9 min

Moving from Estimates to Forecasts Grounded in Real Delivery Data

Distinguishing estimates, forecasts, targets, and commitments to build delivery projections backed by observed history and explicit assumptions.

Answering "when can we reasonably expect to deliver this scope?" does not require abandoning estimation conversations. It simply requires separating two uses: estimating to understand and slice individual items, and relying on observed past deliveries to project a timeline. That separation makes planning trade-offs far clearer, provided you state your assumptions explicitly and account for variability.

The essentials in 30 seconds

  • Estimation and forecasting answer different questions and can coexist: one clarifies an individual item, the other projects a collective timeline.
  • Delivery history offers a factual starting point under explicit assumptions (comparable scope, slicing granularity, and working conditions).
  • An average alone does not describe uncertainty: a useful forecast reflects period-to-period spread and is updated as work progresses.

1. Estimate, forecast, set a target, and commit

Even when all 30 tickets in an upcoming milestone have been discussed and estimated — in Story Points, relative sizes, or days — adding up those estimates does not directly yield a delivery date. Between an item’s effort and its release sit the real dynamics of the system: collective completion pace, waiting times in review or QA, unplanned incidents, and scope adjustments. Keeping estimation conversations to surface unknowns and split work while grounding timeline projections in delivery history prevents a single number from carrying four conflicting roles.

Separating four frequently conflated concepts immediately clarifies roadmap conversations:

ConceptQuestion answeredDefinition and basisPrimary purpose
Estimate“What is the rough order of magnitude of this item (effort, size, complexity, or duration)?”Approximate assessment of a quantity based on available information. Relative sizing discussed in a team is one useful practice for sharing understanding and surfacing unknowns.Clarify the need, uncover technical unknowns, and split oversized items.
Forecast“What outcome or timeframe can we anticipate, and under what assumptions?”Projection of a future outcome based on available data and explicit assumptions. This guide focuses on forecasts grounded in actual delivery history and their probabilistic expression.Inform trade-offs with transparent scenarios rather than a single date presented as certain.
Target / Goal“What business outcome or date do we want to reach?”Product ambition, commercial deadline, external event, or strategic milestone.Provide shared direction and guide scope trade-offs if the forecast diverges from the desired date.
Commitment“What objective or responsibility do we choose to pursue together?”Explicit decision to pursue a shared goal with clear conditions and trade-offs. A commitment does not remove uncertainty: in the Scrum Guide, the team commits to a Sprint Goal, not to delivering a frozen scope of tickets.Focus collective effort and navigate trade-offs calmly when discoveries emerge mid-flight.

2. Start from observed deliveries

Throughput counts the number of tickets genuinely completed (status classified as "Done") over a comparable period. Unlike an upfront estimate, it records an accomplished fact: each counted item has traveled through the entire workflow to completion.

Indicator in CaubaDisplayed unitRequired dataWhat it contributes to forecasting
ThroughputItems / week (normalized)Status + (Sprint or Completed Date)Measures real delivery pace in Scrum and Kanban and feeds the backlog forecast history.
Velocity — "Count completed items" (default)Items / sprintStatus + Sprint (with start/end dates)Shows raw completed volume per iteration on the same ticket basis as Throughput.
Velocity — "Sum estimates"Story Points / sprintStatus + Sprint + Story PointsSums completed points per sprint; the only indicator in Cauba that reads the Story Points column.

A single period is too narrow to judge a delivery pace, so you need several complete, comparable periods. In Cauba, throughput analysis and backlog forecasting require at least 3 usable periods (by default the 3 most recent bounded sprints in iteration mode, or the 6 most recent complete work weeks in calendar mode). A full period with 0 completed items is kept as 0.0, whereas an ongoing trailing calendar period is excluded so it does not artificially drag down the observed rate.

Counting tickets does not require every ticket to be identical in size: across 20 to 50 items, smaller and larger items balance out as long as overall slicing remains consistent (without mixing Epics and Sub-tasks). Finally, neither throughput nor velocity measures business value or individual developer output — both are collective flow indicators.

3. Example: forecasting the delivery of 30 items

Consider a fictional team running 2-week sprints (14 calendar days) that wants to forecast how long it may take to deliver a batch of 30 tickets, based on its last 6 sprints:

PeriodDaysCompleted itemsCauba Throughput (items / wk)Observed context
Sprint 11484.0Steady pace, two items still in code review at sprint close.
Sprint 214126.0Completion of prior review items alongside current sprint work.
Sprint 31463.0Major production incident absorbed part of the team’s capacity.
Sprint 414105.0Return to the team’s typical delivery pace.
Sprint 514147.0Well-sliced batch of tickets with minimal external dependencies.
Sprint 614105.0Median pace with no unusual disruptions.

What the data shows (6 sprints)

  • Total delivered: 8 + 12 + 6 + 10 + 14 + 10 = 60 completed tickets across 6 sprints (12 weeks).
  • Average pace per sprint: 60 / 6 = 10 tickets per sprint, ranging from 6 to 14 tickets depending on sprint events.
  • Average weekly throughput (Cauba formula): 5.0 tickets per week.

Explicit assumptions for projecting the 30 items

  • Remaining scope: 30 tickets sliced at a granularity comparable to the 60 historical items, with no uncompensated net additions.
  • Comparable conditions: same team and same sprint duration (14 days).
  • Operational reality retained: past swings — including a disruption like Sprint 3 — remain possible going forward.

From the average to a probabilistic forecast

Average-based benchmark: dividing 30 tickets by 10 tickets per sprint gives an order of magnitude of 3 sprints (roughly 6 weeks). Yet that average does not guarantee finishing in 3 sprints: if the next three sprints resemble Sprints 1, 3, and 4 (8 + 6 + 10 = 24 tickets), 6 tickets will still remain for a 4th sprint.

