Skip to content
Flow metrics & forecasting

Why Your Delivery Estimates Miss (and How Forecasts From History Fix It)

By Sebastiaan Jansen · · 9 min read

A software delivery forecast is a date range built from how long similar work has actually taken, not a fresh opinion about the next ticket. Single-number estimates miss because they hide variation: blockers, review waits, and scope that turns up mid-build. If a team records start and finish the same way every time, that history becomes a range for what will likely be done by a given day.

Why a single date keeps missing

Most missed dates come from the shape of the estimate. Someone asked "when will this ship?" and got one day back, as if the work had one duration.

The same kind of ticket can finish in two days or in twelve, depending on reviews, interruptions, and who picks it up. A point estimate picks one of those outcomes and calls it the plan. When a slower outcome arrives, the date looks wrong even though the team did ordinary work.

Three habits make this worse.

  • Estimating the happy path. The number assumes the ticket is understood, the reviewer is free, and nothing in production interrupts the week.
  • Padding in private. Each person pads their own piece, the pads stack, and the date still misses, because each pad was a guess about a guess.
  • Replacing the last miss with a bigger number. The next estimate absorbs the last surprise and ignores the tickets that finished on time, so the average moves and the spread stays where it was.

Software estimation still helps when you are choosing between options that don't exist yet, or when a ticket is unlike anything in the backlog. Once you have finished work to look at, it is a weak way to answer "which week does this land?"

What a software delivery forecast uses

A software delivery forecast asks a different question. Given how we have been finishing work, how many items will likely be done by this date, or which date is likely for this many items?

It needs four facts:

  1. When work started (left the queue and someone began it).
  2. When work finished (merged, deployed, or accepted, using one definition).
  3. How many items finished in each period. That count is throughput.
  4. What was still open at the end of each day, if you also want to see where time is sitting.

Cycle time is the elapsed time from start to finish for one item. Throughput is how many items finish in a week. Cycle time shows the spread for a single item; throughput shows how fast the system clears a pile. Throughput and cycle time can move in opposite directions, so reading only one of them can leave you with a confident wrong date.

Fifty to a hundred completed tickets of a similar type is a practical start. With fewer, keep the range wide on purpose. A new product line, or a team that just doubled, is not "similar." Say so, and treat the first run as a sketch.

Percentiles before simulations

Before you simulate anything, sort recent cycle times and read the percentiles.

If the median (50th percentile) finishes in 4 days and the 85th percentile in 11, half the work is done by day 4 and a noticeable share is still open on day 11. Plan the release to the median and you will be late about half the time.

Quote the 85th percentile when you commit, and use the 50th as the day you check whether something is stuck. The gap between the two is the most useful number here. When it's tight, the system is predictable. When it's wide, any single date will keep failing until the process gets steadier.

Lead time and cycle time run on different clocks. Lead time includes the wait before anyone starts. If stakeholders feel that wait, forecast with lead time. If you are staffing people already on the work, use cycle time. Mix the two and the range looks either lazy or magical.

Project forecasting from throughput

For a batch (a release, a cycle, a client milestone), project forecasting is usually a throughput question. You know how many tickets are in the batch and how many finished per week lately, and you want a date.

Use the observed weekly counts. Suppose the last eight weeks finished 8, 11, 6, 9, 10, 7, 12, and 9 tickets. That pace already includes the slow weeks. For 40 remaining tickets, the best recent week (12) clears them in just over 3 weeks, a typical week of about 9 takes roughly 5, and the slow week (6) takes about 7.

Saying "about 3 to 7 weeks" is more honest than "done on the 20th." It still understates the risk, because the next eight weeks won't replay the last eight, and slow weeks tend to cluster. A cumulative flow diagram shows whether the batch is growing while you forecast it. If arrivals outrun finishes, the 40 will not stay 40.

A worked example

These numbers are an illustration, not a benchmark.

A team has 36 tickets left for a marketing launch. Throughput over the last ten weeks was 5, 7, 6, 8, 4, 7, 6, 9, 5, and 7. None of those weeks was a holiday shutdown, so all ten stay in.

Sample those weeks with replacement a few thousand times, and count how many weeks each sample needs to reach 36 finishes. This shows how often each possible future comes up if the coming weeks look like recent ones.

In this illustration, about half the samples clear 36 items in 6 weeks, about 85 of 100 clear them in 8, and a thin tail needs 10, usually because several draws of 4 landed in a row.

