I think that one of the factors making it so difficult to produce accurate estimates on software projects is their usually high levels of variability. There are often so many big differences between projects and it's hard to consistently draw lessons from past experiences.
However, it seems that when such variability is low, it gets easier to make predictions.
I've seen an ERP company which used to provide good estimates for most of their customization projects. However, they used a home-built, suited-to-the-task programming language, the team was experienced and relatively stable, and the projects' scopes were usually narrow.
Those factors are not easily reproducible, for sure. But I believe they indicate that taking measures to reduce variability and standardize processes may pay dividends in the long run -- projects may take longer than they would take under agile approaches, but the improved predictability may be desirable.
However, it seems that when such variability is low, it gets easier to make predictions.
I've seen an ERP company which used to provide good estimates for most of their customization projects. However, they used a home-built, suited-to-the-task programming language, the team was experienced and relatively stable, and the projects' scopes were usually narrow.
Those factors are not easily reproducible, for sure. But I believe they indicate that taking measures to reduce variability and standardize processes may pay dividends in the long run -- projects may take longer than they would take under agile approaches, but the improved predictability may be desirable.