How do you build a Monte Carlo project forecast?

WHAT THIS TESTS: This tests your ability to move beyond simple estimates (like story points or averages) and apply a statistical model to project forecasting. The interviewer is looking for three things: 1) your understanding of what data is valid input (cycle time, not effort estimates), 2) your grasp of the simulation process itself, and 3) your ability to translate a complex statistical output into a simple, actionable forecast for non-technical stakeholders. It's a test of data literacy and communication as much as technical knowledge.

A GOOD ANSWER COVERS: A strong answer walks through four key steps. First, define the inputs: a set of remaining work items (e.g., 50 stories) and a historical dataset of cycle times for similar, completed work items (ideally 20+ data points). Second, describe a single simulation run: for each of the 50 stories, randomly select a cycle time from the historical data and sum them to get one possible total project duration. Third, explain the scaling: repeat this simulation 1,000 to 10,000 times to generate a distribution of possible completion dates. Fourth, describe the output: present the results as a histogram or a probability distribution, answering questions like "What date gives us an 85% confidence of completion?" or "What's the probability of finishing by Q3?".

COMMON WRONG ANSWERS: A major red flag is using story points as the input for the simulation. Story points are a measure of relative effort or complexity, not duration, and are notoriously unreliable for time-based forecasting. Another common mistake is using the average cycle time instead of randomly sampling from the full distribution of historical data; this ignores the variability, which is the entire point of the simulation. Finally, weak answers provide a single date as the output. The value of a Monte Carlo simulation is that it produces a range of possibilities with associated probabilities, not another single-point estimate.

LIKELY FOLLOW-UPS: Be ready for "How would you handle different types of work, like bugs vs. features, that have different cycle time profiles?" (Answer: segment the historical data and sample from the appropriate distribution for each work item type). Another is "What do you do if you have no historical data for a new type of project?" (Answer: start with expert estimates to create an initial distribution, but state the high uncertainty and prioritize collecting real data as soon as possible).

ONE CONCRETE EXAMPLE: Imagine we have 40 stories left in a project. We look at the last 100 stories our team completed and find their cycle times ranged from 1 to 15 days. For one simulation run, we randomly pick 40 cycle times from that historical data (e.g., 3 days, 1 day, 8 days, ...) and sum them. Let's say it totals 210 days. We repeat this 10,000 times. We then count how many results fall on or before certain dates. If 8,500 of our 10,000 simulated totals are 240 days or less, we can tell our stakeholder: "We have an 85% confidence we will finish within 240 days."