My question is related to bootstraping and extracting confidence interval for classification problems. Let's say I have 25 number of data points and 2 features and use Gradient Boosting Machine (GBM) to develop a binary classifier. Then I extract AUC to examine what's the accuracy of my model. When I do bootstraping by reshuffling the data points (basically changing order of data points), I could get different AUC and I do this work for 1000 times and then will find 95% confidence interval and p-value. But, in my opinion, it's something nonsense. Essentially, why my predictive model (i.e. GBM) should depend on order of data points that I use for training? I mean how do I know which order is correct cause my data points do not have any order. I can't wrap my head around this and I will be very grateful if someone could explain it to me.