If one is converting least square regressions into AIC using the 
following formula,

AIC = n log(RSS/n) + 2K

(with RSS being the residual sum of squares, n the sample size, and K 
the fitted parameters), then negative AIC are not uncommon. Really the 
sign of the AIC is inconsequential since it only the differences in AIC 
values that matter.

Barney Luttbeg

Reply via email to