Three-Step Forecast Build
Part of the Planning Week chain · previous: Anatomy of a ROC Standard · next: Forecast Lock Process
Three-Step Forecast Build is the forecasting section of a function's standard, filled: the long-range, mid-term and short-term forecasts as three steps, each consuming the lock of the step above it, each built in the same three layers (a baseline from the engine, a driver adjustment only where a driver has left its historical pattern, an event adjustment from the collision calendar), and each recorded as a vintage. It matters because a function whose three horizons are built by different people on different definitions cannot say which layer or which horizon missed, and so cannot improve any of them. The page produces section S-3.1.1 of the Standard on the ten-block template (template SS) and is worked in a session with Wiki:Packs/ROC Standard Authoring (CP-OPS-003).
The section is the first the chain fills, on Day 2 afternoon, as the strategic pattern beside the real-time one on Incident Management for Contact Centers. It fills the ten blocks that Anatomy of a ROC Standard defines and restates nothing the forecasting pages it cites already carry: the method families are on Forecasting Methods, revision governance on Reforecast and Rolling Forecast Methodology, the vintage column on Forecast Vintages, the annual rhythm on Workforce Planning Calendar and Annual Planning Cycle, and the daily loop an agent team runs on The Short-Term Forecasting Loop with an Agent Team.
The vocabulary is the chain's: a node is a place work can sit: a hub (in-country depth), a service center (badged arbitrage at scale), a partner (speed, market access, flex), or automated; a node owner is the owner of the body of work placed at a node, never the operator of the node; a seat is a post on the org chart (a director-graded or manager-graded box, its scope and its sunset condition); the capacity cycle is the six-stage monthly cycle of Capacity Planning Cycle, in which clearing is the stage and netting the computation inside it.
Three steps, one build
The three steps are horizons. The long-range step produces the annual envelope (volume, handle time, shrinkage and the staffing they imply, by month and by book) on the long-term capacity engine, and its output is the plan the budget is submitted against. The mid-term step produces the month, inside the envelope, on the capacity cycle's clock, and its output is the locked plan of record that recruiting and finance execute against. The short-term step produces the week and the day by interval on the short-term WFM engine, inside the locked month, and its output is the forecast the schedule of record is built on and the intraday steer is measured against.
Each step consumes the one above. The mid-term build opens by reading the annual envelope; the short-term build opens by reading the locked month. A step may redistribute inside what it was handed and may not exceed it; a build that breaches its envelope is not a forecast but a variance, and it travels through the door Forecast Lock Process describes. This is the rule that makes the three horizons one forecast rather than three opinions.
Inside every step the build has the same three layers. The baseline is the engine's time-series output on the register's definitions. The driver adjustment applies only to a driver that has fallen outside its historical pattern; a recurring driver already inside the history (a billing cycle, a weekly release) is in the baseline and adjusting for it again counts it twice. The event adjustment applies each event on the Forecast Collision Calendar with its effect window and its grade. The layers are the same at every horizon; what differs is the engine, the resolution and the clock. The judgmental layers are governed as Reforecast and Rolling Forecast Methodology governs them: a documented reason for every adjustment above a stated threshold. The threshold is not there because adjustment is bad but because the adjustments that damage accuracy are the small, unexplained ones, and because adjustments made upward are the ones least likely to help.[1]
The horizon and regressor treatment follows Hyndman and Athanasopoulos;[2] the adjustment discipline follows Fildes and colleagues;[1] the vintage follows Croushore and Stark;[3] the three horizons are the practitioner's.[4]
The ten blocks, filled
Purpose
Produce, at each of three horizons, the forecast of volume, handle time and shrinkage by book and channel that the next horizon and the schedule of record are built on, and score each horizon against what arrived so that the miss is attributed to a layer and a horizon rather than argued over. Accuracy is treated here as a property of the process that addresses a miss rather than of the number itself, and the vintage is the instrument that makes that treatment auditable.
Inputs
The lock handed down from the step above; actuals by channel and interval, reconciled, from the data core (interface row: Operations); the collision calendar and the event ledger (interface rows: Operations, Technology, Commercial); the driver refresh from the capacity cycle's stage 1; the definitions register, one definition per metric, cited on every number.
