For projects where an estimate is very important (not pointing every story in a sprint), I will do an 80% confidence bounds where on one end everything goes perfectly, and the other where everything goes horribly wrong. For both, it should only be outside the particular bound in 10% of possible worlds. Then I take the geometric mean of the two values.

This has worked with pretty good consistency in the past. I believe that a good deal of the value comes from factoring in the estimate where everything goes wrong. You can also then specifically work to prevent the things you’ve thought about going wrong from occurring.