Quantile Functions and the Inverse CDF

Value at risk is a sentence like "there is a 1 percent chance of losing more than $2 million tomorrow." The number $2 million wasn't found by asking "how likely is a $2 million loss?" — that's the cumulative distribution function (CDF), and it runs the wrong direction. It was found by asking "what loss level cuts off the worst 1 percent?" — which is the CDF's inverse, and it is a distinct mathematical object worth naming.

The analogy: a sorted line of outcomes

Imagine every possible outcome sorted from worst to best and lined up along a ruler from 0 to 1, where position 0 is the worst thing that can happen and position 1 the best. The CDF tells you, for any outcome, how far along that ruler it sits. The quantile function does the reverse: point at a spot on the ruler — say, "the worst 1 percent mark" — and it tells you which outcome lives there. VaR is a request to point at 0.01 and read off the outcome.

Writing it down

The CDF $F(x) = P(X \le x)$ is nondecreasing and runs from 0 to 1. Its inverse, the quantile function, is

$$Q(p) = F^{-1}(p) = \inf\{x : F(x) \ge p\}.$$

In words: $Q(p)$ is the smallest value $x$ such that at least a $p$ fraction of the probability sits at or below it. Feed it $p = 0.5$ and you get the median. Feed it $p = 0.01$ and you get the 1st percentile — a 99 percent VaR level, read directly off the loss distribution's left tail.

Slide the mean and standard deviation above and watch the shape change; the quantile function is the same picture read sideways — instead of "how tall is the curve at $x$," it answers "at what $x$ has the area under the curve up to that point reached $p$."

The same object has a second job entirely unrelated to risk reporting: inverse transform sampling. If $U$ is a uniform random number between 0 and 1 — the one thing every computer can generate cheaply and reliably — then $X = Q(U)$ has exactly the distribution you wanted. This is how a random number generator that only knows how to produce uniform numbers gets turned into normal returns, exponential wait times, or any shape a Monte Carlo simulation needs.

Worked example 1: VaR as a quantile, step by step

A portfolio's daily P&L is modeled as $\mathcal{N}(0, 100)$ (variance 100, so a standard deviation of $10 million). The 99 percent VaR is the loss at the 1st percentile: $Q(0.01) = \mu + \sigma \cdot z_{0.01}$, where $z_{0.01} = -2.326$ is the standard normal quantile at $p=0.01$ (a standard, tabulated number). Plugging in: $Q(0.01) = 0 + 10 \times (-2.326) = -23.26$. The 99 percent VaR is $23.26 million — read directly as $|Q(0.01)|$, no separate calculation of "how likely is a $23.26 million loss" required, because the quantile function was built to answer this exact question.

Worked example 2: sampling an exponential from a uniform number

You need to simulate wait times between order fills, exponentially distributed with rate $\lambda = 2$ per second (mean 0.5 seconds). Its CDF is $F(x) = 1 - e^{-\lambda x}$. Solve for the inverse: set $p = 1 - e^{-\lambda x}$, so $e^{-\lambda x} = 1-p$, giving

$$Q(p) = -\frac{\ln(1-p)}{\lambda}.$$

In words: to sample a wait time, draw a uniform number $p$ between 0 and 1 and pass it through this formula. Suppose your random number generator hands you $p = 0.7$: $Q(0.7) = -\ln(0.3)/2 = -(-1.204)/2 = 0.602$ seconds. Draw a different uniform number, $p = 0.2$: $Q(0.2) = -\ln(0.8)/2 = -(-0.223)/2 = 0.112$ seconds. Repeat this thousands of times and the resulting wait times are, provably, exponentially distributed with rate 2 — built entirely from a uniform number generator and one formula.

What this means in practice

• Every headline VaR or CVaR number is a quantile function evaluation. "99 percent VaR" literally means "evaluate $Q$ at $p=0.01$."
• Monte Carlo engines lean on inverse transform sampling whenever the quantile function has a closed form (uniform, exponential, and — via a numerical approximation — normal), because it needs only one uniform draw per sample.
• Copula-based simulation (correlated risk factors with different marginal shapes) works by simulating correlated uniforms first, then passing each through its own quantile function — the technique only works because $Q$ exists for each marginal.