Monte Carlo Forecasting
Monte Carlo forecasting repeatedly samples historical delivery data to estimate probability ranges for how much work may finish or when a body of work may complete.
Download Markdown (.md)Soroush’s Product Wiki · https://wiki.soroush.uk/topic/monte-carlo-forecasting/
01What it means
Monte Carlo forecasting runs many randomized simulations using observed delivery data. A how-many forecast estimates the number of items likely to finish by a date; a when forecast estimates completion dates for a specified number of items. Results are expressed as probability percentiles, not a single guaranteed date or quantity.[1][2]
- 01Frame one questionHow many by a date, or when a quantity may finish
- 02Select historyRepresentative throughput or cycle-time observations
- 03Run many trialsRepeatedly sample observations under the model
- 04Build outcome distributionPossible dates or quantities and their frequency
- 05Report percentilesProbability statements with data window and assumptions
- 06UpdateRerun as work completes or conditions change
The forecast is conditional on historical representativeness, scope rules, workflow stability, and model assumptions—not a guaranteed commitment.
02Why it matters
Delivery work varies, and a simple average discards that variation. Simulation retains the range found in representative historical observations and makes uncertainty discussable. Stakeholders can compare options such as an 85-percent date versus a 95-percent date, then choose based on risk appetite rather than receiving false precision.
03When to use it
Use Monte Carlo when a team has sufficient relevant cycle-time or throughput history and needs a probabilistic delivery forecast. It suits release, backlog, and service-level questions. Reconsider the input sample when team composition, workflow, item mix, demand, or completion policy changes enough that old data no longer represents the future.
04How it works
- Frame one clear question: how many items by a date, or when a defined number of items may finish.
- Select a representative historical data set, preserving low-output periods and legitimate outliers.
- Run thousands of trials by repeatedly sampling throughput or cycle-time observations according to the model.
- Summarize outcomes as percentile dates or quantities and state the data window, assumptions, and scope rules.
- Update the forecast as work completes or conditions change, and pair it with explicit risk decisions.
05A practical example
A team has 60 comparable migration items remaining and 20 weeks of throughput data. Simulation suggests completion by November 6 at 70 percent confidence, November 20 at 85 percent, and December 4 at 95 percent. Leaders choose the 85-percent planning date and fund a fallback rather than presenting November 20 as certain.
06Where it breaks
Simulation cannot rescue unrepresentative or poor-quality data. It assumes sampled history is relevant and usually does not model every dependency, scope change, item-size shift, or rare disruption. Confidence percentages describe the model's outcomes, not certainty about reality. Cherry-picking recent good weeks produces deceptively optimistic forecasts.
Common mistakes
- Feeding estimates into a model while presenting the output as evidence from actual delivery performance.
- Removing zero-throughput periods or inconvenient outliers solely to improve the forecast.
- Communicating one percentile date as a promise without assumptions, alternatives, or ongoing updates.
07Sources
- Forecasting ResourcesProKanban.org · Professional resource collectionForecasting Resources
- Probabilistic Forecasting and Flow with ScrumScrum.org · Professional white paperProbabilistic Forecasting and Flow with Scrum