Software estimation is systematically biased because developers estimate the median completion time well, but the mean is far higher due to the log-normal distribution of actual vs. estimated time. Using empirical data, the median blowup factor is exactly 1x while the mean is 1.81x. When summing multiple tasks, high-uncertainty tasks dominate the total time — a single task with σ=2 can dwarf all others combined. Fitting a Student's t-distribution to real project data reveals that the 99th percentile blowup is 32x, and the theoretical mean blowup factor for any unknown task is actually infinite. The practical takeaway: instead of summing estimates, identify and focus on the highest-uncertainty tasks, as they will dominate delivery timelines.

9m read timeFrom erikbern.com
Post cover image
Table of contents
Software estimationWhere’s the empirical data?Let’s go grab the statistics toolboxWhy software tasks always take longer than you thinkSummaryNotes

Sort: