Core Path · L6.1.02

Fitting distributions by maximum likelihood

14 minMaximum likelihoodStudent tAIC and BICQQ plotGoodness of fit

By the end you will be able to

  1. 1Fit the normal and the Student t to real daily returns by maximum likelihood (M-D08: Maximum likelihood estimation).
  2. 2Report each estimate with its standard error, and the tail weight ν with a Wald and a profile-likelihood interval.
  3. 3Compare the fits with AIC and BIC, and explain why the likelihood ratio's χ² rule fails here.
  4. 4Judge the fit with QQ plots, the Kolmogorov-Smirnov and Anderson-Darling distances (M-D05: Rank correlation and distribution tests) and tail quantiles, and check the method by simulation.

Why it matters

L6.1.01: Stylised facts of financial returns showed that a normal is wrong for daily returns. Wrong by how much, and what instead? A normal fitted to GBP/USD puts a daily fall of 4% or more at once in 6.5 × 10¹⁰ days; twenty years of fixings had 3. The Student t fitted the same way says once in 2,919 days. The choice of distribution moves a tail probability by a factor of over twenty million, so it had better be made by a method that also says how sure it is.