What I know now is that smaller AIC values and larger likelihood values indicate better fit. I am trying to compare two models fitted to the same data. Model 2 should fit the data better, but not necessarily.

I use R, forecast package and msts- and tbats functions for fitting the models.

> fit1$likelihood
[1] 90854.67
> fit1$AIC
[1] 90884.67
> par1 # par1 <- length(fit1$parameters$vect), number of fitted parameters
[1] 4
> 2*par1-2*log(fit1$likelihood) # AIC
[1] -14.83403
> fit2$likelihood
[1] 90766.44
> fit2$AIC
[1] 90824.44
> par2 # par2 <- length(fit2$parameters$vect), number of fitted parameters
[1] 7
> 2*par2-2*log(fit2$likelihood) # AIC
[1] -8.83209

So, AIC value for fit2 is smaller than for fit1 (that is what I hope to get). But, the likelihood for fit1 is bigger than for fit2. What should I believe?

Also, if I have done the calculation of the parameters correctly, why are those AIC values calculated "by hand" so different that those which I get from the fits directly?

fit1 and fit2 are tbats-objects.


Looking at the source code of the package:

aic <- likelihood+2*(length(param.vector$vect)+nrow(x.nought))

x.nought is some additional parameter particular to the tbats model, and I can't really comment on it.

It becomes clear though the likelihood object is actually equal to $-2\ln(L)$.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.