Forecasting in Scrum and Agile: Velocity, Throughput and Monte Carlo Simulation
"When will it be done?" is the most common question a Scrum Master hears, and the one most often answered badly. Teams either refuse to answer ("Agile doesn't do dates"), which loses stakeholder trust, or give a single confident date that turns out to be fiction. A better approach is to forecast: give an evidence-based range along with the assumptions behind it. This lesson covers three practical methods, velocity-based forecasting, throughput-based forecasting and Monte Carlo simulation, and explains when to use each. It builds on the definitions in Scrum Metrics Explained.
Forecast vs Commitment
A commitment is a promise; a forecast is a probabilistic statement about the future. In Scrum, the Sprint Goal is a commitment; the number of items the team will finish is a forecast. The items the Developers select for a Sprint are a forecast of what can be done; the Sprint Goal is what the team commits to. Treating forecasts as promises is a major source of dysfunction.
Method 1: Velocity-Based Forecasting
Velocity is the amount of work, usually measured in story points, a team completes in a Sprint. To forecast a release:
- Total the size of the remaining Product Backlog items for the target release, say 240 points.
- Look at the team's recent velocities, say 28, 32, 25, 30, 35 and 27 over the last six Sprints.
- Use a range, not an average: the lowest realistic velocity (around 25) and the highest (around 35), perhaps the average (about 30).
- Divide: 240 / 35 is about 7 Sprints; 240 / 30 is 8 Sprints; 240 / 25 is about 10 Sprints.
- Present: "We expect to finish in 7 to 10 Sprints, most likely around 8."
Limitations
- It depends on a stable team and a consistent estimating scale.
- Story points are subjective; they can drift or inflate, especially when velocity is used as a target.
- It assumes that backlog size is fixed; in reality scope grows.
- Comparing velocity across teams is meaningless. See Common Scrum Myths.
Method 2: Throughput-Based Forecasting
Throughput is the number of items completed per time period, for example per week or per Sprint. It ignores size entirely: you simply count finished items. This works surprisingly well when items are of broadly similar size, which good story splitting encourages.
- Count the remaining items: say 60 stories.
- Look at recent throughput per Sprint: 9, 12, 8, 11, 10, 13.
- Use lower, typical and higher: about 8, 10 and 13.
- Divide: 60 / 13 is about 5 Sprints; 60 / 10 is 6; 60 / 8 is about 8.
- Present a range: "5 to 8 Sprints, most likely 6."
The advantage is that it needs no estimation sessions, which saves time and avoids debates about points. The disadvantage is that a mix of very large and very small items will distort results.
Method 3: Cycle Time Percentiles (For Single Items)
When someone asks, "How long will this one item take?", use cycle time history. Sort the cycle times of recently finished items and read off percentiles. For example, if the 50th percentile is 4 days, the 85th is 9 days and the 95th is 15 days, you can say: "Items like this are finished within 9 days about 85 percent of the time." This is a service-level expectation, a common practice in Kanban. See Kanban for Beginners.
Method 4: Monte Carlo Simulation
Averages hide variation. Monte Carlo simulation addresses this by randomly sampling from your actual history thousands of times to produce a probability distribution of outcomes.
How it works for "how many items by a date?"
- Take historical throughput samples, say per week: 3, 5, 2, 4, 6, 3, 4, 5.
- For a four-week horizon, randomly pick four values (with replacement) and add them. That is one simulated outcome.
- Repeat 10,000 times.
- Sort the totals. The 85th percentile (the value that 85 percent of outcomes meet or exceed) is a reliable planning number.
How it works for "when will we finish N items?"
- For each trial, keep sampling weekly throughput until the sum reaches N.
- Record the number of weeks needed.
- After thousands of trials, the distribution shows, for example, 50 percent chance by week 9, 85 percent by week 11, 95 percent by week 13.
Many free spreadsheets and tools implement this, and a basic version is possible in a spreadsheet with random sampling. The key skill for a Scrum Master is not the mathematics but communicating results: "We have an 85 percent chance of finishing by the end of March, and a 50 percent chance by mid-March."
Comparing the Methods
| Method | Needs estimates? | Best for | Main weakness |
|---|---|---|---|
| Velocity | Yes (story points) | Teams already using points with stable estimating | Point drift; subjective |
| Throughput | No (counting items) | Teams with similarly sized items | Distorted by very uneven item sizes |
| Cycle time percentiles | No | Single-item expectations; service levels | Needs enough history |
| Monte Carlo | No | Communicating probabilities and risk | Garbage in, garbage out if history is unrepresentative |
Making Forecasts Trustworthy
- Use real history, ideally the last six to twelve Sprints, excluding periods that are no longer representative such as a major team change.
- Express uncertainty, with ranges or percentages rather than single dates.
- State assumptions: stable team, no major scope growth, no new dependencies.
- Update regularly. A forecast is a living view and should be refreshed every Sprint.
- Show scope change. Use a burnup chart so additions are visible. See Sprint Tracking in Scrum.
- Do not turn forecasts into targets. If velocity or throughput becomes a goal, it will be gamed.
A Short Conversation Template for Stakeholders
- Scope: "The current backlog for this release has about 60 items."
- Evidence: "Over the last 8 Sprints we finished 8 to 13 items per Sprint."
- Forecast: "That gives us a likely range of 5 to 8 Sprints, with 85 percent confidence of finishing within 8."
- Risks: "New scope, the vendor API and holidays could change that."
- Options: "If the date is fixed, we can reduce scope or protect the team from other work."
This approach connects with planning at the roadmap level; see Agile Release Planning and Product Roadmaps.
Common Mistakes
- Using average velocity alone and ignoring variation.
- Forecasting with too little history, such as two Sprints.
- Including items of wildly different sizes in throughput counts.
- Reporting a forecast as a deadline.
- Forgetting that work that is not Done does not count toward velocity or throughput.
- A forecast is a probabilistic range based on evidence, not a promise.
- Velocity needs estimates; throughput just counts items; Monte Carlo shows probabilities.
- Use ranges, state assumptions and refresh forecasts every Sprint.
- Do not use forecasting metrics as targets.