Mean Squared Error¶

Introduction¶

The Mean Squared Error (MSE) is the most widely used criterion for evaluating the quality of a statistical estimator. It measures the average squared deviation of an estimator from the true parameter value, capturing both systematic error (bias) and random fluctuation (variance) in a single quantity.

MSE serves as the default loss function in estimation theory, regression analysis, and many optimization problems throughout statistics and quantitative finance.

Definition¶

Let \(\hat{\theta}\) be an estimator of parameter \(\theta\). The Mean Squared Error is:

\[\text{MSE}(\hat{\theta}) = E\left[(\hat{\theta} - \theta)^2\right]\]

This is the expected value of the squared difference between the estimator and the true parameter, averaged over all possible samples.

Equivalent Expressions¶

The MSE can be computed in several equivalent ways:

\[\text{MSE}(\hat{\theta}) = E[\hat{\theta}^2] - 2\theta E[\hat{\theta}] + \theta^2\]

\[= \text{Var}(\hat{\theta}) + [E[\hat{\theta}]]^2 - 2\theta E[\hat{\theta}] + \theta^2\]

\[= \text{Var}(\hat{\theta}) + (E[\hat{\theta}] - \theta)^2\]

\[= \text{Var}(\hat{\theta}) + [\text{Bias}(\hat{\theta})]^2\]

This last form is the bias–variance decomposition.

Properties of MSE¶

Non-negativity¶

MSE is always non-negative: \(\text{MSE}(\hat{\theta}) \geq 0\), with equality only if \(\hat{\theta} = \theta\) with probability 1 (the estimator is perfect).

MSE of Unbiased Estimators¶

If \(\hat{\theta}\) is unbiased (\(\text{Bias}(\hat{\theta}) = 0\)), then:

\[\text{MSE}(\hat{\theta}) = \text{Var}(\hat{\theta})\]

For unbiased estimators, MSE and variance are identical. Comparing unbiased estimators by MSE is equivalent to comparing them by variance.

Consistency and MSE¶

An estimator is MSE-consistent if \(\text{MSE}(\hat{\theta}_n) \to 0\) as \(n \to \infty\). By the decomposition, this requires both: - \(\text{Bias}(\hat{\theta}_n) \to 0\) - \(\text{Var}(\hat{\theta}_n) \to 0\)

MSE-consistency implies consistency in probability (convergence in probability to \(\theta\)), by Chebyshev's inequality.

MSE Comparisons Between Estimators¶

Relative Efficiency¶

The relative efficiency of estimator \(\hat{\theta}_1\) compared to \(\hat{\theta}_2\) is:

\[\text{RE}(\hat{\theta}_1, \hat{\theta}_2) = \frac{\text{MSE}(\hat{\theta}_2)}{\text{MSE}(\hat{\theta}_1)}\]

If \(\text{RE} > 1\), then \(\hat{\theta}_1\) is more efficient (lower MSE).

For unbiased estimators, this simplifies to:

\[\text{RE}(\hat{\theta}_1, \hat{\theta}_2) = \frac{\text{Var}(\hat{\theta}_2)}{\text{Var}(\hat{\theta}_1)}\]

Admissibility¶

An estimator \(\hat{\theta}\) is inadmissible under MSE if there exists another estimator \(\hat{\theta}'\) such that:

\[\text{MSE}(\hat{\theta}') \leq \text{MSE}(\hat{\theta}) \quad \text{for all } \theta\]

with strict inequality for at least one \(\theta\). An estimator that is not inadmissible is admissible.

James-Stein result: When estimating a multivariate normal mean \(\mu \in \mathbb{R}^p\) with \(p \geq 3\), the sample mean \(\bar{X}\) is inadmissible — the James-Stein estimator dominates it uniformly in MSE.

Worked Examples¶

Example 1: MSE of the Sample Mean¶

Let \(X_1, \ldots, X_n \sim \text{iid}\) with mean \(\mu\) and variance \(\sigma^2\). The sample mean is \(\bar{X} = \frac{1}{n}\sum X_i\).

Bias: \(E[\bar{X}] = \mu\), so \(\text{Bias}(\bar{X}) = 0\) (unbiased).

Variance: \(\text{Var}(\bar{X}) = \sigma^2/n\).

MSE: \(\text{MSE}(\bar{X}) = 0 + \sigma^2/n = \sigma^2/n\).

Example 2: MSE of the Naive Variance Estimator¶

The naive variance estimator is \(\tilde{S}^2 = \frac{1}{n}\sum_{i=1}^n (X_i - \bar{X})^2\).

For a normal population:

Bias: \(E[\tilde{S}^2] = \frac{n-1}{n}\sigma^2\), so \(\text{Bias}(\tilde{S}^2) = -\sigma^2/n\).

Variance: \(\text{Var}(\tilde{S}^2) = \frac{2(n-1)}{n^2}\sigma^4\).

