Biases of Estimators in Multivariate Non‐gaussian Autoregressions

Expressions for the bias of the least‐squares and modified Yule‐Walker estimators in a correctly specified multivariate autoregression of arbitrary order are obtained without assuming that the innovations are Gaussian. Instead, the innovations are assumed to form a martingale difference sequence which is stationary up to sixth order and which has finite sixth moments. The errors in the expressions are shown to be O( n ‐3/2 ), as the sample size n under some moment conditions. The expressions obtained are the same in the Gaussian and non‐Gaussian cases.