Control is a closed loop process.

You have some input, something happens, you look at the output and adjust your input again. This is implicit in quality procedures everywhere. This is what testing is.

Over the last few days I’ve been asked to estimate how long various bits of work for a potential customer will take. I’ve tried to make good estimates, but the truth is that I have never seen a comparison of how long was estimated at the beginning of a project, and how long was actually taken.

In other words, I still only have my gut feeling, my perceptions, to guide me as to how accurate I think my estimates are, despite the fact that this is a clearly measurable metric. The control loop is still open.

At the moment I provide estimates like “I think it’ll take about 10 days, but I’m not really very sure”. Wouldn’t it be better to know “Andy normally under estimates by and average of 10%, with a standard deviation of 10%”. Then if, for example, I estimated 10 days, we’d know that it’d actually 10-12 days, but only 68% of the time, or 9-13 days 96% of the time.

Perhaps this exercise is performed at a management level already – but as the developers are being the ones asked to estimate for technical details, it’s us that need our accuracy fed back for those aspects.

Now, I realise that there’s always going to be a human aspect to providing an estimate, that projects often aren’t that clear and easy to cut up, and that often we’re working with new things that we simply aren’t sure about. There’s also a cost-benefit question as to the effort in feeding back accuracy data. But I think that there are broad trends we should be able pull out.

I guess what I’m thinking is that we should have a process for optimising our estimates. I guess that this would involve examining statistical methodology. The best we’ve got at the moment is an informal “well that look longer/less time than I thought” after we’ve completed some work – and over projects that can span years, you just lose track.

It does strike me that it should be straight-forward to collate this data, and then feed it back at the end of the project. Close the loop. Regain some control.