MSE:

\[\text{MSE}(\tilde{S}^2) = \frac{2(n-1)}{n^2}\sigma^4 + \frac{\sigma^4}{n^2} = \frac{2n-1}{n^2}\sigma^4\]

Example 3: Comparing Biased vs Unbiased Variance Estimators¶

The Bessel-corrected estimator is \(S^2 = \frac{1}{n-1}\sum (X_i - \bar{X})^2\).

For a normal population:

MSE of unbiased \(S^2\): \(\text{MSE}(S^2) = \text{Var}(S^2) = \frac{2\sigma^4}{n-1}\)

MSE of biased \(\tilde{S}^2\): \(\text{MSE}(\tilde{S}^2) = \frac{(2n-1)\sigma^4}{n^2}\)

Compare: \(\frac{2n-1}{n^2}\) vs \(\frac{2}{n-1}\)

Cross-multiplying: \((2n-1)(n-1)\) vs \(2n^2\), i.e., \(2n^2 - 3n + 1\) vs \(2n^2\).

Since \(-3n + 1 < 0\) for \(n > 0\), we have \(\text{MSE}(\tilde{S}^2) < \text{MSE}(S^2)\).

The biased estimator has lower MSE than the unbiased one! This is a concrete illustration of the bias–variance tradeoff. The optimal estimator (minimizing MSE among estimators of the form \(c \cdot \sum(X_i - \bar{X})^2\)) divides by \(n+1\), not \(n\) or \(n-1\).

Example 4: MSE-Optimal Variance Estimator¶

Consider \(\hat{\sigma}^2_c = \frac{1}{c}\sum_{i=1}^n(X_i - \bar{X})^2\) for constant \(c > 0\).

For normal populations:

\[\text{MSE}(\hat{\sigma}^2_c) = \left(\frac{n-1}{c} - 1\right)^2 \sigma^4 + \frac{2(n-1)}{c^2}\sigma^4\]

Differentiating with respect to \(c\) and setting to zero:

\[c^* = n + 1\]

So the MSE-optimal estimator divides by \(n+1\):

\[\hat{\sigma}^2_{n+1} = \frac{1}{n+1}\sum_{i=1}^n (X_i - \bar{X})^2\]

This is biased (underestimates \(\sigma^2\)) but has lower MSE than both \(\tilde{S}^2\) (divide by \(n\)) and \(S^2\) (divide by \(n-1\)).

Connections to Other Loss Functions¶

Mean Absolute Error (MAE)¶

\[\text{MAE}(\hat{\theta}) = E\left[|\hat{\theta} - \theta|\right]\]

MAE is less sensitive to outliers than MSE. However, MSE is mathematically more tractable and directly connects to the bias-variance decomposition.

Risk Function¶

In decision theory, \(\text{MSE}(\hat{\theta})\) is the risk of \(\hat{\theta}\) under squared error loss \(L(\hat{\theta}, \theta) = (\hat{\theta} - \theta)^2\):

\[R(\hat{\theta}, \theta) = E[L(\hat{\theta}, \theta)] = \text{MSE}(\hat{\theta})\]

Cramér-Rao Lower Bound¶

For unbiased estimators, the MSE (= variance) is bounded below by the Cramér-Rao bound:

\[\text{Var}(\hat{\theta}) \geq \frac{1}{I(\theta)}\]

where \(I(\theta) = -E\left[\frac{\partial^2}{\partial\theta^2}\log f(X;\theta)\right]\) is the Fisher information. An unbiased estimator achieving this bound is called efficient.

MSE in Finance¶

MSE appears throughout quantitative finance:

Forecast evaluation: MSE is the standard metric for comparing return, volatility, or risk forecasts. RMSE = \(\sqrt{\text{MSE}}\) puts the error in the same units as the target.
Tracking error: The MSE between a portfolio's returns and its benchmark captures both systematic deviation (bias) and random deviation (variance).
Model calibration: MSE between model-implied and market-observed option prices is the objective function in calibrating volatility models.
Regression: OLS minimizes \(\sum(y_i - \hat{y}_i)^2/n\), the in-sample MSE.

Summary¶

MSE is the fundamental criterion for evaluating estimator quality. Its decomposition into variance and squared bias reveals the inherent tradeoff in estimation and provides a principled framework for choosing between competing estimators. While unbiasedness is desirable, MSE reminds us that the best estimator minimizes total error — and a little bias can be worth a lot of variance reduction.

Key Formulas¶

Quantity	Formula
MSE	\(E[(\hat{\theta} - \theta)^2]\)
Decomposition	\(\text{Var}(\hat{\theta}) + [\text{Bias}(\hat{\theta})]^2\)
MSE of \(\bar{X}\)	\(\sigma^2 / n\)
Relative Efficiency	\(\text{MSE}(\hat{\theta}_2) / \text{MSE}(\hat{\theta}_1)\)
Cramér-Rao Bound	\(\text{Var}(\hat{\theta}) \geq 1/I(\theta)\)