PM Predictive Analytics (G9 -- velocity, forecasts, risk)
An in-development predictive analytics layer for PM boards: velocity tracking over a configurable window that also says when the open work finishes at this pace, Monte Carlo completion forecasts, what-if scenarios with resource, priority, deadline and cut-scope levers whose risk change is stated in days, and risk heatmaps with cascading cross-board impact.
On this page
What it is
In-development predictive analytics layer: velocity tracking, Monte Carlo completion forecasts, what-if scenarios, and risk heatmaps.
Watches how fast work moves through PM boards and produces four kinds of predictions:
- Velocity tracking -- counts completed items over a configurable window
(default 14 days). Falls back to done-status column values when no lead-time
observations exist. Trend uses 7-day buckets with linear regression to classify
direction as increasing, decreasing, or stable.
"Completed" is decided from each status cell's label, which resolves the same way the bottleneck
risk below does: the stored text if the cell has it, otherwise the label its own column gives the
cell's selected index -- so a board whose status cells were imported without text still counts its
finished work. It also counts an item finished by MOVING it into a done group. See
pm-cross-board-intelligence.md, "Reading a status".
The Velocity History table and
GET /pm/predictive/velocitylist one saved snapshot per day (the latest of that day), so a board with months of history returns a short list rather than every scheduler tick (SPM-BUG-001). - "At this pace" line -- a velocity answer says what the pace means, not only the number:
"At this pace, the open work finishes around early December." It divides the board's open
items (the shared
isFinishedItemSqlrule) by the pace that answer already shows, and names the early (days 1-10), mid (11-20) or late (21+) part of the finish month, adding the year when it is not this year. It appears on the Velocity Trend card, the Velocity History section, the forecast write-up ("has 9 remaining items at a pace of 0.2 items/day ... At this pace, ..."), and AI answers about pace ("7 items finished in the last 7 days and the pace is steady, and at this pace the open work finishes around late November").GET /pm/predictive/velocity/trendand thepm:predictive:velocity:trendchannel return it aspace: { openItems, perDay, finishDate, summary }. It isnull-- never a guessed date -- when nothing is open, nothing finished recently, or the finish would be more than 1095 days away. - Deadline line in the forecast write-up -- the write-up warns about a deadline only when the work due by some date is more than the recent pace can finish by then: "Warning: at the recent pace, about 14% of the open items due by 11/7/2026 will finish in time. Open items due by then: 21." It checks every upcoming due date (overdue items count first) and names the weakest one, using the same calculation and percentage as On-time confidence, so one item due tomorrow on a big project no longer raises a warning. Items with no due date never count against a date; with no upcoming due date it says "Can't tell yet whether the deadlines will be met". In a what-if, a faster or slower pace is weighed, added items carry no date, and removing items says "can't tell yet" when the date is still short, because the what-if does not say which items go.
- Days start at the user's local midnight -- the velocity window, the weekly trend buckets and the forecast's day-by-day finish counts all start each day at local midnight, so a task finished at 8 PM counts on that day, not the next (BUG-SIM-077; before this they started at UTC midnight, 7 PM in Central time).
- Monte Carlo simulation -- runs 1000 simulations per board. WHICH model it uses depends on
what is being asked, not on which data happens to exist. "Remaining" everywhere below means the
board's open work under the shared
isFinishedItemSqlrule -- the same one the risk heatmap and the overdue scanner use -- so an item finished by MOVING it into a done group is not remaining, and a board with no status column at all is still counted correctly (BUG-SIM-070).- More than one item remaining -- simulates the board day by day from how much work it
actually finishes in a day (its finish history over the last 90 days, needing at least 7
finishes), reported as
confidenceSource: 'throughput'. This is a throughput question, so a board that finishes several items a week is forecast as finishing several items a week. Each finish is dated by WHEN IT HAPPENED -- the day the status changed, or the day the card moved into a done group -- so a board whose data was just re-synced still reports the days its work really landed (BUG-SIM-073). An item that was already finished when it was imported carries no such record and is left out of the rate rather than dated to the sync. - Exactly one item remaining -- samples that item's own lead time, from empirical data
(>= 10 observations), a triangular distribution from best/typical/worst estimates (< 10
observations), or a normal distribution centred on the board's velocity when neither exists
but work is completing (
confidenceSource: 'velocity-derived'). Extracts p50/p85/p95 completion dates and a deadline-hit probability. The deadline is the active sprint's end date, or the latest due date among the remaining work, or nothing at all -- never a single item's near-term due date for a multi-item forecast.
