Let say I'm building a model to predict the number of accidents in a insurance portfolio of automobiles. The problem is that my model is is very very sensitive to the seed (for each time I change the seed, the significance of each predictor changes) (this is due to the fact that there is only a small percentage of accidents among the portfolio). So I come up with an "invention": I change the seed 1000 times, and I calculate the empirical probability of significance at level of 5% for each predictor. By doing that, I will know that what predictor is really significant (for example, with a probability of significance at level of 5% greater than 70%). This "invention" is really similar to the idea of bagging, but I'm not sure that it is correct or not? Can someone help explain to me?