Outputs
A versioned forecast per horizon with its assumption register, its lineage and its vintage; the lock, published on the calendar Forecast Lock Process describes; the loaded forecast file for the short-term WFM engine; the variance decomposition (volume, handle time, mix, supply) when a register row is open.
Roles
The forecaster builds; the planner who owns the book approves at the gate; the standardization lead owns the method for every book; the node owner consumes the forecast and owns the decision it feeds. Where an agent team runs the short-term step, the roster and the gates are those of The Short-Term Forecasting Loop with an Agent Team and nothing here changes except who moves the steps.
Method
The L1 flow is one page, thirteen steps, four of them decisions phrased as questions. It runs once per horizon on that horizon's clock: annually, monthly at the cycle's stage 1, weekly and daily for the short-term step.
| # | Step | Owner | Next |
|---|---|---|---|
| 1 | Open the build for the horizon due | forecaster | 2 |
| 2 | Read the lock handed down from the step above (the envelope, or the locked month) | forecaster | 3 |
| 3 | Pull the baseline from the engine for this horizon, on the register's definitions | forecaster | 4 |
| 4 | Has any driver fallen outside its historical pattern? | forecaster | yes → 5 · no → 6 |
| 5 | Apply the driver adjustment; name the driver; register the assumption with its grade | forecaster | 6 |
| 6 | Read the collision calendar for this horizon's window | forecaster | 7 |
| 7 | Apply each event's adjustment with its effect window and grade | forecaster | 8 |
| 8 | Does the build breach the envelope handed down? | forecaster | yes → 9 · no → 11 |
| 9 | Raise a variance through the lock door of the step above; hold this build until it is answered | forecaster | 10 |
| 10 | Does the owner of the horizon above accept the breach? | the step above's owner | yes, the envelope moves → 11 · no → 3, the build is rebuilt inside the envelope and the refusal is recorded |
| 11 | Version the forecast with lineage and vintage; register every assumption | forecaster | 12 |
| 12 | Planner gate: approved? (no unlabeled carried assumption; grades present; definitions cited) | the planner | yes → 13 · no → 3, reason recorded |
| 13 | Publish the lock to the step below and to every interface on the lock calendar | standardization lead | End |
The L2 step table decomposes each step on the fifteen columns of Process Decomposition (L0–L3); it runs about forty rows for this section and is produced by the pack, not reproduced here. The job aids it names are the export procedure, the reconciliation checklist, the vintage retrieval, the decomposition method and the control limits.
Metrics
Forecast accuracy by layer (baseline, driver, event) and by horizon, read from vintages; the share of revisions carrying a registered assumption; days to detection of a level shift on the short-term step. Accuracy by layer belongs to Level 3; detection of a regime break before the trend shows it belongs to Level 4.
Tools
The long-term capacity engine for the long-range step; the short-term WFM engine for the short-term step and the loaded file; the Python analytics platform for the decomposition and the control chart; the data core; the event ledger.
Controls
Every number cites its definition and carries a grade; nothing carried across a channel, platform or cohort change is presented as measured, per Human Gates and Number Grades; the planner gate on any version that changes a plan; the acceptance gate the section passed under Process Standardization Lifecycle.
Maturity
Level 1: a single inherited number. Level 2: each heritage forecasts on its platform's field definitions and reports one accuracy figure. Level 3: three steps on one definition set, accuracy by layer, vintages kept, yesterday scored daily. Level 4: the agent team runs the short-term step and regime detection is automated; the planner approves. Level 5: reforecast on trigger within governance bounds.
Node attribute and vendor paragraph
All nodes, per book. At a partner supply seat (a partner capacity block: one supplier, one body of work, one commercial form) the lock carries the billable view Forecast Lock Process describes, a partner-held forecaster post is transitional until automated, and its output is scored on the same instrument and the same vintages as a badged post's.
