Student's t-Distribution

Motivation: why this matters in quant finance

The Student's $t$ -distribution has two lives. In statistics, it is the distribution behind small-sample inference when variance is estimated. In finance, it is a practical fat-tailed alternative to the normal distribution.

Both lives come from the same construction: a normal shock divided by an uncertain scale estimate. The centre can look Gaussian, but the tails are much heavier.

Definition

Let $Z\sim\mathcal{N}(0,1)$ and $V\sim\chi^2_\nu$ be independent. Then

T=\frac{Z}{\sqrt{V/\nu}}\sim t_\nu.

Its density is

f(t)=\frac{\Gamma((\nu+1)/2)}{\sqrt{\nu\pi}\Gamma(\nu/2)}\left(1+\frac{t^2}{\nu}\right)^{-(\nu+1)/2}.

The parameter $\nu$ is the degrees of freedom. Smaller $\nu$ means heavier tails.

Key Properties

For large $|t|$ , the density decays like a power law:

f(t)\sim C|t|^{-(\nu+1)}.

Moments exist only below the degrees-of-freedom threshold:

\mathbb{E}[T]=0 \quad (\nu>1), \qquad \operatorname{Var}(T)=\frac{\nu}{\nu-2} \quad (\nu>2).

As $\nu\to\infty$ , $t_\nu$ converges to $\mathcal{N}(0,1)$ . For finite $\nu$ , the MGF does not exist for nonzero arguments.

In Quant Finance

A common fat-tailed return model is

r_t=\mu+\sigma_t\varepsilon_t, \qquad \varepsilon_t\sim t_\nu.

With dynamic $\sigma_t$ , this becomes the GARCH- $t$ idea: conditional tails are fat and volatility changes through time.

For one-day VaR, a $t_5$ lower 1% quantile is about $-3.365$ , compared with the Gaussian $-2.326$ . If the $t_5$ variable is rescaled to unit variance, the 1% multiplier is still about $3.365\sqrt{3/5}=2.607$ , larger than Gaussian.

ISL's regression chapter uses the $t$ distribution for coefficient tests and confidence intervals. That matters for strategy backtests: with small samples, normal critical values are too optimistic.

Worked Example: Tail Capital Difference

For a zero-mean portfolio with daily volatility $1\%$ , Gaussian 99% VaR is about $2.326\%$ . A unit-variance $t_5$ model gives about $2.607\%$ . The difference is not a formatting detail; it is extra capital driven by tail shape.

Common Confusions and Pitfalls

The $t$ distribution is only for hypothesis testing. It is also a return innovation and copula building block.

Higher degrees of freedom means fatter tails. The opposite:

\nu\to\infty

gives the normal.

$t_\nu$ has variance 1. Only after rescaling, unless

\nu=\infty

A $t$ innovation handles volatility clustering. It handles conditional tail thickness; clustering needs a dynamic volatility model.

Where This Goes Next

Normal Distribution: the thin-tailed benchmark.
Chi-Squared and Related Distributions: supplies the random denominator in the $t$ construction.
Improper Integrals: explains moment existence through tail integrability.

References

Bertsekas and Tsitsiklis, Introduction to Probability, 2nd ed., Ch. 3 Sec. 3.3 and Ch. 4 Sec. 4.4.
Mosteller, Fifty Challenging Problems in Probability, Problems 27-34, for approximation and tail-probability habits.
James, Witten, Hastie, and Tibshirani, An Introduction to Statistical Learning, 2nd ed., Ch. 3, for $t$ statistics and confidence intervals.