Consistency of Estimators

You have written a formula that turns data into a number: an average return, a volatility, a hedge ratio, a factor loading. Before betting money on that number you want some assurance it is aimed at the right target. The most basic assurance available is this: if I collected more and more data, would this formula home in on the true value, or wander off somewhere else forever? A formula that homes in is consistent. One that does not is broken in a way no amount of data will fix — and plenty of natural-looking formulas are in that category.

An analogy: the camera and the crooked scale

Think of an estimator as a camera photographing a distant sign. With a short exposure the image is grainy and barely readable. Lengthen the exposure — collect more light, which is what collecting more data does — and the grain shrinks. A consistent estimator is a camera whose picture sharpens onto the actual sign.

Two things can go wrong. The picture might sharpen beautifully onto the wrong sign, because the camera is pointed slightly off. Or the grain might never shrink at all, however long the shutter stays open, because the camera keeps sampling the same handful of photons. Consistency rules out both: the picture must sharpen, and sharpen onto the truth.

The definition, one symbol at a time

Write $\theta$ (theta) for the true, unknown quantity — the real average return, the real volatility. Write $\hat{\theta}_n$ ("theta-hat sub n") for your estimate computed from $n$ observations. The hat means "estimated from data", and the subscript is a reminder that the estimate changes as the sample grows. The estimator is consistent if

$$P\big(|\hat{\theta}_n - \theta| > \varepsilon\big) \longrightarrow 0 \quad \text{as } n \to \infty ,$$

for every $\varepsilon > 0$ ("epsilon", any small tolerance you care to name).

In plain English: name any margin of error you like, however tight. With enough data, the probability that your estimate misses by more than that margin falls to zero. Statisticians write this $\hat{\theta}_n \xrightarrow{p} \theta$ and say the estimator converges in probability to $\theta$.

Notice what this is not. It never promises your estimate equals the truth. It promises a shrinking probability of being far off: the estimate stays random forever, the randomness just gets squeezed into a narrower band around the right answer.

The test you actually use

Checking that definition directly is awkward. In practice almost every consistency proof runs through one simple sufficient condition:

$$\text{bias}(\hat{\theta}_n) \to 0 \quad \text{and} \quad \text{Var}(\hat{\theta}_n) \to 0 \ \Longrightarrow \ \hat{\theta}_n \xrightarrow{p} \theta .$$

In plain English: if the estimator's average error shrinks to nothing and its wobble shrinks to nothing, it is consistent. Bias is where it sits on average; variance is how much it bounces from sample to sample. Kill both in the limit and there is nowhere left for the estimate to hide.

Drag the sample size in the explorer below and watch the running average converge on the dashed truth. The envelope narrows like $1/\sqrt{n}$ — that shrinking envelope is consistency, made visible.

Worked example 1: the sample mean of returns

A stock's daily returns have unknown mean $\mu$ and standard deviation $\sigma = 1.2\%$. You estimate $\mu$ with the sample average $\bar{r}$. Is that consistent?

Step 1 — bias. The average of the sample average is the true mean, so the bias is zero at every $n$. Nothing to shrink.

Step 2 — variance. The variance of a sample mean is the population variance divided by $n$:

$$\text{Var}(\bar{r}) = \frac{\sigma^2}{n} = \frac{1.44}{n} ,$$

using $\sigma^2 = 1.2^2 = 1.44$ in units of percent-squared. At $n = 100$ this is $0.0144$, giving a standard error of $\sqrt{0.0144} = 0.12\%$. At $n = 10{,}000$ it is $0.000144$, a standard error of $0.012\%$. It goes to zero. Both conditions hold, so the sample mean is consistent.

Step 3 — put a probability on it. Chebyshev's inequality bounds the chance of missing by more than $\varepsilon$:

$$P(|\bar{r} - \mu| > \varepsilon) \le \frac{\sigma^2}{n\varepsilon^2} .$$

Take $\varepsilon = 0.05\%$. At $n = 100$ the bound is $1.44 / (100 \times 0.0025) = 5.76$, bigger than 1 and so useless. At $n = 10{,}000$ it is $1.44 / 25 = 0.058$; at $n = 1{,}000{,}000$ it is $0.00058$. The chance of missing by more than five hundredths of a percent is being driven to zero — the definition, with numbers in it.

Worked example 2: biased but consistent, and unbiased but not

Now the two cases that separate consistency from its neighbours.

A biased estimator that is still consistent. The maximum-likelihood estimate of variance divides by $n$ rather than $n-1$:

$$\hat{\sigma}^2 = \frac{1}{n}\sum_{i=1}^{n}(r_i - \bar{r})^2 , \qquad E[\hat{\sigma}^2] = \sigma^2\left(1 - \frac{1}{n}\right) .$$

In words: on average this formula lands a little below the true variance, by a factor of $1 - 1/n$. Suppose the true variance is $4$ (a 2% daily volatility). Then:

• $n = 5$: expected value $4 \times 0.8 = 3.2$ — biased low by 20%.
• $n = 100$: $4 \times 0.99 = 3.96$ — low by 1%.
• $n = 10{,}000$: $4 \times 0.9999 = 3.9996$ — low by one part in ten thousand.

The bias is real at every finite $n$, and it goes to zero. Its variance goes to zero too. So this estimator is biased at every sample size yet perfectly consistent.

An unbiased estimator that is not consistent. Estimate the mean return using only the first day of data: $\hat{\mu} = r_1$. Its average is $\mu$, so it is unbiased — at $n = 10$, at $n = 10^9$, always. But its variance is $\sigma^2 = 1.44$ at every $n$; collecting a million more days changes nothing, because the formula throws them away. Feed it into Chebyshev and the bound never shrinks. This estimator is unbiased and useless: the dashed curve in the figure above.

What this means in practice

Consistency is a floor, not a ceiling. It tells you an estimator is not fundamentally aimed at the wrong thing. It says nothing about how fast the convergence happens, and speed is where real quant work lives.

The sharpest illustration is the mean return itself. With an annual volatility of 20%, the standard error of the estimated mean over $T$ years is $20/\sqrt{T}$ percent. To pin the mean down to within 1% you need

$$\frac{20}{\sqrt{T}} = 1 \quad \Longrightarrow \quad T = 400 \text{ years} .$$

The sample mean is consistent — and effectively unusable for expected returns, because the required sample outlives the firm and probably the asset. Variance estimates converge far faster and improve further with intra-day sampling. That asymmetry is why serious portfolio construction leans on covariance estimates and treats expected returns as something you assert rather than measure.

Three more places it bites:

• Misspecification breaks consistency. OLS is consistent for the slope only if the errors are uncorrelated with the regressors. When an omitted variable drives both, the estimate converges cleanly onto the wrong number.
• Overlapping data inflates your effective $n$. Ten thousand overlapping 20-day returns are not ten thousand independent observations, and the variance shrinks like one over the effective sample size, which may be a few hundred.
• A moving target defeats consistency entirely. The framework assumes a fixed $\theta$ waiting to be found. If the parameter drifts with the regime, older data actively hurts and the honest answer is a rolling window, not a bigger sample.