Statistics Glossary → Error distribution (normal, Student-t)
Error distribution (normal, Student-t)concept
The probability distribution assumed for a model's random error (innovation) term; it determines the likelihood, the tail behaviour and the forecast intervals.
In models estimated by maximum likelihood, the distribution of the error term must be specified explicitly. The normal distribution is the most common assumption, but many real data sets, financial returns in particular, contain extreme values more often than the normal predicts (fat tails). The Student-t distribution adjusts the thickness of the tails through its degrees of freedom ν: with small ν extreme values are more likely, and as ν grows the distribution approaches the normal. In GARCH models the t distribution is usually scaled to unit variance and ν is estimated from the data.
Assuming normal errors for fat-tailed data understates the probability of extreme events and makes forecast intervals (such as value at risk) too narrow; parameter estimates can also become overly sensitive to extreme observations. The estimated ν is informative: values between about 4 and 8 point to clearly fat tails. In parametric survival models, the choice of error distribution directly determines the distribution of times (an extreme-value error gives Weibull times, a normal error log-normal times). Distributions can be compared with AIC and with Q-Q plots of the standardized residuals.
An everyday example
An investor works out the probability that the portfolio loses more than 5% of its value in a single day. A model assuming normal errors finds this so rare that it should happen once every few years, yet past data show several such days a year. A model with Student-t errors estimates ν = 5 and gives a probability of extreme losses much closer to reality.
In MerQur
In GARCH, the 'Error distribution (Normal / Student-t)' option; the t distribution has unit variance, ν is estimated and forecast intervals use the quantiles of the model distribution. In Parametric Survival, the chosen distribution determines the error distribution of the log-time model.
Analyses where this concept appears
Parametric Survival (AFT)GARCH (Volatility Model)