I'm using maximum likelihood estimation to fit a model to time series data. I later want to use the model to do forecasting. Let $L$ be the likelihood function, $$L_i = L(f_k(\theta), data_i).$$ It depends on some distribution $f_k$ of the parameter(s) $\theta$ and it depends on the observed data for timeseries i. I use MLE to find the optimal $\hat{\theta}$ for several different distributions $f_k$ and want to to pick the best model, i.e $f_k$ in combination with $\hat{\theta}$. For example I want to compare $f_k$ being a lognormal distribution or $f_k$ being a gamma distribution. How would I use the Akaike information criterion(AIC) to do this? I assume that I can calculate and use (AIC) to pick the best type of distribution $f_k$ and $\hat{\theta}$ for the same $data_i$. Is this correct?

If I wanted to find the best $f_k$ for several time series $data_i$ for i 1 to N. Would it bet better to take the $f_k$ which has had most often the lowest AIC? Or would I instead take $f_k$ with the lowest average AIC?

My confusion comes from the fact, that I can construct a likelihood function for the whole dataset which would be $L = \prod_i L_i$. Since $ \ln L = \sum_i \ln(L_i)$ the AIC for the whole dataset essentially is the sum of the individual AICs. However, this can easily be dominated by a single time series for which a certain $f_k$ does not fit well even if it performs much better on all other data time series.



Your Answer

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