- More than one item remaining -- simulates the board day by day from how much work it
actually finishes in a day (its finish history over the last 90 days, needing at least 7
finishes), reported as
- What-if scenarios -- four lever types: resource (velocity multiplier),
priority (item reorder), deadline (day shift) and scope (the number of open items cut or
deferred; a negative value adds items). Scope scales the remaining work by what is left:
cutting 2 of 5 open items leaves 3/5 of the time, cutting all of them finishes now, and it
combines with the resource lever. The What-If dialog's "Items to cut or defer" slider (0-20)
sends it. Every impact also carries
riskSummary, the risk change in plain words and days: "7 more days of breathing room before the deadline", "3 fewer days of breathing room before the deadline", "Finishes 4 days later" when the project has no deadline, or "No change to the finish date or the deadline" under half a day.riskDeltastays the same number in days (positive = closer to or past the deadline). The dialog shows the scenario board's sentence above the table and on each risk badge's hover. Cascades impacts to related boards via parent-item relationships only: a board counts as related when it holds child items of the scenario board's items. A lead-time setup on another board never makes it related. A scenario names only the boards you are a member of, the scenario board included; a board you cannot see is left out of its impacts and cascade lines entirely. The answer is capped at 20 impact rows and 20 cascade lines: the scenario board plus its first 19 related boards, ordered by board id. Only those boards are simulated, so a hub board with dozens of dependents no longer costs one forecast per dependent. The payload carriesimpactsTotal(the true number of impacted boards, from aCOUNT(*)beside the capped query) andtruncated, and the dialog says "Showing the first 20 of N impacted projects" when the cap bit. A badly shaped request is refused with a message that names the wrong field and spells out the body it accepts:leversis a list of 1 to 5{"type": "resource" | "priority" | "deadline" | "scope", "value": <-365 to 365>}. - Risk heatmap -- four dimensions per board, each clamped to [0, 1]:
overdue risk (0.3 weight), velocity risk (0.25), bottleneck risk (0.25),
dependency risk (0.2). Overall = weighted sum. Bottleneck risk is the largest
share of items sitting in any one status, so it reads each item's status as its
column's label — a status cell stores an index into that column's label list,
not the text, and the index is resolved against the column it came from (see
pm-cross-board-intelligence.md, "Reading a
status"). Finished items are not counted — the metric measures
concentration among work still to do, and counting finished ones inverted it
(BUG-SIM-037: a board with 10 of 12 items Done scored 0.833). Exclusion uses
the shared
isFinishedItemSqlrule, so an item in a done group and an item whose status reads as done are both excluded. With done no longer forming a second bucket, a board whose open work all sits in ONE status is judged by the lone-pile rule: if that one status is Stuck or Blocked, the board reads as fully bottlenecked (1), because a board where every open item is stuck is the worst case the metric exists to show; if it is anything else (for example every open item "Working on it"), it reads 0, since everyone being at the same stage is not a pile-up. A Stuck or Blocked pile also counts for more than its size — it is weighted 1.5x when the largest pile is chosen, so ten items Stuck outscores ten items "Working on it" (which reads the same as it always did). "Blocked" is the sharedBLOCKED_STATUS_KEYWORDSlist, matched exactly and case-insensitively, so a label like "Blocked pending review" is not counted as blocked. The score still stays between 0 and 1, and a board with nothing stuck is unchanged. - The bottleneck explains itself -- open a project's row and, under the Bottleneck bar, one plain sentence names the cause: "6 of 10 open items are sitting in “Review”.", or for stuck work "3 of 8 open items are stuck in “Stuck”, and stuck work counts extra." A status with no label says so instead of naming one. A bottleneck of 0% shows no sentence.
- Edits show at once -- after you change an item, the next time the heatmap (or the Mission Control AI, or the CLI) reads that project it is scored from the new data, not from a cached score. The one exception: a different project whose Dependency bar waits on the changed work can take up to five minutes to catch up.
- The heatmap never invents a score -- a project that has never finished anything shows "Not enough data yet" with no percentage and sits in its own group below every scored project. A project that finished work before but nothing in the last five weeks is "stalled" (pace 100%), and one with no open work has pace 0%.
- Heatmap filters -- a workspace dropdown, a timeframe (Next 7 days -- the default -- Next 30 days, or All open work) and an "At risk only" switch. All three are remembered between visits and cleared when you sign out.
Where to find it
In-development, gated behind the pm-predictive-analytics unreleased feature
(pmPredictiveAnalyticsEnabled). Background computation is gated. The Risk Heatmap,
Risk Heatmap shows in Omni's Operations view and in Stats → Mission Control →
Omni. The deadline forecast (the simulation one) shows in Stats → Mission Control →
Projects for the board you pick, next to Velocity History. Stats → Mission Control → Overview shows
on track and at risk from the heatmap's own ratings and a team pace.
