When you have 2000 observations and 1900 variables, what is the best performance measurement?

Question

When you have 2000 observations and 1900 variables, what is the best performance measurement? Assume we have numeric output (i.e. not a classification problem)