Most delivery dashboards are rear-view mirrors. Average cycle time, last sprint’s velocity and a burndown chart all describe work that has already finished, which is the work least likely to cause a surprise. When we set out to build Risk Radar, we wanted signals that move before a project slips, not after. So before designing a single chart, we measured which flow metrics carry predictive information, using our own project history as the dataset.
Why Most Flow Metrics Arrive Too Late
The five flow metrics in Portfolio Analytics have precise definitions, and those definitions determine what each metric can tell you about the future:
- Cycle time: elapsed time from an item’s first move into an in-progress state until completion.
- Lead time: elapsed time from creation to completion.
- Throughput: items completed per week, per team or project.
- WIP age: elapsed time since an unfinished item first entered an in-progress state.
- Flow efficiency: active time divided by cycle time, where waiting and blocked states count as inactive.
Cycle time, lead time and flow efficiency can only be computed for items that have finished. By the time an item contributes a long cycle time to your dashboard, the delay has already happened. WIP age is the only one of the five defined for work still in progress, which is exactly where future delays live. Throughput sits in between: it is lagging on its own, but its week-to-week variance says a lot about how far any date built on it can be trusted.
Four weeks before a target date, the share of work older than its team’s P85 cycle time separated slipped projects from on-time ones better than anything on a typical sprint dashboard.
How we tested the metrics
We assembled 187 projects and large epics from our own workspace, delivered between January 2023 and March 2026, each with a committed target date recorded at kickoff. A project counted as slipped if it finished more than five working days after that date. For each project we reconstructed the board as it stood four weeks before the target date and computed every candidate signal from that snapshot alone.
Replaying history from events is the unglamorous part, and it is where most of the effort went. A metric computed from today’s field values leaks the future: a task that was re-scoped last month looks small in hindsight. Rebuilding each board from the task event history removed that leakage, and it changed our conclusions. Two signals that looked strong on current data lost most of their predictive value once computed as of the snapshot date.
Of the 187 projects, 61 slipped. For each signal we measured how well it separated slipped from on-time projects, reported as ROC AUC, where 0.5 is a coin flip and 1.0 is perfect separation:
- Share of WIP older than the team’s P85 cycle time: 0.79
- Scope added after commit, as a share of committed scope: 0.74
- Median age of currently blocked items: 0.71
- Open upstream dependencies with aging WIP: 0.68
- Throughput coefficient of variation, trailing eight weeks: 0.63
- Trailing average cycle time: 0.58
- Sprint velocity in story points: 0.52
Velocity was effectively noise. Practitioners have argued as much for years, but seeing 0.52 on our own data ended an internal debate quickly. Average cycle time was barely better, largely because an average over finished work is dominated by what already went well.
Computing WIP age against the right baseline
Raw WIP age is not comparable across teams. Five days in progress is normal for a data pipeline migration and alarming for a copy change. The useful signal is age relative to the same team’s historical cycle-time distribution for items of the same type. We compute percentiles over a trailing window and compare each open item against them:
WITH history AS (
SELECT team_id, item_type,
percentile_cont(0.85) WITHIN GROUP (ORDER BY cycle_days) AS p85
FROM completed_items
WHERE completed_at >= now() - interval '120 days'
GROUP BY team_id, item_type
HAVING count(*) >= 30
)
SELECT w.project_id,
count(*) FILTER (WHERE w.age_days > h.p85)::float / count(*) AS aging_share
FROM open_wip w
JOIN history h USING (team_id, item_type)
GROUP BY w.project_id;
The HAVING count(*) >= 30 clause does real work. With fewer than 30 completed items, a P85 estimate swings too much to trust. The join above simply drops those item types; in production they fall back to the team’s all-types distribution, and the dashboard marks the baseline as low-sample. Leads told us that label mattered to them more than a marginally more precise number.
Cleaning the event stream before you trust it
Flow metrics inherit every quirk of how teams use their boards. Three corrections changed our results more than any modeling choice:
- Bulk closures. Backlog cleanups close dozens of stale items in minutes and inflate throughput. We exclude closures made in batches of 25 or more within ten minutes when the items had no activity in the prior 60 days.
- Reopened items. A reopened item continues its original cycle rather than starting a new one; otherwise rework looks like fast delivery.
- Working calendars. Ages and cycle times are counted in working days using each team’s calendar and PTO data, so a public holiday does not read as a stalled item.
Survivorship bias deserves its own mention. Cycle-time distributions only include items that finished. Items abandoned after weeks in progress disappear from the history, which makes the baseline look faster than reality. We keep canceled items that spent more than two working days in progress and treat them as censored at their cancellation date, so production baselines come from a Kaplan-Meier estimate rather than the plain percentile in the query above.


From signals to a Risk Radar score
Risk Radar scores every project on four dimensions: schedule slip, scope churn, blocked-work age and dependency risk. Schedule slip combines forecast movement against the committed date with the aging-WIP share, scope churn uses scope added after commit, blocked-work age uses the median age of blocked items, and dependency risk uses open upstream dependencies with aging WIP. Each dimension is normalized to a 0 to 100 scale against the workspace’s own history.
The obvious way to combine them is a weighted average, and an early internal build did exactly that. It hid the projects we most needed to see. A project with stable scope, no blockers and a single critical upstream dependency stuck for three weeks averaged out to a comfortable score. The combination we use now starts from the highest dimension and adds a smaller contribution from the rest, so one severe risk always surfaces while broad moderate pressure still raises the total:
d = [schedule_slip, scope_churn, blocked_age, dependency_risk] # each 0-100
overall = min(100, max(d) + 0.25 * (sum(d) - max(d)) / 3)
# default bands (adjustable): 0-39 steady, 40-69 watch, 70-100 at risk
We tuned the 0.25 coefficient on the 120 projects delivered before 2025 and scored the remaining 67 as a holdout. There, the combination reached an AUC of 0.83, against 0.77 for the best weighted average we tried, and the projects it ranked highest were the ones leads agreed they would have wanted to hear about. Every score in the weekly executive digest appears with the dimension that drove it and the specific items behind it, such as the four tasks older than P85 or the dependency blocked for 16 working days. A number without its items is not something anyone can act on.
What we deliberately do not measure
Every metric in this article is computed per team and per project, never per person. That boundary is written into our AI & Responsible Use Policy, which rules out using planning analytics for automated employment decisions, and it is also what keeps the metrics predictive. Once throughput becomes a personal target, work gets split into smaller tickets, throughput rises and the metric stops predicting anything. Tracking the distribution of item sizes alongside throughput is a cheap way to catch the same effect at team level.
Where to start on Monday
- Stop presenting average cycle time and velocity as delivery indicators. Report percentiles and WIP age instead.
- Compare each open item’s age with its own team’s P85 for the same item type.
- Record committed scope at kickoff so that scope added later is measurable.
- Clean bulk closures, reopened items and calendars before trusting any trend.
- Backtest against your history as it looked on the date, not as it looks now.
Put these signals in front of your portfolio
WIP age baselines, scope churn tracking, the four-dimension Risk Radar score and weekly executive digests are part of Portfolio Analytics & Risk Radar, included in the Enterprise plan (see pricing). Full metric definitions, dashboard options and scheduled exports are described on the Portfolio Analytics & Risk Radar page.