Asymptotic Bias in Parameter Estimation of AR‐Processes Using Recursive Least Squares with Exponential Forgetting (Q4956072)

For autoregression processes with i.i.d. (not necessarily Gaussian) perturbations a recursive least squares (RLS) estimator of AR-parameters is considered. In RLS an exponential forgetting of old information is used with a constant forgetting factor \(\lambda\). Using second order Taylor expansion the authors obtain an approximate formula for the bias of the RLS estimator. E.g., for AR(1) process the bias is \[ -2a\lambda(1-\lambda)(1-a^2)(1-a^2\lambda)^{-1}(\lambda+1)^{-1} +O((1-\lambda)^2)\sim -a(1-\lambda) \] as \(\lambda\to 1\). Simulation results for AR(1) and AR(2) are presented.