1. LandOffer
  2. Blog
  3. All articles
  4. Forecast Delivery with Cycle Time Data Without Promising an Exact Date

Product & Project Delivery

Forecast Delivery with Cycle Time Data Without Promising an Exact Date

Build a single-item delivery forecast from a defined cycle-time sample and an explicit percentile convention. Two completed synthetic cases show how historical proportions differ from calibrated future probabilities, why aging work changes the question, and when a changed approval process invalidates the old forecast.

10 min readLandOffer team

Use cycle-time data to describe how long comparable work has taken from a defined start to a defined finish, with explicit probability and assumptions. Do not turn one percentile into a guaranteed delivery date. First check whether the current work resembles the history, whether the workflow boundaries match, and whether anything has changed in staffing, work in progress or external approvals.

This LandOffer.ai guide builds a single-item forecast and then revises the decision when the historical work no longer matches. All work records, team conversations and future planning dates are original synthetic examples. The calculations were checked on those records; they do not validate future delivery probabilities, test a real planning product or promise an actual release.

A craftsperson examines three wooden boat models at different stages of completion on a workshop bench.

Decide which delivery question the data can answer

The Kanban Guide, May 2025, defines cycle time using the workflow's start and finish boundaries. It distinguishes completed-item cycle time from the age of unfinished work, the number of items in progress and throughput over time. The same guide describes a service level expectation as a time-and-probability forecast informed by historical cycle time.

That boundary matters. An item started today is different from a request waiting in a queue or a partly completed item already nine days old. If your clock starts at implementation, state explicitly that an unknown wait in the queue before implementation is outside that forecast. A finish defined as “merged” does not establish when the change reaches a customer if deployment or acceptance happens later.

A whole-release question is different again. Scrum.org's probabilistic forecasting paper, published in May 2024, discusses cycle time, throughput and Monte Carlo forecasting for delivery questions. A single item's 85th-percentile cycle time is not automatically the time needed to finish six items, nor should you multiply it by six without modeling how work actually flows.

For a batch, use a suitable approach to scope and historical throughput, including periods with zero completions, while examining capacity and dependencies. This article's numeric example concerns one newly started item. It does not run or claim a release-level Monte Carlo forecast.

Define the sample before calculating a percentile

Fictional team HarborGrid maintains small internal reporting changes. For its ordinary-review workflow, “started” means the team begins implementation after pulling the item, and “finished” means the change has completed internal review and been deployed to the internal users. Waiting for that internal review remains inside cycle time.

The unit is elapsed calendar days, measured as finished timestamp minus started timestamp in UTC. A two-day difference means two elapsed days; the calculation does not add an extra day for inclusive date counting. Business-day calculations could be useful elsewhere, but switching conventions mid-sample would produce a different distribution.

The team selects 20 completed items from the same ordinary-review class, with synthetic completion dates between September 3 and October 27, 2026. It retains the long durations, including one 18-day item. Retain legitimate long durations; investigate their records instead of deleting them to produce a more attractive forecast.

The sorted durations are 2, 2, 3, 3, 3, 4, 4, 4, 5, 5, 6, 6, 7, 7, 8, 9, 10, 12, 14 and 18 calendar days. These are the full original example values, not results from a customer tracker. A real extract should also preserve the item identifiers and timestamps needed to verify the calculation and explain exclusions.

Calendar duration also includes weekends and waiting inside the selected boundaries; it is not a measure of developer effort. An eighteen-day item need not represent eighteen days of uninterrupted coding. HarborGrid retains that observation and examines its record before deciding whether it is a data error or a legitimate part of the workflow. If a staffing interruption is expected to recur, silently removing its historical effect would make the sample less representative of the conditions being forecast.

Do not put unfinished items into this list with their current age treated as completed duration. Keep them visible separately so that a completion-only view does not hide a growing queue of old unfinished work. Check duplicate records, inconsistent completion states and changes in the meaning of “done” before using the sample.

Completed single-item forecast with an explicit calculation rule

HarborGrid uses the nearest-rank convention for this example: sort the durations, multiply the sample size by the requested percentile, round the position upward to the next whole number, and select that observation using positions starting at one. Record the convention with the forecast so another person can reproduce the selected observations rather than relying on an unspecified dashboard interpolation.

Percentile Position calculation in 20 items Selected cycle time What the sample actually shows
50th Ceiling of 20 × 0.50 = 10. 5 days. 10 of the 20 completed items finished within five days.
85th Ceiling of 20 × 0.85 = 17. 10 days. 17 of 20 finished within ten days.
95th Ceiling of 20 × 0.95 = 19. 14 days. 19 of 20 finished within fourteen days; one took eighteen.

For this even-sized sample, the conventional median obtained by averaging the two middle observations would be 5.5 days: the tenth value is five and the eleventh is six. The nearest-rank 50th-percentile value is five. Both calculations can be defined, but they are not the same operation. HarborGrid reports the nearest-rank convention consistently and does not relabel five as the averaged median. A reviewer can reproduce the table directly from the displayed durations without guessing a dashboard default.

Those are empirical proportions in a small authored dataset. Calling ten days an 85% planning forecast for the next comparable item additionally assumes the historical observations are informative about that item's future workflow. The historical table alone cannot establish that the next item has the same probability distribution or that future outcomes will match the forecast's stated coverage.

