Today, dr. Hamparsum Bozdogan (also see here), one of our own Faculty members, delivered a talk in our PhD seminar entitled 'Predictive Subset Vector Autoregressive (VAR) Modeling Detection of Influential Observations and Forecasting using the Genetic Algorithm and Information Complexity'. This research uses ICOMP, a measure developed by dr. Bozdogan, as an extension of Akaike's Information Criterion (AIC). Dr. Hirotugu Akaike was Dr. Bozdogan's mentor. ICOMP is a perfect measure for VAR model selection as they are very easily over-parameterized (overfit). Here is the video and some visual impressions of his talk. Thank you dr. Bozdogan for a very interesting afternoon.
I finally learned where the 2 comes from in the penalization term 2m(k). Dr. Bozdogan says that this is the result of empirical analyses by Akaike, but one can definitely cross-validate that constant on new data. ICOMP integrates AIC but also takes into account the dependence between the independent variables.
Here we see the effectiveness of ICOMP and the Genetic Algorithm in creating a more parsimonious solution than what we had in slide 44.
