The retro opens with a slide: “Forecast accuracy: missed.” The release committed for 14 March shipped on 19 March. Someone says the forecast was wrong. Someone else says the vote was too optimistic. The meeting moves on to what to do differently, and the team quietly decides to pad the next date.
Nobody asked the question that decides whether the forecast was wrong: how often do dates committed at that confidence level actually land?
Forecast accuracy is a property of a set of commitments
A single date is either met or missed. A date committed with a stated confidence level makes a different claim. An 85% commitment says that, across many similar releases, about 85 of every 100 should land on or before the date, and about 15 should not. Judging that claim from one release is like judging a weather forecast from one rainy day.
The useful measure is calibration: compare the confidence you committed at with the share of releases that actually landed on time.
- Commit at 85% and land about 85% of the time: well calibrated.
- Commit at 85% and land 50% of the time: the forecast is overconfident, or something is eroding the date after the vote.
- Commit at 85% and land 100% of the time: the team is leaving margin on the table, and probably committing at a higher confidence than it needs to.
The team’s confidence vote is expert engineering judgment, and calibration is how that judgment gets better. It is a learning loop, not a verdict on the people who voted.
Why a well-made forecast still misses
The forecast on vote day can be sound and the release can still land late. The assumptions behind the vote were true on that day. Then the three things that erode a committed release date go to work:
- Scope creep adds work the forecast never saw.
- Dependencies surface too late and add a chain nobody sequenced.
- Capacity changes: people are pulled away, absences land, interrupt work grows. See what overallocation costs a committed release date.
This matters for how you read your own history. A late release tells you one of two things: the vote was miscalibrated, or the plan changed after the vote and nobody re-read the date. Those need opposite fixes, and a hit-or-miss slide cannot tell them apart.
Grade the commitment, not the snapshot
An illustration, not customer data. A team keeps a simple log of ten releases: committed date, confidence level, and the date each actually shipped. Eight landed on or before the committed date at 85%. That is 80% against an 85% claim, close enough that ten releases cannot tell it from noise. The team keeps its method.
Now suppose only five of the ten landed on time. The team goes back through the five late ones and finds three where scope was added after the vote, one where a dependency surfaced in week five, and one where two engineers were reassigned. The votes were not the weak link. The date was never re-read once the plan moved. That points at a different fix from a more conservative vote.
A workable routine:
- Log every commitment with its date and the confidence level it was made at.
- After release, record whether it landed on or before the committed date.
- For each late release, note which of the three erosion forces moved it, if any.
- Once you have around ten, compare the confidence you voted at with the share that landed on time.
The log is the team’s own practice and lives wherever the team keeps its records.
Where Advanced Release Planning fits
Advanced Release Planning for Jira works on the live side of the loop. It runs a Monte Carlo simulation over the team’s real Jira throughput and reports the committed date with its confidence band: 50%, 85% and 95%. As scope, dependencies and capacity change, it re-runs the simulation and reports the gap between the committed date and the forecast at the committed confidence, in days. Green at or above P85, amber between P50 and P85, red below P50.
The team still owns the assumptions: which throughput window, whose capacity, what is in scope. The app does the arithmetic at Monte Carlo scale and keeps it current. That is the difference between a confidence vote as a snapshot and one kept live: when the date drifts, the team can see it in week three and renegotiate while there is margin to trade, instead of discovering it at the retro. For a companion view, see why a schedule risk analysis run once at planning goes stale. For exactly what the app reads, see the forecast accuracy reference doc.
Key takeaways
- Forecast accuracy for a probabilistic forecast is judged across many commitments, by calibration, not by one hit or miss.
- A date committed at 85% should land about 85% of the time. One late release does not prove the forecast wrong.
- Late releases come from miscalibrated votes or from erosion after the vote: scope creep, late dependencies, capacity changes. Log which, so the fix matches the cause.
- Keep the date current between commitments so drift is visible while it can still be renegotiated.
FAQ
What is forecast accuracy in project management?
Forecast accuracy is how closely forecasts match what actually happened. For a single date it is a hit or a miss. For a forecast made at a confidence level, it is judged across a set of commitments by comparing the stated confidence with the share that landed on time.
How do you measure the accuracy of a release forecast?
Log each committed date with its confidence level and its actual release date. After around ten releases, compare the share that landed on or before the committed date with the confidence level used. A large, persistent gap means the forecast is miscalibrated or the plan is changing after the vote.
Is a missed date proof the forecast was wrong?
No. At 85% confidence, about one release in seven is expected to land late. A single miss is expected. A pattern of misses at a stated confidence level is the signal worth investigating.
Keep your confidence vote. Keep the date current too. See the committed date and its live confidence band in Advanced Release Planning, or start a free trial on the Atlassian Marketplace.




2 Comments
Leave your reply.