If today is 6 October and the count starts next Monday, the 85 percent date falls in early December. That is a different answer from "six weeks from now" given as a single day. Marketing books the launch against the 85 percent date and uses the earlier date to check that the batch hasn't grown. What you hand over is a range, the chance you're quoting, and a rule for what happens when the remaining count rises.

The example rests on two assumptions: the 36 tickets are about the same size as the historical ones, and the team won't take on a large second batch in the same weeks. If either fails, rerun the numbers.

Falrow runs this from ticket history: cycle time percentiles, a Monte Carlo simulation of 2,000 runs, and cumulative flow so you can see whether the batch is stable. A spreadsheet can do the same job with the same history, so the method matters more than the tool. Sampling delivery dates explains why an average hides the weeks that blow a launch.

How fresh the history should be

Monte Carlo here means repeated draws from your own completed work. A few thousand runs is enough to settle the percentile. Adding more runs won't sharpen the date. Steadier work and a shorter queue will.

Use the last 8 to 12 weeks when the team and the ticket mix are steady. After a re-org or a new definition of done, start the window at the change. Split ticket types that move at different speeds. A one-day bug and a two-week feature need separate samples, and the critical path has to respect the slower class.

When the range is too wide to use

If you can't honestly promise a day, work on the system that produces the spread.

  • Finish older items before starting new ones. A lot of work in progress stretches cycle time and makes weekly throughput jump around.
  • Cap how many items are in progress, then watch whether the 85th percentile moves in.
  • Forecast the first slice that can ship on its own. With fewer items, a slow week has fewer chances to ruin the date.
  • Remove tickets that are in the release only because they happened to be nearby.

If the business needs one finished item every two days and your 85th percentile cycle time is nine days, the range will keep landing late until the pace or the demand changes. Deployment frequency tells you how safely you ship. It says nothing about when a 36-ticket batch will clear.

Where judgment still belongs

History can't see an unrecorded dependency, a vendor who has gone quiet, or a design people are still arguing over. Put those on the ticket as blockers with a named owner.

Be careful about the sample. Drop weeks when the office was shut. Drop cancelled tickets too, or throughput looks healthier than the product is. Keep rework if rework is normal for the team.

The Scrum Guide treats the sprint backlog as the developers' forecast of what they believe they can do, and expects that picture to change as they learn more. Publish the range, name the chance you are quoting, and replace it when the count or the pace changes. A date that can no longer be revised has become a promise.

How to say the date

Put the percentile in one sentence: "We are 85 percent confident this batch finishes by 4 December, based on the last ten weeks, if the ticket count stays at 36." Then name what would move it: new scope, two slow weeks, or a blocker on the oldest items.

Keep that assumption next to the date, so when the date moves you can see which assumption broke. Update a month-long batch once a week, and again when ten or more tickets are added. Marketing plans on the later date and keeps a smaller announcement ready in case the earlier one hits. The roadmap uses the same remaining-ticket count, so everyone quotes the same chance.

The charts and the 2,000-run sample can live with the tickets. The insights view puts cycle time, cumulative flow, and the forecast in one place. Launch planning that starts from that date is the handoff described for marketing teams.

Start this week

  1. Apply one definition of done to the last ten weeks and count finishes per week.
  2. Note the 50th and 85th percentile cycle times for those same items.
  3. Count only the tickets the next date actually requires.
  4. Publish the item count, the history window, both dates, and the assumption that the count will not grow.

Check it weekly. If you renegotiate it every day, the range turns back into a guess.

FAQ

What is a software delivery forecast?

It is a range of dates, or a range of completed items, taken from recently finished work. It uses throughput, cycle time, or both, and it states how often similar history would have hit that date. It describes the system you have now, so replace it when the system or the batch changes.

How is this different from software estimation?

Software estimation assigns a size or a duration up front, usually from judgment or story points, before the work is underway. A forecast looks back at completed tickets and projects that pace forward. Estimation helps when you have no comparable history. Once you have it, forecasting is the better default.

How much history do you need?

A practical start is 8 to 12 recent weeks, or about 50 to 100 finished items of a similar type. A shorter window still produces a number, but keep the range wide. After a major change in team, tooling, or the definition of done, drop the weeks from before the change.

How often should you update it?

Update a live batch once a week, and again when the remaining count jumps or a week falls well outside recent throughput. The 85 percent date only means something while the assumptions next to it still hold. If you never rerun the range, it turns back into a fixed promise.

More on flow metrics & forecasting