Creating portfolio risk and return models [18]

Go to: Summary | Previous | Next   
Bullet points include: Can be thought of as an ‘all-at-once version’ of projection pursuit Involves working out the maximum likelihood estimator of the entire unmixing matrix, assuming the signals are independent Needs an a priori distributional form to assume for the individual signals Often choose one with very high kurtosis, e.g.  Or ‘infomax’ ICA Identify how ‘surprising’ (and therefore meaningful) is the observed data given some a priori multivariate distribution in which each individual series is independent, measured using, say, relative entropy (aka Kullback-Leibler divergence)

Contents | Prev | Next | ERM Lecture Series

Desktop view | Switch to Mobile