tommymorgan Flow Metrics Guidance
Apply these principles from Daniel Vacanti's framework whenever discussing flow metrics, forecasting, or predictability.
Core Principles
Never use averages for forecasting
- The Flaw of Averages: "plans based on average fail on average"
- Cycle Time data is NOT normally distributed — it is skewed right
- An average hides the variability that matters most for prediction
- Instead: use percentiles (50th, 70th, 85th, 95th)
Percentiles ARE forecasts
- "85th percentile = 12 days" means "there is an 85% chance this item will finish in 12 days or less"
- No estimation required — the historical data speaks for itself
- Percentiles are robust against outliers (unlike averages)
Every forecast must include three elements
- A range of possible outcomes (not a single number)
- A probability for that range
- A time-to-live — when should this forecast be regenerated?
Work Item Age is the most important metric
- Age = how long an in-progress item has been in the system
- The real reason to control WIP is to prevent unnecessary aging
- Items aging beyond the 85th percentile SLE warrant immediate investigation
- Two best ways to prevent aging: finish items, and don't start items you're not ready for
Size does not matter for forecasting
- There is almost no correlation between Story Point estimates and actual Cycle Time
- You do NOT need same-sized items for flow metrics to work
- Estimation usually makes processes LESS predictable, not more
- Stop estimating. Start measuring.
When someone asks "when will it be done?"
For a single item:
- Use Cycle Time percentiles from historical data
- Present as: "Based on our data, there's an 85% chance this will finish within {n} days"
- Improve the forecast by: reducing WIP, managing Work Item Age, improving flow efficiency
For multiple items:
- Use Monte Carlo simulation on historical Throughput data
- Present as: "There's an 85% chance all {n} items will be complete by {date}"
- NEVER divide remaining items by average throughput — that's the Flaw of Averages
Anti-patterns to correct
| If you hear... | Respond with... |
|---|---|
| "It will be done by {date}" | "What's the probability? What range of dates is possible?" |
| "Our average cycle time is X days" | "What are the percentiles? The 85th is more useful for setting expectations." |
| "We need to estimate this in story points" | "Story points don't correlate with actual delivery time. Use historical Cycle Time data instead." |
| "Velocity says we can do X points per sprint" | "Throughput (items completed per week) is more reliable than velocity for forecasting." |
| "This sprint we did fewer points so we're slower" | "Is the change within normal variation? Check the XmR chart before reacting." |
Reforecast when new information arrives
- Every forecast has a shelf life
- As work progresses, new information validates or invalidates assumptions
- Reforecasting is not a sign of failure — it's responsible practice
- The NHC updates hurricane forecasts every 8 hours. You should update delivery forecasts regularly too.