Information criteria for choosing best predictive models
29th May 2012
Usually I’m using 10-fold (non-stratified) CV to measure the predictive power of the models: it gives consistent results, and is easy to perform (at least on smaller datasets).
Just came across the Akaike’s InforÂmaÂtion Criterion (AIC) and Schwarz Bayesian InforÂmaÂtion Criterion (BIC). Citing robjhyndman,
AsympÂtotÂiÂcally, minÂiÂmizÂing the AIC is equivÂaÂlent to minÂiÂmizÂing the CV value. This is true for any model (Stone 1977), not just linÂear modÂels. It is this propÂerty that makes the AIC so useÂful in model selecÂtion when the purÂpose is prediction.
…
Because of the heavÂier penalty, the model choÂsen by BIC is either the same as that choÂsen by AIC, or one with fewer terms. AsympÂtotÂiÂcally, for linÂear modÂels minÂiÂmizÂing BIC is equivÂaÂlent to leave–v–out cross-​​validation when v = n[1-1/(log(n)-1)] (Shao 1997).
Want to try AIC and maybe BIC on my models. Conveniently, both functions exist in R.