Worked example
The chain's function, in the June 2026 cycle, the first on the single calendar decision D-08 set. The mid-term step for July opened Mon 1 Jun 2026, read the annual envelope, and locked at pencils-down on Tue 9 Jun 2026 (BD+7). The short-term step for the week of Mon 6 Jul 2026 was built on Wed 24 Jun 2026 as version 2026-06-24.v1 and locked Fri 26 Jun 2026. Its baseline for the migrating book carried voice handle time at 452 seconds [M] from the demand ledger; the pre-migration 412 seconds remained in the version's lineage relabeled "carried" [A], as the March versions 2026-03-05.v1 through 2026-03-25.v1 had set. At step 4 one driver was outside its pattern: the partner cohort's ramp, registered as an assumption at [E]. At step 6 the calendar carried one event in the window, a release on Thu 9 Jul 2026 with a two-day effect window [A]. At step 8 the build was inside the locked month. Scored in August, the week's volume miss was 1.4 percent [C] and the handle-time miss 0.6 percent [C], read from the vintage, and attributed: the event layer over-adjusted, the baseline did not. The numbers are the example's, not a measured result.
The artifact this page produces
Section S-3.1.1 on the ten-block template, template SS. One filled example row of the section register:
| ID | Section | Class | Owner (seat) | Blocks filled | L1 steps | L2 rows | Gate status | Skeleton until |
|---|---|---|---|---|---|---|---|---|
| S-3.1.1 | §3.1.1 Three-step forecast build | strategic; short-term both | standardization lead | 10 of 10 | 13 | about 40 [E] | peer review | Fri 25 Sep 2026 |
Produced in a working session with Wiki:Packs/ROC Standard Authoring (CP-OPS-003); the filled set is part of blueprint v0.1.
What would change this
The page claims that building three horizons as steps that consume each other, on one definition set with vintages, lets a miss be attributed to a layer and a horizon, and that the attribution is what improves the forecast. If a function ran the section for four cycles and the vintage record showed the attribution changing nothing (the same layer missing by the same amount each cycle with no revision to that layer's method), the record would be an audit trail and not a learning instrument, and the section's Purpose block would have to be rewritten to say so. Two cycles are not enough to see it; four are.
How this connects
- Previous in the chain: Anatomy of a ROC Standard — the template this section fills
- Next in the chain: Forecast Lock Process — the door a build uses when it breaches its envelope, and the calendar step 12 publishes to
- Also in §3.1: Forecast Collision Calendar — the third layer's source
- Defers to: Forecasting Methods (the method families) · Reforecast and Rolling Forecast Methodology (revision cadences, triggers and the adjustment threshold) · Forecast Vintages (the as-of column and its storage grain) · Workforce Planning Calendar and Annual Planning Cycle (the annual rhythm the long-range step sits in) · The Short-Term Forecasting Loop with an Agent Team (the daily loop and its roster)
Maturity Model Position
The section as written is a Level 3 practice on the WFM Labs Maturity Model™: one definition set, accuracy by layer, vintages kept. A Level 2 function adopts it by issuing the section as a skeleton and filling the short-term step first. Four scales on this wiki use the word "level"; Planning Week for a Workforce Function states which is which. This page uses the documentation levels for the L1 flow and the L2 table, and the maturity Levels 1–5 in the Maturity block.
See Also
- Planning Week for a Workforce Function
- Anatomy of a ROC Standard
- Forecast Lock Process
- Forecast Collision Calendar
- Capacity Planning Cycle — the clock the mid-term step runs on
- Human Gates and Number Grades — the number grades used on this page
References
- ↑ 1.0 1.1 Fildes, R., Goodwin, P., Lawrence, M., & Nikolopoulos, K. (2009). Effective forecasting and judgmental adjustments: an empirical evaluation and strategies for improvement in supply-chain planning. International Journal of Forecasting, 25(1), 3–23. doi:10.1016/j.ijforecast.2008.11.010 — across more than 60,000 forecasts in four supply chains, the relatively larger adjustments tended to improve accuracy while the smaller ones often damaged it, and upward adjustments were much less likely to improve accuracy than downward ones, suggesting a bias toward optimism.
- ↑ Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.). OTexts. https://otexts.com/fpp3/ — the forecast origin, the horizon, and the treatment of holidays and events as regressors (chapters 5 and 7).
- ↑ Croushore, D., & Stark, T. (2001). A real-time data set for macroeconomists. Journal of Econometrics, 105(1), 111–130. doi:10.1016/S0304-4076(01)00072-0 — the vintage as the forecast that was current at the time, and evaluation against the data then available.
- ↑ Cleveland, B. (2012). Call Center Management on Fast Forward (3rd ed.). Colorado Springs: ICMI Press. ISBN 978-0-9854611-0-2 — the long-, mid- and short-term forecasting horizons as practitioners run them.
