Confidence Intervals

A point estimate gives one number.

A confidence interval gives a range that reflects sampling uncertainty.

For example, saying an accuracy estimate is (82%) is less informative than saying the estimate is around (82%), with a plausible range from (78%) to (86%).

lowerestimateupperrange shows sampling uncertainty
An interval places a range around a point estimate instead of pretending one number is exact.

Working Intuition

A wider interval means the estimate is less precise.

More data usually narrows intervals, though the exact formula depends on the quantity being estimated.

An interval is not a promise about one particular future example. It is a way to report uncertainty in the estimate itself.

In the usual frequentist reading, the interval comes from a procedure. Over many repeated samples, that procedure is designed to cover the true quantity at the stated rate. The interval is not saying that one fixed unknown value is randomly moving around inside this particular range.

The important habit is to read the point estimate and the uncertainty together. An estimate of 82% with a wide interval should not feel as settled as 82% with a narrow interval.

In ML

Confidence intervals are useful when comparing models with close scores. If two scores differ by a tiny amount, the difference may not be meaningful.

They are also useful for honest reporting. A benchmark number without any sense of uncertainty can make a small difference look larger than it is.

MATH-C10-T09-001Exercise: Interval width

A reported interval runs from 78 to 86. What is its width?

Compute it first, then check your number.

Hint

Subtract the lower end from the upper end.

Solution
8678=886 - 78 = 8

The width is 8. A wider interval means the estimate is being reported with less precision.

MATH-C10-T09-002Exercise: Wide or narrow

Which estimate is less precise: an interval from 78 to 86, or an interval from 81 to 83?

Answer it first, then check.

Hint

Less precise means wider.

Solution

The interval from 78 to 86 is less precise because it has width 8. The interval from 81 to 83 has width 2.

MATH-C10-T09-003Exercise: Point estimate only

Does a point estimate by itself show sampling uncertainty?

Answer it first, then check.

Hint

One number has no visible range.

Solution

No. A point estimate gives one number. An interval adds a visible range around that number.

MATH-C10-T09-004Exercise: Tiny difference

Two models report accuracies of 82.0% and 82.3%, but the uncertainty intervals are wide.

Should the small difference be treated carefully before claiming one model is better?

Answer it first, then check.

Hint

The lesson warns about close scores.

Solution

Yes. If uncertainty is wide, a tiny difference in point estimates may not be a meaningful difference in model quality.

MATH-C10-T09-005Exercise: Procedure idea

Enter 1 if a confidence interval is best read as uncertainty from a sampling procedure, not as a promise about one future example.

Compute it first, then check your number.

Hint

The interval surrounds an estimate, not an individual future case.

Solution

Enter 1. A confidence interval reports sampling uncertainty in an estimate. It does not promise what will happen on one future example.

Before Moving On

Prefer estimates with uncertainty attached when a decision depends on small differences.