[Gretl-devel] constant mis-specification and arima convergence problem

Dear all, ########## redundant constant open denmark.gdt # note a decent model: arima 2 1 0; LRM --nc modtest --autocorr 4 # # Ljung-Box Q' = 2.40961, # with p-value = P(Chi-square(2) > 2.40961) = 0.2997 # Real Imaginary Modulus Frequency # ----------------------------------------------------------- # AR # Root 1 -1.5828 0.0000 1.5828 0.5000 # Root 2 1.4501 0.0000 1.4501 0.0000 # ----------------------------------------------------------- # so models below # are seasonally over-differenced set bfgs_toler default catch arima 1 1 0; 0 1 1; LRM err = $error eval errmsg(err) arima 1 1 0; 0 1 1; LRM --nc ma = 10^-5|$coeff tol = 10^-9 set bfgs_toler tol set initvals ma arima 1 1 0; 0 1 1; LRM --verbose arima 1 1 0; 0 1 1; LRM --x-12-arima # Note # Scaling y by 93455.8 set bfgs_toler default catch arima 1 1 0; 1 1 0; LRM err = $error eval errmsg(err) # arima 1 1 0; 1 1 0; LRM --verbose # Scaling y by 93455.8 ols diff(LRM) 0 -s arima 1 1 0; 1 1 0; LRM --nc ma = 10^-5|$coeff set bfgs_toler default set initvals ma arima 1 1 0; 1 1 0; LRM arima 1 1 0; 1 1 0; LRM --x-12-arima # here we have almost the same estimates # and log-lik # pol. roots are very decent #### the main indicator is high p-value at const # So another situation with bad # convergence is when we have a const # with very high p-value # i.e. redundant const # some indicators: slowly growing lnl # and small gradients ############# missing constant open greene5_1.gdt logs * # a decent model arima 0 1 2; l_realcons # looking at p-value for theta_1 (0.9356) # select matrix qvec = {2} arima 0 1 qvec; l_realcons modtest --autocorr # Test for autocorrelation up to order 4 # # Ljung-Box Q' = 1.34035, # with p-value = P(Chi-square(3) > 1.34035) = 0.7196 # pol. roots are well behaved arima 0 1 qvec; l_realcons --nc modtest --autocorr # no problem with convergence but autocorrelation # We consider arima 1 1 1; l_realcons modtest --autocorr # well-behaved roots but autocorrelation # Note p-value on const is 1.05e-029 set bfgs_toler default scalar reallyhuge = 2*$huge set bfgs_maxgrad reallyhuge arima 1 1 1; l_realcons --nc arima 1 1 1; l_realcons --nc --x-12-arima # some indicators: large norm of gradient # and nearly unit circle roots Oleh