Hello,
Part of potential answer to your question transcends domains, model types... with binary target. For example, "percent of variance explained" typically should be over 80% or 75% at least, regardless. But be careful not to engage too many variables because if you add many enough, you can always get over 80%.
One aspect of answer to your question is intrinsic, technical, inside the model mechanics, such of goodness of fit. When a model is mainly towards predicting, like digital stream response, or traditional camp response, we do not check this kind of fitness much. Instead, AUC, ROC,..... dominates while the model does often need to watch out model drivers. In traditional models like regression where your original question touches on, BIC, AIC, models often, out of domain requirement or not, do watch our model specification. With machine learning models, BIC, AIC... are often considered as secondary, defensive measure, not as primary criteria since it is practically hard to link them to popular accuracy. Still, domain matters as far as your question is concerned. If the model is 'pure prospect' targeting model, you should be happy with fairly low, weaker measures, while if the model is 'behavioral', the yardstick is much higher. Lastly, in high power DNN like models, BIC, AIC may not be part of the performance foundation at all. In the case of SVM, you don't normally get a score output. In some DNN models where all the events are captured in top, say, 1%, in other words very accurate, then there is little meaning to measure anything because you don't have traditional depth to calculate those 'old' measurement. + before running HPREDUCE, you may need to check correlation. Jia
