How forecasts are calculated
A bootstrap of your own past throughput, with scope growth and your real sprint calendar. What 50%, 85% and 95% mean, and how the backtest is read.
Sprint Radar does not use a formula for your velocity. It replays your own past, many times, and counts how long the work took in each replay. This page says exactly what it does.
The method: bootstrap of your own throughput
- Periods. History is cut into periods. A period is a sprint when your data has dated sprints, otherwise a calendar week. Periods in which nothing was finished stay in as zeros: a bad week is part of your past.
- Throughput. For each period, Sprint Radar counts what was finished (story points, or items) and what was added (issues created in that period). The unit is chosen automatically: story points when at least 80% of the remaining issues have an estimate, otherwise item count.
- One simulated future. Start with the remaining work. For each future period, draw one past period at random, with replacement, subtract what it finished and add the scope it added. Stop when nothing is left. The number of periods is one result.
- Many futures. This is repeated 10,000 times by default. The runs are seeded, so the same data gives the same forecast.
- Window. The draw uses a window of recent periods. The window is chosen from the shape of your history, and the report states which one was used.
Scope growth
Scope growth is measured, not assumed. Issues created in a period count as work added in that period. Issues closed with the resolution Won't Do, Won't Fix, Duplicate, Cannot Reproduce or Declined are not counted as done and not counted as growth. Because every simulated period brings back the scope added in the drawn past period, a team whose backlog grows as fast as it shrinks gets a forecast that says so. If no past period ever reduced the work, no run can finish and the forecast says “may not finish”.
The sprint calendar
Periods become dates with the real cadence of your team: the length of a period is measured from your sprints or weeks, and the first future period is the part of the active sprint or week that is left. The date of the n-th period is the end of the first one plus n − 1 periods.
What 50%, 85% and 95% mean
After all runs are sorted from fewest to most periods, the 50% date is the one that half of the runs finished by. The 85% date is the one that 85 of every 100 runs finished by, and the 95% date is the one that 95 of 100 finished by. The report also shows 70%. Plan with 85% when a missed date is costly. The 50% date is, by its definition, missed about half of the time.
The backtest
Sprint Radar checks itself on your history. For each past point in time with at least 6 periods before it, it hides the future, forecasts how long the work the team usually finishes in 3 periods would take, and compares that with what happened next. The same simulation as the real forecast is used.
The result is a sentence such as “Past forecasts 3 sprints ahead at 85% confidence held 22 of 27 times”. It means: in 27 past checks, the actual result was inside the 85% forecast 22 times. About 23 of 27 would be expected, so 22 is on target. The count is shown only when there are at least 8 checks. Sprint Radar calls the forecast optimistic or conservative only when the hit rate differs from the stated level by more than ordinary chance (a 5% two-sided test).
The backtest checks forecasts about 3 periods ahead. Longer forecasts are not validated, and the report says so.
Warnings you may see
- Less than 8 periods of history. The forecast is less reliable than its confidence levels say.
- Throughput is falling (or rising). With at least 8 periods, a Mann-Kendall trend test at the 5% level finds a clear trend. A falling trend means resampling the past is optimistic, a rising one that it may be conservative.
- Issues without an estimate count as 0 in the remaining work and in past throughput, when the unit is story points.
- Past forecasts held less (or more) often than their level says: treat the forecast as optimistic (or conservative).
- Notes about changelogs that were too long to read, or a status list that could not be read, when cycle time is affected.
What it does not do
It does not model people, holidays or a planned change in team size. It assumes the future resembles the recent past, and the trend test tells you when that is doubtful.