Probabilistic reading: when simulating thousands of possible sequences drawn with replacement from these 6 observed rates, roughly 59% of scenarios finish the 30 tickets in 3 sprints or fewer — placing the median horizon (P50) at 3 sprints — and nearly 98% finish within 4 sprints. Because sprints are counted in whole iterations, both the 85% and 95% thresholds fall on 4 sprints here. Keep in mind that these results remain conditional on the assumptions above and are never a guarantee: planning around 3 or 4 sprints depends on how much delay risk the milestone can absorb.

To dive deeper into how sampling works and how to interpret P50, P85, and P95, read our guide to forecasting backlog delivery with Monte Carlo (full calculation verification details for this example appear at the end of section 5).

4. Put the approach into practice

Adopting delivery-grounded forecasting requires neither a reorganization nor dropping estimation if it helps refinement. Four steps are enough to establish a clear habit:

Step 1 —Define the target scope
Count the remaining tickets belonging to the milestone without double-counting parent containers (Epics) alongside their child items, and note whether additional discovery work is expected.
Step 2 —Verify historical data quality
Confirm that only genuinely delivered items sit in "Done" (classify cancelled items as "Discarded"), check that resolution dates or sprint boundaries are filled in, and gather at least 3 to 6 comparable periods.
Step 3 —Share the forecast alongside its assumptions
Communicate a median horizon and a more prudent horizon rather than a single date, explicitly stating the counted scope, historical periods used, assumed conditions, and risk margin chosen.
Step 4 —Refresh as deliveries happen
At the end of every sprint or week, update both the delivery history and the remaining item count: a useful forecast evolves as actual scope and pace unfold.

5. Limitations and going further

A forecast grounded in past data loses relevance whenever the near future stops resembling the observed history. Five situations call for explicit caution:

Changes in team staffing or availability
Onboarding, departures, holiday seasons, or partial reallocation to another product change upcoming capacity.
In practice: State this capacity gap explicitly in your assumptions instead of mechanically applying full-team throughput.
Shift in work nature or slicing granularity
Moving from routine incremental enhancements to a complex technical overhaul alters typical ticket size.
In practice: Review the delivered work-type breakdown and slice upcoming batches to a comparable granularity before projecting dates.
Unusual external dependencies or blockers
An atypical wait on a third-party vendor or regulatory sign-off falls outside past team pace alone.
In practice: Track those blockers separately: a historical model never resolves an organizational bottleneck.
Ongoing scope growth
A forecast computed for 30 tickets becomes outdated if 15 items are added without an equivalent trade-off.
In practice: Recompute the forecast whenever the remaining item count changes materially.
Short history or uneven, unbounded periods
With only 1 or 2 periods, or when mixing sprint lengths without start and end dates, dispersion cannot be gauged reliably.
In practice: Map sprint boundaries in Cauba and wait for at least 3 complete usable periods.

Do not confuse individual Cycle Time with a Backlog Forecast

In Cauba, Cycle Time percentiles (P50 and P85, computed once at least 10 completed items have both start and completion dates) are expressed in whole calendar days for a single ticket. Backlog Forecast percentiles (P50, P85, and P95) are expressed in sprints or 7-day periods for an entire batch of remaining items. See also the Cycle Time vs Lead Time guide.

Verifying the example calculations

Under uniform random draws with replacement from the 6 throughputs [8, 12, 6, 10, 14, 10] for a 30-item backlog, exact combinatorics yields 0% completion in 2 sprints (maximum 14 + 14 = 28), 58.8% in 3 sprints (127 combinations out of 216), and 98.5% in 4 sprints (1,277 out of 1,296). A 10,000-run Monte Carlo simulation reproduces those frequencies (~59% at 3 sprints and ~98.5% at 4 sprints). In the Cauba engine, calculate_monte_carlo_from_history followed by calculate_percentile (linear interpolation cast to integer) returns P50 = 3 sprints, P85 = 4 sprints, and P95 = 4 sprints.

Forecasting Backlog Delivery with Monte Carlo

Understand how Monte Carlo backlog simulations work, what P50, P85, and P95 horizons actually mean, and how to use them without false certainty.

Preparing a Jira Export to Analyze Software Delivery Flow

Recent-activity and sprint-scoped JQL queries, columns to export, and classifying Cauba’s 4 status categories.

Cycle Time vs Lead Time: Definitions, Formulas, and Distributions

Learn how to measure individual ticket durations (in whole calendar days) and interpret their P50 and P85 percentiles.

Sprint Spillover & Delivery Predictability

Analyze the systemic causes of sprint carryover (WIP, slicing, interruptions) and sprint scope dynamics.

Frequently Asked Questions on Estimates and Delivery Forecasting

No. If Story Point conversations help your team uncover unknowns and split work during refinement, keep using them for that purpose. In Cauba, you can choose between "Count completed items" (the default) and "Sum estimates" for Velocity, while Throughput and the Backlog Forecast operate on completed ticket counts.

Methodological References & Further Reading

  • Daniel S. Vacanti — When Will It Be Done? Lean-Agile Forecasting to Answer Your Customers' Most Important Question (2018). Foundational reference on separating estimation from probabilistic forecasting grounded in historical throughput.
  • Ken Schwaber & Jeff Sutherland — The Scrum Guide (2020). Presents the Sprint Goal as the commitment for the Sprint, distinct from the selection of Sprint Backlog items which remains adaptable.
  • Scrum.org & ProKanban — The Kanban Guide for Scrum Teams (2020). Joint use of Throughput and Cycle Time to inform planning in Scrum teams.
  • Troy Magennis — Forecasting and Simulating Software Development Projects (2011). Application of Monte Carlo simulations to historical throughput and scope evolution risks.