Why are software development task estimations regularly off by a factor of 2-3?

Because speculation tends to be wrong.