Kernel Density Estimation

You have a year of daily returns for a stock and want to know the shape of the distribution they came from — is it roughly normal, or does it have a fat left tail from crash risk? A histogram is the obvious first move: chop the range into bins, count how many returns fall in each, draw bars. But histograms are jumpy and arbitrary. Shift the bin edges by half a bin width and the picture changes — a smooth bell shape can suddenly look bimodal, or a real second hump can vanish, purely because of where you happened to draw the lines. You want a picture of the underlying distribution that doesn't depend on an arbitrary grid.

An analogy: streetlights on a foggy road

Imagine each data point is a streetlight on a dark road, and each streetlight casts a small pool of light around itself rather than a single pinprick. Where streetlights cluster close together, their pools of light overlap and add up, making that stretch of road brighter — a visual signal that lights (data) are dense there. Where streetlights are sparse, the road stays mostly dark. If you stood back and looked at the total brightness along the road, you'd get a smooth glow that's bright where lights cluster and dim where they don't, with no sharp edges anywhere — unlike a histogram, which would report a count for each fixed slab of road regardless of where inside it the lights actually sit.

Kernel density estimation is exactly this: replace each data point with a small smooth "bump" of probability, and add all the bumps together to get one smooth curve.

The estimator, one symbol at a time

Let $x_1, \dots, x_n$ be your $n$ observed returns. Let $K$ be the kernel — a smooth, symmetric bump function centered at zero, most commonly the standard normal density, which integrates to 1 and is highest at its center. Let $h$, the bandwidth, control how wide each bump is spread.

The estimated density at any point $x$ is

$$\hat{f}(x) = \frac{1}{n h} \sum_{i=1}^{n} K\!\left( \frac{x - x_i}{h} \right).$$

In plain English: for every observation $x_i$, place a bump of width $h$ centered at $x_i$; evaluate how tall that bump is at the point $x$ you care about; add up the heights from all $n$ bumps; divide by $nh$ so the whole curve still integrates to 1, i.e., is still a proper probability density.

The bandwidth $h$ is the single knob that matters. A tiny $h$ makes each bump nearly a spike sitting right on top of its own data point — the resulting curve is jagged and shows every quirk of the sample, including pure noise (this is undersmoothing). A huge $h$ spreads every bump so wide that they all blur into one another — the resulting curve is a featureless mound that hides real structure, like a second mode (this is oversmoothing). Choosing $h$ is a bias–variance trade-off: small $h$ has low bias but high variance, large $h$ has low variance but high bias.

Worked example 1: five points by hand

Take five daily returns (in %): $\{-2, -1, 0, 0.5, 3\}$. Use a simple triangular kernel, $K(u) = \max(1 - |u|, 0)$, with bandwidth $h = 1$. Evaluate the density estimate at $x = 0$:

$$\hat{f}(0) = \frac{1}{5 \cdot 1}\Big[ K(2) + K(1) + K(0) + K(-0.5) + K(-3) \Big].$$

Compute each term: $K(2) = \max(1-2,0)=0$ (too far, no contribution). $K(1) = \max(1-1,0)=0$ (exactly at the edge, zero). $K(0) = \max(1-0,0)=1$ (the point sitting exactly at 0 contributes fully). $K(-0.5) \to K(|-0.5|)=\max(1-0.5,0)=0.5$. $K(-3) = 0$ (far outside the bump's width). Sum: $0+0+1+0.5+0 = 1.5$. So

$$\hat{f}(0) = \frac{1.5}{5} = 0.3 .$$

Only the two returns near 0 (the point at 0 itself, and the point at 0.5) contribute — the return at $-2$ and the outlier at $3$ are too far away for their bumps (width $h=1$) to reach $x=0$ at all. This is the mechanism by which KDE stays local: distant points simply don't influence the estimate far from where they sit.

Worked example 2: bandwidth changes the picture

Using the same five returns, compare the estimated density at $x=1$ under two bandwidths with the triangular kernel. With $h=0.5$: only points within distance 0.5 of $x=1$ contribute, which is just the return at 0.5 (distance 0.5, right at the kernel's edge, contributing 0), giving $\hat f(1) \approx 0$ — the curve reads essentially zero probability at $x=1$ despite two nearby points at 0 and 0.5. With $h=2$: the kernel reaches from $x=-1$ to $x=3$, now picking up contributions from $-1$ (distance 2, edge, 0), $0$ (distance 1, $K=0.5$), $0.5$ (distance 0.5, $K=0.75$), and $3$ (distance 2, edge, 0). Sum $= 0.5+0.75=1.25$, so $\hat f(1) = 1.25/(5 \times 2) = 0.125$. The wider bandwidth pulls in more neighbors and produces a noticeably different, smoother reading at the exact same point — which is why every KDE plot should be reported with its bandwidth, the same way a histogram should be reported with its bin width.

What this means in practice

Quants use KDE to look at the actual shape of a return distribution before assuming normality — it's a natural first check before fitting a parametric model, and often reveals fat tails, skew, or a second mode (e.g., a mixture of calm-regime and stressed-regime days) that a normal-distribution assumption would hide. It's also used to visualize the distribution of a signal's information coefficient across many backtests, or the distribution of slippage across trades. The main practical decision is the bandwidth; common defaults (Silverman's rule of thumb, cross-validated bandwidth selection) exist precisely because eyeballing $h$ is unreliable. KDE says nothing about what happens beyond the range of your observed data — it's a smoothed picture of what you saw, not an extrapolation.