The Wald Test

You fit a model and get a coefficient — say, a factor's estimated exposure to momentum is 0.12. Is that a real effect, or could it plausibly be zero, with 0.12 just being noise from a finite sample? Every practitioner reaches for the same tool: divide the estimate by its standard error and see if the ratio is "big." That instinct, formalized, is the Wald test — arguably the single most-used piece of statistical machinery in applied work, precisely because it needs almost nothing beyond an estimate and its standard error.

An analogy: how many arm-lengths away is it

Imagine you're trying to tell whether an object sitting some distance from you is genuinely far away, or whether it just looks far away because your depth perception is a bit fuzzy today. The natural question is: how many "units of your own fuzziness" away is it? If the object is ten fuzziness-units away, you're confident it's really there; if it's only half a fuzziness-unit away, you can't tell it apart from being right next to you. The Wald test asks exactly this about an estimate: how many standard errors away from the hypothesized value does it sit?

The rule, one symbol at a time

Let $\hat{\theta}$ be your parameter estimate, $\theta_0$ the value you're testing against (often zero), and $\text{SE}(\hat{\theta})$ its standard error. The Wald statistic is

$$W = \frac{\hat{\theta} - \theta_0}{\text{SE}(\hat{\theta})} .$$

In plain English: take the gap between what you estimated and what the null hypothesis claims, and measure that gap in units of standard error. Under the null hypothesis, and for a large enough sample, $W$ behaves like a standard normal random variable — it should typically land between about $-2$ and $2$. A value far outside that range is evidence the null is wrong. Squaring $W$ gives a statistic that follows a chi-squared distribution with one degree of freedom, which generalizes cleanly to testing several parameters jointly using $W^2 = (\hat{\theta}-\theta_0)^\top \hat{V}^{-1} (\hat{\theta}-\theta_0)$, where $\hat{V}$ is the estimated covariance matrix of the parameters — the multi-parameter version of "how many standard errors away."

Under the null, the Wald statistic behaves like a draw from the curve above — usually landing near zero, rarely landing far out. A $\hat\theta$ that sits many standard errors from $\theta_0$ corresponds to a $W$ deep in the tail, exactly the sort of outcome the curve says should almost never happen if the null were true.

Worked example 1: testing a single regression coefficient

A regression gives a momentum-factor loading of $\hat{\beta} = 0.12$ with standard error $\text{SE}(\hat{\beta}) = 0.04$, from a large sample. Test $H_0: \beta = 0$.

$$W = \frac{0.12 - 0}{0.04} = 3.0 .$$

Under the null, $W \sim N(0,1)$ approximately, so a value of $3.0$ has a two-sided p-value of about $2 \times P(Z>3.0) = 2 \times 0.00135 = 0.0027$. That's well below any conventional 5% threshold, so you reject $H_0$: the loading is unlikely to be pure noise.

Worked example 2: testing two coefficients jointly

Suppose a model estimates two coefficients, $\hat{\beta}_1 = 0.10$ and $\hat{\beta}_2 = -0.05$, with estimated variances $0.0025$ and $0.0016$ and zero covariance between them (a diagonal covariance matrix, for simplicity), and you test $H_0: \beta_1=\beta_2=0$ jointly.

$$W^2 = \frac{0.10^2}{0.0025} + \frac{(-0.05)^2}{0.0016} = \frac{0.01}{0.0025} + \frac{0.0025}{0.0016} = 4.0 + 1.5625 = 5.5625 .$$

Compare $5.5625$ to a chi-squared distribution with 2 degrees of freedom; the 5% critical value there is $5.99$. Since $5.5625 < 5.99$, you fail to reject the joint null at the 5% level — even though $\hat{\beta}_1$ alone might look individually significant, the joint test doesn't clear the bar, illustrating that testing coefficients one at a time and testing them together can give different answers.

What this means in practice

• It's everywhere. Regression t-statistics, GARCH parameter significance, logistic regression coefficient tests — nearly all of them are Wald tests in disguise, because they all reduce to "estimate over standard error."
• Only needs one fitted model. Unlike the likelihood-ratio test, the Wald test doesn't require refitting the model under the null — a real advantage when refitting is expensive.
• Sensitive to the parameterization. Testing $\beta = 0$ versus testing $e^{\beta} = 1$ (an equivalent hypothesis, algebraically) can give a different Wald statistic, because the standard error transforms differently under each parameterization — an artifact the likelihood ratio test doesn't share.