For a hypothetical new ordinary-review item starting October 28, the team communicates the result as follows:

Our 20 comparable completed items show a nearest-rank 50th percentile of five calendar days, an 85th-percentile duration of ten and a 95th-percentile duration of fourteen. We propose using the ten-day point as a working planning forecast for this newly started item, subject to the same scope, internal-review process and capacity. It can take longer. This forecast starts at implementation, includes internal review through deployment, and is not a contractual delivery commitment.

If the start timestamp were October 28 at midnight UTC, those five-, ten- and fourteen-day elapsed endpoints would fall on November 2, 7 and 11. They are illustrative translations of durations, not three promised delivery dates. An unknown start time, queue wait or release gate would require its own treatment.

The team chooses which planning risk to discuss with stakeholders rather than presenting the 85th percentile as a universal standard. A high-consequence event might need a different decision, reduced scope, an alternative launch plan or explicit contingency. Merely choosing a higher percentile does not remove dependencies the data never measured.

Do not subtract an item's age from the percentile

ProKanban's discussion of service level expectations treats aging work as a reason to inspect new information and intervene. An SLE is not a countdown that guarantees completion when its duration is reached.

Suppose an ordinary-review item is already nine days old. “Ten minus nine equals one day left” is not a sound inference. In the synthetic completed sample, only four durations exceed nine days: 10, 12, 14 and 18. Their additional durations beyond day nine are one, three, five and nine days. Just one of those four finished within one additional day.

That tiny tail is not a calibrated forecast for the active item either. It illustrates why its age changes the question. Review the actual blocker, remaining scope and current workflow rather than repeating the original ten-day statement. A nine-day item waiting on an approval outside the original process may no longer be comparable at all.

Uneven wooden pieces and a cord pause at a small gate model on a pale workbench.

Completed mismatch review: external approval changes the forecast

HarborGrid next plans a customer-facing reporting change that requires external partner approval. The old ordinary-review finish did not include that gate. The team has eight synthetic completed examples from an external-approval workflow, measured through external partner approval and deployment to the intended customers, using the same elapsed-day convention: 6, 8, 9, 12, 14, 16, 20 and 24 days.

Nearest-rank descriptive percentiles are 12 days at the 50th, 20 at the 85th and 24 at the 95th. With only eight observations, the selected 20-day point includes seven of eight items, or 87.5%; the 24-day point includes all eight. Treat these as descriptive sample cutoffs: they do not establish precise future 85% or 95% coverage, and a future item can exceed the observed maximum.

The team completes this adjustment record instead of quietly replacing “ten days” with “twenty days guaranteed.”

Review question Finding in the fictional case Decision and owner
Does the finish mean the same thing? Ordinary history ends after internal review and deployment; current work also needs external approval before deployment to its intended customers. Analyst Mina separates the histories and records the new finish boundary.
Is the work class comparable? Twenty ordinary items and eight externally approved items describe different flows. Delivery lead Anwar withdraws the ordinary ten-day forecast for this item.
Is the newer sample sufficient for a confident replacement? Eight completed cases provide a rough descriptive signal, not demonstrated calibration or a stable future guarantee. Mina shares 12/20/24-day descriptive points with the small-sample limitation; no precise future probability is asserted.
Are current dependencies resolved? Partner reviewer availability and the specific evidence required remain unconfirmed. Partner contact Ruth obtains the requirements and available review window; receipt of a request is not an approval.
What planning action follows? The launch assumption depended on an internal-only workflow that does not apply. Product owner Cora keeps the external launch date uncommitted and discusses a narrower internal demonstration separately.
What triggers a new review? Scope, approval requirements, capacity, active-item age or observed completion patterns change. Anwar and Mina review the forecast when those changes occur and retain the previous version with its assumptions.

Anwar withdraws the inapplicable ten-day launch assumption and Ruth takes ownership of obtaining the partner-review requirements, while Mina preserves the exploratory comparison. It identifies the failed assumption, preserves the exploratory data and assigns the missing dependency work. The hypothetical internal demonstration is a separate option to assess, not evidence that a customer release or approval has happened.

Maintain the forecast as a decision record

ProKanban's delivery-forecasting guidance emphasizes whether past performance reasonably predicts the current system. More rows do not repair a change in workflow meaning. A narrow-looking distribution can still be irrelevant if the team, item mix, staffing, work-in-progress policy or approval path has changed.

Record the data window, work class, start and finish rules, time convention, quantile method, active dependencies and review date with the forecast. Explain why exclusions are justified. Do not remove difficult completed work solely to obtain a convenient date, and do not repeatedly redefine “started” to reset an old item's age.

When enough future outcomes are observed, compare what actually happened with the forecasts made at the time. Retain missed outcomes and changing assumptions so the review can distinguish a poor method from a changed system. A plausible calculation today is not evidence that the forecast has achieved its stated coverage over future work.

Mina next records Ruth's confirmed partner-review requirements and collects comparable completion records; Anwar keeps the customer launch uncommitted until the dependencies and planning assumptions can be reviewed. LandOffer.ai provides related workplace guides; the immediate benefit here is a reproducible planning conversation that shows uncertainty, the cost of a missed assumption and who will resolve it.

Sources and Further Reading