On a paired phone and on the web, Stats → Mission Control shows the same Overview, the AI Project Manager heatmap, and a Projects tab with the velocity trend and the deadline forecast for one project you pick. Velocity History and the other project charts stay on the computer, and the Projects tab says so in one line. A phone or web view only ever includes projects you are a member of, in workspaces that have an owner. The server decides this, so nothing the phone sends can widen it, and if it cannot tell who is a member it shows nothing. Being an admin does not widen it on the phone either. The computer shows exactly what it showed before.
How it behaves
Risk heatmap filters, timeframes and business days
- Workspace -- "All workspaces" shows every project your account can see; choosing a workspace shows only its projects. A workspace with no name record shows as "Unnamed workspace".
- Timeframe -- "Next 7 days" scores each project only over its open work due on or before seven days from today, including work that is already late. Items with no due date and finished items are left out, and a project with nothing due is left off the list with a count underneath ("3 projects have nothing due in this timeframe"). "Next 30 days" works the same way; "All open work" is the original whole-project view.
- What each bar means inside a timeframe -- Overdue is the share of that work already late; Bottleneck is how piled up that work is in one status; Pace becomes "Pace vs. what's due": can the project finish what is due at its recent pace (for example 10 items due in 7 days at 1 finished a day scores 30%); Dependency is the share of that work waiting on unfinished work.
- Business days -- Settings -> Mission Control -> "Count my week as business days only" turns the timeframes into "Next 5 business days" and "Next 20 business days" (an even four work weeks) and judges pace per weekday, so weekends never count as working time. The card shows "Counting business days (Mon-Fri) - Change", and Change opens the setting.
- Company holidays -- with the Lab switch "Mission Control company holidays" on, each project's company holidays are skipped as well (see "Company holidays" in pm-cross-board-intelligence.md). A holiday never shortens a window: "Next 5 business days" is still five working days, so a holiday inside it pushes the window's last day later. Because projects can close on different days, each project's window ends on its own day. The card says "through <date>" when every project's window ends the same day, and "end date varies by project" when they differ; a row whose end date differs from the others shows its own end date. A holiday is also left out of the working days that pace is divided by, so a short holiday week never makes a team look slow. With the switch off, nothing changes from before.
- At risk only -- hides every project under 40% (the same line where a score turns amber) and every "Not enough data yet" project, and says how many it hid with a "Show all" link.
- Never a dead end -- an empty result says why ("No projects in this workspace yet.", "Nothing is due in this timeframe.", "No projects are at risk right now.", or "N projects couldn't be scored right now.") and offers "Clear filters", which returns to All workspaces, All open work, everything shown.
- Freshness -- scores are reused for up to five minutes, so switching filters is instant; an edit can take up to five minutes to show.
Results are cached in the G2 snapshot store and refreshed on a three-tier scheduler: hot (on board sync), warm (15 minutes), cold (24 hours with 90-day velocity data retention).
When the computer is busy, the warm refresh may be skipped, but never more than
two times in a row: under steady load every other refresh still runs, so a
forecast is refreshed long before its cached copy expires. A refresh that runs
on a busy computer leaves the cleanup of old velocity rows for a quieter moment.
GET /pm/predictive/status reports consecutiveLoadSkips, the number of
refreshes skipped in a row right now (0 once one runs).
Lead-time setups
A lead-time setup says how long one status-to-status transition normally takes on a board, and the estimates above are computed against it. Setups can be added and removed:
- Remove one from the Lead-Time Analytics table in a board's analytics — each row
has a delete control that asks first and names the exact transition. Or over the CLI:
DELETE /pm/intelligence/lead-time/config/:id. - Removal is permanent. There is no in-app undo, so the confirmation is the only safety rail and it says so plainly.
- Nothing measured is destroyed. The setup stops counting toward that board's estimates and dependency risk, and its in-flight measurements stop, but every already-completed measurement keeps its row and its record of how long it took.
- Removing a setup that is not there, or one already removed, answers "not found" rather than reporting a success.
- Whoever can add a setup on a board can remove one; a view-only member can do neither.
Related
The widgets this layer produces are mounted in the Ops Engine panel, documented in pm-operations-engine.md, and Omni's advisor turns the same completion forecasts into plain-language narratives — see pm-intelligence-advisor.md. Its cross-board view of the data is described in pm-cross-board-intelligence.md. The full list of library pages is in INDEX.md.
Last verified 2026-10-24