A non-parametric test for dependence based on the entropy rate

18 February 2014
Pedro Vitoria (Stochastic Analysis group) and Galen Sher (Economics)
A non-parametric test for dependence between sets of random variables based on the entropy rate is proposed. The test has correct size, unit asymptotic power, and can be applied to test setwise cross sectional and serial dependence. Using Monte Carlo experiments, we show that the test has favourable small-sample properties when compared to other tests for dependence. The ‘trick’ of the test relies on using universal codes to estimate the entropy rate of the stochastic process generating the data, and simulating the null distribution of the estimator through subsampling. This approach avoids having to estimate joint densities and therefore allows for large classes of dependence relationships to be tested. Potential economic applications include model specification, variable and lag selection, data mining, goodness-of-fit testing and measuring predictability.
  • Junior Applied Mathematics Seminar