In my experience time-based estimates pretty much always wildly inaccurate, but people can get pretty good (good enough for project management) using complexity

In my experience time-based estimates pretty much always wildly inaccurate, but people can get pretty good (good enough for project management) using complexity-based estimates. Benchmark a small task whose complexity the group agrees on and go up in orders of magnitude from there. Then look at velocity over time. You get a pretty accurate picture of your throughput in a high-, medium-, and low-confidence scenario (high confidence we can do this much, medium confidence we can do this much more, etc.).