I wrote an article about the method we use at Fog Creek for making software schedules which I've seen work very precisely, consistently, on projects from a week to two years. The basic approach is to make sure that the granularity of the tasks that you are estimating is fine. If you itemize the tasks at the procedure level (write subroutine x), where each task is less than 2 days, your schedule will work. The reason most people's estimates don't work is because they pull them out of the air, instead of actually thinking about what tasks they will need to complete. Getting down to the procedure level forces you to figure out what you're actually going to do, which is how you get a real estimate.