The basic concept
Student’s t-distribution is symmetric around zero and has heavier tails than a standard normal distribution. Its degrees of freedom ν (nu) govern tail thickness. Lower positive degrees of freedom produce more extreme outcomes. As ν grows, the distribution approaches the standard normal.
Two uses need separate explanations: inference about an unknown population mean, and a distributional model for heavy-tailed observations.
Use 1: estimating an unknown mean
For independent, identically distributed normal observations with unknown population variance, the following statistic has a t-distribution with n − 1 degrees of freedom.
Here n is the sample size and s is the sample standard deviation. The denominator is the estimated standard error of the mean. Non-normal data may permit an approximation, but small samples and dependence require care.
Worked example: a confidence interval
Suppose 25 independent observations have a sample mean of 8% and sample standard deviation of 10%. Under the stated normal-sampling assumption:
| Step | Calculation |
|---|---|
| Degrees of freedom | 25 − 1 = 24 |
| Standard error | 10% / √25 = 2% |
| Two-sided 95% critical value | 2.0639 |
| Confidence interval | 8% ± 2.0639 × 2% = [3.87%, 12.13%] |
The interval estimates the population mean. It is not an interval containing 95% of individual observations, and it does not imply a 95% frequentist probability that the fixed unknown mean lies in this particular realised interval.
Use 2: modelling heavy-tailed returns
A location-scale model is X = μ + aT, where T has ν degrees of freedom and a > 0 is a scale parameter. For ν > 2:
With ν = 5 and a = 1%, the standard deviation is about 1.291%. To target a 1% standard deviation instead, set a = 1% × √(3/5) ≈ 0.7746%.
Moments and tail comparison
| Property | Condition or illustration |
|---|---|
| Mean | 0 for ν > 1 |
| Finite variance | ν/(ν−2) for ν > 2 |
| Finite excess kurtosis | 6/(ν−4) for ν > 4 |
| P(T > 3), ν = 5 | Approximately 1.505% |
| P(Z > 3), standard normal | Approximately 0.135% |
The last two rows compare the conventional standard t scale with a standard normal. They do not hold the variances equal. At ν = 1 the t model becomes the Cauchy distribution, whose mean does not exist.
Excel illustration
| Task | Formula |
|---|---|
| 95% two-sided critical value | =T.INV.2T(0.05,24) |
| Lower confidence bound | =0.08-T.INV.2T(0.05,24)*0.10/SQRT(25) |
| Right-tail probability beyond 3, ν = 5 | =T.DIST.RT(3,5) |
| Cumulative probability at 3 | =T.DIST(3,5,TRUE) |
Check your understanding
Does “small sample” automatically justify a t confidence interval?
No. The sampling assumptions matter. Dependence, outliers and severe departures from normality can invalidate the simple small-sample calculation.
Does a heavier-tailed model automatically solve volatility clustering?
No. Tail thickness describes the unconditional shape. Time-varying volatility requires additional modelling.
Interactive distribution laboratory
Change the parameters and threshold, then switch between cumulative probability and density or mass. Values are numerical approximations for learning.
Horizontal axis: the variable’s value. Vertical axis: density, probability mass, or cumulative probability as labelled. Shading highlights values at or below the threshold. The plot shows a finite viewing window.
Sources and further reading
- National Institute of Standards and Technology: Gallery of Distributions
- Distribution definition and properties
- Microsoft: Statistical functions reference
All numerical finance and project examples are hypothetical teaching illustrations. Probability assumptions require validation before practical use.