The Jackknife

You've computed a statistic — a correlation, a Sharpe ratio, a regression coefficient — and you want to know two things: is this estimate systematically biased, and how much would it wobble if you'd collected a slightly different sample? You don't have a formula for either, and simulating from an assumed distribution feels circular, since you don't fully trust the distributional assumption either. The jackknife answers both questions using nothing but the one dataset you already have: recompute the statistic with each observation removed, one at a time, and read the bias and spread off of how the answers move.

An analogy: testing a bridge by removing one support at a time

An engineer wants to know how much any single support beam matters to a bridge's overall stability, and whether the bridge's design has some hidden systematic weakness. One way to find out: remove each support beam, one at a time, and measure how much the bridge sags each time. If removing any single beam causes roughly the same, small amount of sag, the bridge is stable and no beam is doing outsized work — you'd also expect the "with all beams" measurement to be reasonably trustworthy. If removing one particular beam causes a huge sag, that beam — and by extension your original "all beams" measurement — was hiding an outsized dependency. The jackknife applies the same "remove one, see what changes" logic to a dataset instead of a bridge.

The procedure, one symbol at a time

Let $\hat\theta$ be your statistic computed from the full sample of $n$ observations. Define $\hat\theta_{(i)}$ as the same statistic recomputed after deleting observation $i$ — so you compute $n$ "leave-one-out" versions of your statistic, one for each data point removed. The jackknife estimate of bias is

$$\widehat{\text{bias}}_{jack} = (n-1)\left(\bar\theta_{(\cdot)} - \hat\theta\right), \quad \text{where } \bar\theta_{(\cdot)} = \frac{1}{n}\sum_{i=1}^n \hat\theta_{(i)}$$

Read $\bar\theta_{(\cdot)}$ ("theta-dot-bar") as simply the average of all $n$ leave-one-out estimates. In plain English: compare the average of the leave-one-out estimates to your original full-sample estimate; if they differ, that difference — scaled up by $(n-1)$ — is the jackknife's estimate of how much bias is baked into $\hat\theta$. The jackknife estimate of the standard error is

$$\widehat{\text{SE}}_{jack} = \sqrt{\frac{n-1}{n}\sum_{i=1}^n \left(\hat\theta_{(i)} - \bar\theta_{(\cdot)}\right)^2}$$

In plain English: measure how much the leave-one-out estimates spread out around their own average, and that spread — with a specific scaling factor $\frac{n-1}{n}$ built in to correct for the fact that leave-one-out samples overlap heavily — is your estimate of how much $\hat\theta$ itself would wobble from sample to sample.

The explorer above shows a running estimate settling as more data arrives — the jackknife asks a related but different question about the same fixed dataset: not "what happens as $n$ grows," but "how much does the answer move if I quietly remove one observation at a time from the $n$ I already have?" Both are ways of probing an estimator's stability, from two different angles.

The explorer above shows how much a statistic bounces around across genuinely different samples — the spread you see there is the thing the jackknife is trying to estimate using only one sample, by simulating "slightly different data" through deletion instead of through drawing a fresh sample entirely.

Worked example 1: jackknife standard error of a sample mean

Five daily returns, in percent: $1, 2, 3, 2, 2$. Full-sample mean: $\hat\theta = \frac{1+2+3+2+2}{5} = \frac{10}{5}=2.0$.

Leave-one-out means: drop the $1$: $\frac{2+3+2+2}{4}=2.25$. Drop a $2$ (first one): $\frac{1+3+2+2}{4}=2.0$. Drop the $3$: $\frac{1+2+2+2}{4}=1.75$. Drop a $2$ (second one): $\frac{1+2+3+2}{4}=2.0$. Drop a $2$ (third one): $\frac{1+2+3+2}{4}=2.0$.

So the five leave-one-out values are $2.25, 2.0, 1.75, 2.0, 2.0$. Their average: $\bar\theta_{(\cdot)} = \frac{2.25+2.0+1.75+2.0+2.0}{5}=\frac{10.0}{5}=2.0$ — matching $\hat\theta$ exactly, so jackknife bias $=(5-1)(2.0-2.0)=0$, correctly signalling the sample mean is unbiased.

Sum of squared deviations from $2.0$: $0.25^2+0+0.25^2+0+0=0.0625+0.0625=0.125$. $\widehat{\text{SE}}_{jack} = \sqrt{\frac{4}{5}\times0.125} = \sqrt{0.1}=0.316$. Compare to the ordinary textbook standard error of the mean, $s/\sqrt n$: sample standard deviation of the five points is $s\approx0.707$, giving $0.707/\sqrt5=0.316$ — an exact match, confirming the jackknife recovers the known formula here without ever using it directly.

Worked example 2: jackknife catching bias in a ratio estimator

Suppose you estimate a strategy's win/loss ratio from 4 days of P&L: wins of $3, 5$ and losses of $-2, -1$ (in $thousands). Define $\hat\theta = \dfrac{\text{sum of wins}}{|\text{sum of losses}|} = \dfrac{8}{3}=2.667$.

Leave-one-out: drop the win of $3$: wins sum $5$, losses sum $3$, ratio $=1.667$. Drop the win of $5$: wins sum $3$, losses sum $3$, ratio $=1.0$. Drop the loss of $-2$: wins sum $8$, losses sum $1$, ratio $=8.0$. Drop the loss of $-1$: wins sum $8$, losses sum $2$, ratio $=4.0$.

Average of leave-one-out ratios: $\bar\theta_{(\cdot)}=\frac{1.667+1.0+8.0+4.0}{4}=\frac{14.667}{4}=3.667$.

Jackknife bias: $(4-1)(3.667-2.667)=3\times1.0=3.0$ — a substantial estimated bias relative to the point estimate of $2.667$ itself, flagging that ratio estimators built from a handful of observations (especially with a small denominator like the loss sum here) can be badly biased, exactly the kind of warning sign the jackknife is built to surface without any distributional assumption about P&L.

What this means in practice

• Ratio and other nonlinear statistics (Sharpe ratios, win/loss ratios, correlation coefficients) are the cases where the jackknife earns its keep most, since these are exactly the statistics without simple, closed-form bias or standard-error formulas.
• With small samples, the jackknife's bias and standard-error estimates are themselves noisy — computing $n$ leave-one-out values from a small $n$ doesn't manufacture new information, it only re-examines the information already present, so the jackknife is a diagnostic tool, not a way to make a small sample behave like a large one.
• The bootstrap is the jackknife's more flexible cousin — instead of leaving out exactly one observation at a time, it resamples with replacement many times, generally giving better standard-error estimates for statistics that respond in a jumpy, non-smooth way to individual data points, where the jackknife's smoother leave-one-out approach can misbehave.