Monte Carlo Forecasting

Traditional planning: Estimate each item, sum the estimates, commit to a date.

Why it fails:

• Individual estimates are unreliable (optimism bias)
• Errors compound (many estimates → big errors)
• Single-point estimates ignore variability
• Nobody tracks whether estimates were accurate

The result: Dates that nobody believes, pressure to "commit," and erosion of trust.

Monte Carlo offers a different approach: use historical data to simulate many possible futures, then report probabilities.

Basic concept:

1. Collect historical data: How many items did we complete each week for the past N weeks?
2. Simulate many futures: Randomly sample from history thousands of times
3. Aggregate results: What percentage of simulations finished by each date?
4. Report probabilities: "There's an 85% chance we'll finish by March 15"

Example:

Historical throughput (items/week): 5, 8, 6, 4, 7, 6, 9, 5

To forecast completing 20 items:

• Simulation 1: Week 1 draws 6, Week 2 draws 5, Week 3 draws 4, Week 4 draws 8 = done in 4 weeks (6+5+4+8=23)
• Simulation 2: Week 1 draws 4, Week 2 draws 4, Week 3 draws 5, Week 4 draws 5, Week 5 draws 4 = done in 5 weeks
• ... (repeat 10,000 times)

Result: "85% of simulations completed within 4 weeks. 95% within 5 weeks."

Monte Carlo gives you confidence levels, not single dates.

How to read results:

• "50% confidence: March 10" — Coin flip; might happen, might not
• "85% confidence: March 17" — Pretty likely, reasonable to commit
• "95% confidence: March 24" — Very likely, buffer for almost anything

Which confidence level to use?

Depends on the consequences of being wrong:

• Low stakes: 50-70% confidence is fine
• Normal business: 85% is common choice
• High stakes: 95% or higher

The honest conversation: "If you need 85% confidence, the date is March 17. If you need 95% confidence, it's March 24. Which matters more—the earlier date or the certainty?"

Question: "When will these 20 items be done?"

Method:

1. Count items to complete: 20
2. Get historical throughput data (last 8-12 weeks)
3. Run Monte Carlo simulation
4. Report date at desired confidence level

Example output:

• 50% confidence: Complete by Week of March 4
• 85% confidence: Complete by Week of March 11
• 95% confidence: Complete by Week of March 18

This is a "When" forecast — fixed scope, variable date.

Question: "How much can we deliver by March 31?"

Method:

1. Calculate weeks until deadline: 4 weeks
2. Get historical throughput data
3. Simulate throughput for 4 weeks, many times
4. Report item count at desired confidence level

Example output:

• 85% confidence: At least 20 items
• 50% confidence: At least 25 items
• 15% confidence: At least 30 items

This is a "How Many" forecast — fixed date, variable scope.

The planning conversation: "By March 31, we're 85% confident we can complete at least 20 items. If you want 30 items, we'd need to either extend the date or add capacity."

Getting started:

1. Track throughput: Count completed items per week for at least 6-8 weeks
2. Use a tool: Spreadsheets, Jira plugins, specialized tools (ActionableAgile, Nave, etc.)
3. Start simple: Raw throughput without decomposition
4. Report honestly: Share confidence intervals, not false precision

Improving forecasts:

• More historical data = better forecasts
• Stable processes = tighter ranges
• Decompose work into similar sizes
• Track accuracy and refine

What Monte Carlo doesn't solve:

• New types of work with no history
• Major process changes (past ≠ future)
• Small sample sizes (garbage in, garbage out)
• Items that depend on external factors

For those situations, use judgment alongside the data.