Inputs and provenance
Use task scope, completed work over time, explicit dependencies and recorded capacity. Keep source timestamps and definitions available. Avoid treating a missing estimate as zero work.
Methodology & evaluation
A useful forecast shows its uncertainty.
Understand the inputs, assumptions and limitations behind predictive delivery. No invented accuracy score or guaranteed date.
Start with evidence.
Delivery estimates depend on the work recorded, the history available and the constraints a team can explain. Missing data, changed process and unusual projects can all make estimates less reliable.
The examples on this site use fictional, seeded workspace records. They demonstrate the interface and an approach to simulation; they are not an evaluation of production model performance.
01Recorded history
02Sampled throughput
03Delivery intervals
The forecast demonstration
A seeded simulation generates 1,000 delivery paths. Weekly throughput is drawn from a lognormal distribution fitted to fictional history. The median and quantiles form the displayed forecast bands.
Evaluation before use
Evaluate on held-out projects in time order, rather than mixing future outcomes into training. Compare with simple historical baselines. Review interval coverage, date error and performance by project type.
Calibration and drift
A 90% interval should cover outcomes near that rate across comparable forecasts. Measure coverage over time, check changing work practices and recalibrate when the historical assumptions no longer fit.
Human review
Review sources and model suggestions before changing scope, ownership or commitments. A confidence value measures a model output under its assumptions; it does not make an answer correct.
Limits worth keeping visible
Small samples, untracked work, seasonality and correlated blockers can distort results. Predictions describe possible outcomes. They should support a planning conversation, not replace professional judgment.
Read the range correctly.
The central 50% band spans the 25th to 75th percentile of the simulated paths. The wider 90% band spans the 5th to 95th percentile. These are simulation intervals, not independently measured accuracy or a promise about a real project.
A completion probability is the fraction of simulated paths that reach the current scope by the deadline. Adding scope recomputes the paths from the same seed, allowing a consistent comparison. It does not predict the behavior of a real team without validated input history.
Try the forecast demonstrationExplore the calculation
Change one assumption. Inspect the result.
Try the scope slider in the forecast demonstration. The same fictional throughput history and random seed are used for each comparison, while the amount of remaining work changes.
Inspect both the probability curve and the completion distribution. A result depends on the inputs and model assumptions, so the chart should travel with an explanation of the data that produced it.
ACDFMC+1
Illustrative workspace · interactive demonstration
Evaluation checklist
Ask how the model behaves on later projects.
| Evaluation question | Evidence to inspect |
|---|---|
| Can it beat a simple baseline? | Compare date errors with historical median throughput or another explicit planning baseline. |
| Are the intervals calibrated? | Check observed coverage across comparable held-out projects, alongside the width of the forecast intervals. |
| Is the test time-aware? | Keep later project outcomes out of the training and parameter-fitting period used for their forecasts. |
| Where does it fail? | Review small samples, changed processes, unusual scope and dependencies that the available history does not represent. |
| What changed over time? | Track data definitions and team practices. Revisit assumptions when the delivery process changes. |
Questions & answers
A few useful answers.
Why show more than one interval?
The central 50% band and the wider 90% band show different parts of the simulated range. Together they help a reader distinguish the central path from more uncertain outcomes.
What does the fixed seed do?
It makes the demonstration reproducible. Comparing the same sampled paths with different scope isolates the effect of that scope change within this model.
What is outside this demonstration?
It does not evaluate production model accuracy, learn from your team’s records or account for every real-world dependency. Relevant input history and independent evaluation are needed before operational use.
Ship what you promised.
Bring the plan, the work and the next decision together.