We have spent ten chapters learning where data comes from. Now we use it to answer the central question of statistics: what is the value of some unknown population quantity? The realistic answer comes in two parts, a best guess and a measure of how unsure we are.
We move from describing data to inferring the unknown population behind it. This chapter sets up the idea; the next chapters build the actual confidence intervals for means, proportions, and differences, and end with the bootstrap.
The Point Estimate
A point estimate uses a sample statistic to stand in for an unknown population parameter: the sample mean x̄ estimates μ, the sample proportion p̂ estimates p. It is the natural best guess, but it is a single number that gives no hint of how far off it might be.
Good point estimators share three properties from estimation theory: they are unbiased (centered on the truth), consistent (converge as n grows), and efficient (low variance). The sample mean is all three. But even the best point estimate is one draw from a process that could have landed elsewhere.
Estimates Wobble: the Standard Error
An estimator is itself a random variable: a fresh sample yields a different point estimate. The standard deviation of those estimates is the standard error, SE = σ/√n, and it is the raw material of every interval.
In the notebook, 5,000 samples of 40 give point estimates centered exactly on the true mean of 100, with a spread of 2.84, matching σ/√n to the decimal. A single point estimate is one pick from this bell, so it is almost never exactly right. The width of the bell tells us how to size an interval.
The Interval Estimate
An interval estimate is the point estimate plus or minus a margin: estimate ± z·SE. A 95% confidence interval is built so that, across many samples, about 95% of the intervals contain the true parameter.
The notebook draws 100 intervals and 96 contain the true mean, close to the promised 95. The few red misses are the cost of 95% (not 100%) confidence. Crucially, "95% confident" describes the procedure over many samples, not a probability about one fixed interval. The interval is wider than a point, and that extra width is its honesty.
It does not mean "there is a 95% probability the true value is in this particular interval." Once computed, the interval either contains μ or it does not. The 95% refers to how often the method succeeds across repeated samples.
Real-World Example: Average Home Price
Estimation is what analysts do every day. Consider a spreadsheet of 220 closed home sales, the kind of export a county assessor or real-estate platform produces. We do not know the true average price of every home in this market, so we estimate it.
Each row is a sale with sale_price, sqft,
bedrooms, neighborhood, and more.
Running the numbers on the real data:
| Quantity | Value | Meaning |
|---|---|---|
| Point estimate (x̄) | $338,159 | single best guess of the market's average price |
| Std dev (s) | $69,417 | spread of individual sale prices |
| Standard error | $4,680 | s/√n, how much x̄ would wobble |
| 95% interval | $328,986 to $347,331 | x̄ ± 1.96·SE, the honest range |
| Median | $335,550 | mean > median signals a right skew |
Reporting only "$338,159" would imply false precision. The interval is the professional answer: the average sale price is very likely between about $329,000 and $347,000. Notice the mean sits above the median, a hint of right skew that will matter when we reach the bootstrap (see Resampling & Simulation), where the median and a distribution-free interval become the better tools.
Point & Interval Estimates in Machine Learning & AI
Every number a model reports is an estimate, and mature ML practice reports the interval, not just the point. A single accuracy figure without an interval is as incomplete as a home price without a range.
| Idea (this chapter) | In ML / AI it becomes | Example |
|---|---|---|
| Point estimate | A single reported metric | "the model is 92% accurate" |
| Standard error | Uncertainty of a metric | SE of accuracy on a test set of size n |
| Interval estimate | Error bars on the leaderboard | "92% ± 1.5%" via a CI or bootstrap |
| Estimator properties | Bias and variance of an estimator | cross-validation reduces variance of the estimate |
A model's reported accuracy, F1, or RMSE is a point estimate computed on a finite test set, so it has a standard error, and a difference between two models can be pure noise if their intervals overlap. Serious benchmarks now report confidence intervals (often via the bootstrap of Resampling & Simulation) rather than a bare number. The discipline is identical to the home-price example: never report a point without the uncertainty around it.
Estimate, point and interval, in Python
The companion notebook builds a population, watches point estimates wobble with standard error
σ/√n, draws 100 confidence intervals to show 95% coverage, and finishes by loading the real
point-vs-interval-estimation--home_sales.xlsx spreadsheet to estimate the average sale price with a 95% interval.
View opens the rendered notebook instantly (no setup). Open in Colab runs &
edits it live in your browser. To run locally, install numpy, pandas, scipy,
matplotlib, and openpyxl (to read the Excel file) and launch jupyter notebook.
🎓 Key Takeaways
- ✓A point estimate (x̄, p̂) is a single best guess of a population parameter; ideally unbiased, consistent, and efficient.
- ✓Estimates wobble sample to sample by the standard error SE = σ/√n.
- ✓An interval estimate = point ± margin (z·SE); a 95% interval is built so the method captures the truth ~95% of the time.
- ✓"95% confident" is a property of the procedure, not a probability about one computed interval.
- ✓Real data: 220 home sales give an average price of $338,159 with a 95% interval of about $329k–$347k; in ML, always report metric intervals, not bare points.
Practice Challenges
Five short challenges, beginner to intermediate. Try them with NumPy and SciPy before checking the solutions.
A point estimate
Draw one sample of 50 from Normal(75, 12) and report the point estimate of the mean.
sample.mean().The standard error
Estimate the standard error two ways: the formula σ/√n, and by simulating many sample means.
Build a 95% interval
From one sample of 50 (σ = 12 known), build a 95% interval as point ± z·SE.
Coverage of the procedure
Build 2,000 such 95% intervals and confirm about 95% contain the true mean.
Real data: mean home price
Load point-vs-interval-estimation--home_sales.xlsx and report the point estimate and 95% interval for the mean sale
price.
pd.read_excel(..., sheet_name="Sales").A fully-worked solutions notebook walks through all five challenges, each verified in code. Try them yourself first, then compare.
Quiz: Test Yourself
Eight quick questions on point and interval estimation. Answer them, hit Check Answers, and keep refining until you score 100%. Your progress is saved, so you can hop back to the chapter and return anytime.
We have the idea of an interval; now we build it precisely. Confidence Intervals for a Mean derives the confidence interval for a mean, the z-interval when σ is known and the t-interval when it is not, and checks the conditions that make it valid.