The Empirical Distribution Function

You have 500 historical daily returns for a stock and want to know: what fraction of days had a return below -1%? You don't need to assume the returns are normally distributed, fit a model, or estimate any parameters — you can just count. The empirical distribution function does exactly this, systematically, for every possible threshold at once, and it turns out that this simple counting exercise is itself a legitimate estimator of the true, unknown distribution the data came from.

An analogy: ranking your class by test score

Imagine lining up every student in a class by their test score and, for any score $x$ you name, reading off what fraction of the class scored at or below $x$. With no model of "how test scores are distributed" required, this ranking directly answers the question for any $x$ you could ask about. As you add more students to the class, this fraction-below-$x$ becomes a more and more reliable estimate of the fraction you'd see in the entire population of students who could have taken the test, even though you never assumed a bell curve or any other shape to begin with.

The idea, one symbol at a time

Given a sample $X_1, \ldots, X_n$, the empirical distribution function is

$$\hat{F}_n(x) = \frac{1}{n} \sum_{i=1}^{n} \mathbb{1}\{X_i \leq x\},$$

where $\mathbb{1}\{X_i \leq x\}$ is 1 if observation $i$ is at or below $x$ and 0 otherwise. In plain English: $\hat{F}_n(x)$ is simply the fraction of your $n$ data points that fall at or below the value $x$, and as $x$ sweeps from $-\infty$ to $+\infty$, $\hat{F}_n$ traces out a staircase that jumps up by $1/n$ at each observed data point and is flat everywhere else. The Glivenko-Cantelli theorem guarantees that $\hat{F}_n(x)$ converges to the true, unknown CDF $F(x)$ uniformly — meaning the worst-case gap between the staircase and the true curve, over all $x$ simultaneously, shrinks to zero as $n$ grows — not just at each individual point separately, but everywhere at once.

Worked example 1: five daily returns by hand

Suppose five sorted daily returns (in %): $-2.1, -0.5, 0.3, 0.8, 1.5$. The ECDF is a staircase: for $x < -2.1$, $\hat{F}_5(x) = 0$; for $-2.1 \leq x < -0.5$, $\hat{F}_5(x) = 1/5 = 0.2$; for $-0.5 \leq x < 0.3$, $\hat{F}_5(x) = 2/5 = 0.4$; for $0.3 \leq x < 0.8$, $\hat{F}_5(x) = 3/5 = 0.6$; for $0.8 \leq x < 1.5$, $\hat{F}_5(x) = 4/5 = 0.8$; and for $x \geq 1.5$, $\hat{F}_5(x) = 1$. So the answer to "what fraction of days had a return below -1%?" is read directly off the step at $x=-1$: $\hat{F}_5(-1) = 0.2$, since only the $-2.1\%$ observation is below $-1\%$.

Worked example 2: watching convergence

Now suppose the true, unknown distribution is standard normal, and you draw increasingly large samples. With $n=10$, the ECDF's staircase might deviate from the true normal CDF by as much as 0.15 at some point (a coarse, jumpy staircase from few observations). With $n=1{,}000$, that maximum deviation typically shrinks to around $1/\sqrt{1000} \approx 0.03$ — a much finer staircase hugging the smooth curve closely. This shrinking maximum gap, on the order of $1/\sqrt{n}$, is exactly what the Kolmogorov-Smirnov test (see The Kolmogorov-Smirnov Test) later formalizes into a hypothesis test.

Drag the sample size in the explorer above and watch the running estimate settle onto the truth — the same logic applies to the ECDF: watch how the staircase's largest deviation from the true CDF shrinks as more data accumulates, exactly the Glivenko-Cantelli guarantee in action.

What this means in practice

The ECDF is the foundation underneath most of nonparametric statistics: it's what a QQ plot compares against a theoretical distribution, what the Kolmogorov-Smirnov test measures deviations from, and what the bootstrap implicitly resamples from when it treats your data as a stand-in for the whole population. Whenever you plot "percent of days with a drawdown worse than X" or build a historical value-at-risk estimate by reading a percentile off your data, you are using the ECDF, whether or not you name it.