The single biggest reason plans slip is that we estimate from the insideThe inside view builds an estimate bottom-up from the specifics of your plan — the tasks you can see. It systematically ignores the unknowns that hit every project. — task by task, as if nothing will surprise us. The fix is the outside viewReference-class forecasting (Kahneman & Lovallo) anchors on how comparable past projects actually turned out, not on the specifics of yours — so the unknowns you can't enumerate are already priced in.: take your estimate and apply the optimism biasThe systematic, well-documented tendency to underestimate cost and duration. The corrections here are the average historic bias measured across many real projects. its kind of work has historically run.
Adding to a product you know — a feature, a screen, an endpoint. The default kind of work, and the tightest of the new-build classes.
Plan to 12.2 weeks to be 85% sure — your bare 6-week estimate sits around the 26th percentile, where only about 26% of comparable projects come in.
Shaded band is the 5th–95th-percentile range; the dashed line is the median and the bold solid line is your 85% (plan-to) number. Click a stage — the range tightens and the optimism fades as you pin the work down. You're at scoped.
85% confident a feature this size lands within 12.2 weeks — +103.3% on your 6-week estimate.
| Confidence | Finish within |
|---|---|
50%confident | ≤ 7.9 weeks |
85%confident | ≤ 12.2 weeks |
95%confident | ≤ 15.8 weeks |
The centre of the forecast is the median overrun for this kind of work (anchored to the ~30–40% effort overruns measured across software studies); the range widens the earlier and more novel the work, and outcomes are modelled log-normal — a long right tailA few projects overrun catastrophically: Flyvbjerg finds roughly one IT project in six is a 'black swan' that runs 200%+ over. That's the long upper tail — which is why the median, not the average, is the honest centre., because work overruns far more often than it lands early. The per-type split is calibrated judgment — there's no clean public dataset of overruns by software work type — so lean on your own track record (the “My own factor” tab) once you have one.