The Delta Method

You estimate something from data and, being careful, you attach an error bar to it. Then someone asks for a different quantity built from the first: you measured a variance and they want a volatility; you measured a hit rate and they want the odds; you measured a mean and a variance and they want a Sharpe ratio. Computing the new number is trivial — just apply the formula. Computing its error bar is not, because the standard error does not travel through the formula unchanged. The delta method is the shortcut that gets it right, and it is short enough to do in your head.

An analogy: uncertainty on a ramp

Imagine standing on a hillside, unsure of your exact position along the ground. Say you know where you are to within one metre. How uncertain is your altitude?

It depends entirely on how steep the hill is where you stand. On a gentle slope, one metre of horizontal doubt buys you a few centimetres of altitude doubt. On a cliff face, that same one metre becomes many metres of altitude doubt. The uncertainty is the same; the terrain converts it differently.

That is the whole idea. Your estimate has some wobble. A transformation is a piece of terrain. The wobble that comes out the other side is the wobble that went in, multiplied by the local steepness. Everything below is that sentence made precise.

The rule, one symbol at a time

Let $\theta$ (theta) be the true quantity and $\hat{\theta}$ ("theta-hat") your estimate of it. Let $g$ be the transformation you want to apply — square root, logarithm, reciprocal, whatever — so the new estimate is $g(\hat{\theta})$. Let $g'(\theta)$ ("g-prime") be the derivative of $g$, which is just the slope of the curve at that point. Then, for a large enough sample,

$$\text{SE}\big(g(\hat{\theta})\big) \approx \big|g'(\hat{\theta})\big| \times \text{SE}(\hat{\theta}) .$$

In plain English: the standard error of the transformed estimate is the standard error of the original, multiplied by the steepness of the transformation at the point you estimated. That is the working form, and it is all you need for most tasks.

The formal statement adds the shape of the distribution. If your estimator is asymptotically normal, meaning

$$\sqrt{n}\,(\hat{\theta} - \theta) \xrightarrow{d} N(0, \sigma^2) ,$$

read as "as the sample grows, the scaled estimation error settles into a bell curve with variance $\sigma^2$", then the transformed estimator is asymptotically normal too:

$$\sqrt{n}\,\big(g(\hat{\theta}) - g(\theta)\big) \xrightarrow{d} N\big(0,\ g'(\theta)^2 \sigma^2\big) .$$

In plain English: the transformed estimate is also bell-shaped, centred on the transformed truth, with its spread scaled by the slope — squared, because variances are in squared units.

Why it works is worth one sentence. Over a band as narrow as a standard error, almost any smooth curve is indistinguishable from its tangent line. A straight line does nothing to a bell curve except shift and stretch it. So the transformed estimate stays bell-shaped, stretched by the slope.

The delta method needs the estimate itself to be roughly bell-shaped first. Below, set a skewed parent and drag the sample size upward to watch the distribution of the estimate become normal — that is the input the rule requires.

Worked example 1: from a variance to a volatility

You estimate daily variance from $n = 250$ days and get $\hat{v} = 4.0$, in units of percent-squared. You want the daily volatility, which is $\sqrt{4.0} = 2.0\%$, with an error bar.

Step 1 — the standard error of the input. For roughly normal returns, the standard error of a variance estimate is

$$\text{SE}(\hat{v}) \approx \hat{v}\sqrt{\frac{2}{n}} = 4.0 \times \sqrt{\frac{2}{250}} = 4.0 \times 0.0894 = 0.358 .$$

Step 2 — the slope of the transformation. Here $g(v) = \sqrt{v}$, so $g'(v) = \dfrac{1}{2\sqrt{v}}$. At $\hat{v} = 4.0$:

$$g'(4.0) = \frac{1}{2\sqrt{4.0}} = \frac{1}{4} = 0.25 .$$

Step 3 — multiply.

$$\text{SE}(\hat{\sigma}) \approx 0.25 \times 0.358 = 0.089 .$$

So the daily volatility is $2.00\% \pm 0.09\%$, and a rough 95% interval runs from about $1.82\%$ to $2.18\%$. Sanity check: the textbook result for the standard error of an estimated volatility is $\sigma/\sqrt{2n} = 2.0/\sqrt{500} = 0.089$. Identical. The delta method just re-derived a standard formula in three lines, which is exactly how it earns its keep.

Worked example 2: from a hit rate to odds

A strategy wins on $\hat{p} = 0.30$ of its $n = 200$ trades. Your risk manager wants the odds of winning, $g(p) = p/(1-p)$, with an error bar.

Step 1 — the input. The estimate is $0.30/0.70 = 0.4286$, and the standard error of a proportion is

$$\text{SE}(\hat{p}) = \sqrt{\frac{\hat{p}(1-\hat{p})}{n}} = \sqrt{\frac{0.30 \times 0.70}{200}} = \sqrt{0.00105} = 0.0324 .$$

Step 2 — the slope. The derivative of $p/(1-p)$ is $\dfrac{1}{(1-p)^2}$, so

$$g'(0.30) = \frac{1}{0.70^2} = \frac{1}{0.49} = 2.041 .$$

Step 3 — multiply.

$$\text{SE}(\text{odds}) \approx 2.041 \times 0.0324 = 0.066 .$$

The odds are $0.43 \pm 0.07$. Notice the amplification: a 3.2-point uncertainty in the probability became a 6.6-point uncertainty in the odds, because the odds transformation is steep here. Push the hit rate up to $0.90$ and the slope becomes $1/0.01 = 100$, so the same sampling noise would be blown up a hundredfold. That is not a flaw in the method; it is a true statement about how badly determined odds are near the extremes.

What this means in practice

• Sharpe ratio error bars. The rule of thumb $\text{SE}(\widehat{SR}) \approx \sqrt{(1 + \widehat{SR}^2/2)/n}$ is a delta-method calculation applied to the ratio of a mean to a standard deviation. It is why a Sharpe of 1.0 from one year of daily data has an error bar of roughly $\pm 0.2$.
• Greeks and risk numbers. Any risk figure computed from estimated parameters — a duration, a delta, a VaR — inherits their uncertainty scaled by a derivative. The delta method is the audit trail for that inheritance.
• When to skip it. If the slope is near zero, or the transformation is sharply curved over the width of your error bar, the linear approximation fails and you should bootstrap instead. A bootstrap costs compute; the delta method costs a derivative.
• Build the interval on the natural scale. Compute the confidence interval for the raw parameter, then push its two endpoints through $g$. That respects the shape of the transformation instead of forcing a symmetric interval onto a skewed quantity.