Sir,

Stratified 10 fold cross- validation is recommended for testing accuracy than hold out and other methods.  It has uniform class distribution of tuples in each fold. Training & testing is performed on each folds iteratively. Accuracy is estimated by calculating the over all number of correct classification from 10 iteration, divided by total number of tuples in initial data.

 Reference : Han & Kamber, " Data Mining - Concepts and Techniques"

Regards,

Anuradha J