What techniques are available to calculate the accuracy of ranged estimates, on units of work that may be irregularly re-estimated? The evaluation should account for estimates made on behalf of another, and count those toward the estimator's rating.
Our tooling allows and encourages us to enter ranged estimates at the 80% confidence level, but critically we can't get the feedback that tells us we're estimating too high/low/wide/narrow etc. Without which, we cannot improve.
One thing that might help you estimate each task for effort (sheer amount of work), difficulty (amount of thinking) and risk/uncertainty (how much can go wrong). Then let those assessments guide your estimation.
The FogBugz approach. Estimate normally, record how long it actually took you and let a computer program use that information to auto-translate all your future estimates. I was incredibly taken with this idea when I first heard about it. Unfortunately logging all those estimates AND actual times is laborious.
I've read of experiments where they've found that more experienced developers tend to create more accurate estimates even when they are not the ones doing the actual work. This has the downside of putting your best workers doing most of the chores. And those who are not good at estimating have no way of getting better at it.
People tend to better at comparing tasks relative to each other than at predicting absolute time required. You still need to track how long it actually took you.
For B and D, tracking actual time only for batches of tasks greatly eases the administrative burden and still gives you some information of "when I estimate X points/hours that tends to equal around Y hours". If you don't end up comparing estimates with outcomes you will never get better. So a process that produces data with less precision is preferable to a process that produces no data because it's too inconvenient and people don't follow it.
If you put all of it together you get something like this:
