedule slack, a position size — build the calibrated interval first, then size the margin to its width (and to the worst relevant end of an asymmetric range). If you find yourself adding a buffer to a single point number, stop: you skipped the step that tells you how big the buffer should actually be.
Saying it out loud
One last, quieter skill: communicating the range without flinching. “3 to 6 million, 80% confident” is more useful than a crisp “4.2 million” — it hands the listener your actual uncertainty so they can size their own decisions and buffers. The crisp number feels more competent and informs worse.
The hard part isn’t drawing the range; it’s the social pressure to collapse it. Rooms reward the person who bangs the table with one confident number and quietly distrust “it’s somewhere between 3 and 6.” Resisting that pull — staying honest about width when everyone wants false precision — is its own discipline. It’s the deep tension between sounding decisive and being calibrated, and it’s exactly where the next lesson picks up.
Recap
What is the practical cure for overprecision when you make an estimate?
Check your answer to continue.
Where this goes next
You can now turn a false-precision number into an honest interval, widen it past your overprecise first instinct, score a batch against the 90% rule, and feed its width straight into a margin of safety. What’s left is turning all of this into a standing practice — and confronting the tension we just teased: being honestly uncertain (calibration) while still being decisive and useful (resolution).
Lesson 5, Getting Calibrated, is the training program: tracking your predictions, running calibration drills, post-morteming your scores, and what the superforecasters do differently. It’s where the whole course becomes a habit you keep — including the courage to say “3 to 6 million, 80% confident” out loud and not flinch.