And the majority of (non-software) engineering projects also involve iteration.
The majority of software projects, I believe, can be estimated with reasonable accuracy as long as you ensure the project is a development project and not a research project. Most of the literature on this subject classify the types of software projects but nearly all recognize a category of "research project" and most agree you cannot estimate such a project. Note this finding applies to all branches of engineering.
A major problem I see with software engineering is developer ego resisting accepting that most of what they do is mundane and has been done many times before. They resist this ego-deflating fact by trying to turn every project into a "research" project; thus instead of using simple boring dependable tools (let's say Apache, MariaDb or Postgres and Php) to deliver working software, many developers/geeks (and I consider myself one) will try their utmost to turn the project into a research project. So it will be proposed to use a Hadoop cluster for storage and the latest cutting edge ORM tool with load-balancing and an XML based messaging bus talking to a front end written using Nimgorust or whatever cool new programming language is out there. Of course, estimating such a project is an order of magnitude more difficult than a project involving mature well-known tools. But this difficulty is a product of technical decisions (which you can control) and not inherent in the task of delivering working software.
FWIW, that procurement inefficiency represents an opportunity for software entrepreneurs who have the patience to deal with the sub-optimal process. It also probably represents an opportunity for large firms, in terms of how to cut down on software expenses (presumably by bringing development in-house, although that obviously comes with its own overhead).