Glossary

The planning vocabulary, defined once

24 terms from the platform’s one metric dictionary and its data model, each in plain words, then what it means in PlanSieve.

A

Aggregation

Rolling values up a hierarchy, items to brands, stores to regions, weeks to months, using the rule each measure declares: quantities sum, prices are averaged or weighted, stock levels take the last period, rates are recomputed from their components rather than averaged. Which hierarchy a value rolls along is a choice, not a given: the same fact can roll up by brand for one measure and by supplier for another.

In PlanSieveEvery measure names its own aggregation rule and the hierarchy it rolls along, so a roll-up is never a guess about what the number means.

B

Bias

The net over- or under-forecast as a share of actual demand, with the sign kept: the sum of (forecast − actual) divided by the sum of |actual|. Above zero the plan ran high; below zero it ran low. Nearer zero is better, and the sign is the whole point: persistent one-sided bias is a model or a process to fix, not luck to wait out. Over- and under-forecasts cancel inside it by design, so a bias near zero can sit on top of a large weighted error; it is read beside WMAPE, never instead of it.

In PlanSieveComputed from the same summed components as WMAPE, at whatever level you pivot to, and shown beside it on every accuracy page, never blended into one number.

C

Closure

The complete set of ancestor-descendant pairs a hierarchy implies: not only that a store belongs to its district, but that it also belongs to the region above and the company above that, at every depth. Materialising the closure is what makes roll-ups, access checks and "everything under this node" questions a single join instead of a recursive walk, and dating it is what lets last year's structure answer last year's questions.

In PlanSieveThe closure is derived from the dated hierarchy edges and kept engine-local beside the facts, so roll-ups, named sets and access grants all cover down through it in one join.

D

Dimension

An axis a business classifies its plan by: product, location, customer, channel, supplier, resource, time, scenario, unit of measure, currency. Each dimension has members (the actual products, the actual stores), attributes on those members, and one or more hierarchies that roll the members up. A plan's shape is the set of dimensions it is built on.

In PlanSieveDimensions are data, not schema. There is no hard-coded item or location table, so adding a channel, a demand stream or a supplier axis is a metadata change with no migration and no downtime.

Disaggregation

Spreading a value set at a higher level down to the finer cells beneath it: a category plan down to items, a monthly number down to weeks. The basis decides how: in proportion to recent sales, by the assortment, evenly, by copying the value to every leaf, or by a rule written for that measure. Done properly it conserves the total, respects the cells a planner has locked, and can say exactly which cells it touched.

In PlanSieveA property of each measure, not of the table: two measures in one group can spread differently, locked cells stay locked, the totals still tie, and the grid shows the exact cells a spread touched.

Disposition

A planner's verdict on a piece of history or a forecast, recorded as a value in its own right rather than as a silent edit: keep it as it is, correct it, mask it out of the model's view, declare it a named and dated event, adjust it, or approve it. The point of recording the verdict is that the engines downstream can read it and act on it, this cycle and the ones after.

In PlanSieveSix verbs behind one write path: a disposition lands as data on the history itself, with a reason and a vintage stamp, so the cleansing engine, the model contest and the next cycle all read it.

F

FVA (forecast value added)

The accuracy gain a step in the forecasting process makes over its input: the weighted error of the step before it minus the weighted error of this step, at a stated lag. The steps are the touchpoints a forecast passes through: a naive baseline, the statistical model, machine learning, consensus, the planner's override. A positive value means the step added value; a negative value means it made the forecast worse. Measuring it is how a process learns which of its steps to keep.

In PlanSieveComputed automatically every cycle over the frozen snapshots, against both the prior step and the naive baseline; a negative value is shown quietly with the number, including when the change was a person's.

G

Grain

The level of detail a set of numbers lives at, stated as one level per dimension: item × store × week is a grain; brand × region × month is another. Every planning number has exactly one grain, and a surprising share of planning mistakes come from adding two numbers whose grains differ without anyone saying so.

In PlanSieveA grain is any combination of dimension × level, with no limit on how many dimensions take part; a measure group is one grain and one physical table, so there is no fixed item-store-week model to bend your business into.

H

Hierarchy

A roll-up path over a dimension's levels: item → brand → category, or store → district → region. A dimension can carry several at once (a product hierarchy for merchandising and another for procurement), and a good one handles the hard cases: alternate paths, ragged depths where some branches stop early, members shared between parents, and dates, so that a re-categorisation applies from when it happened rather than rewriting history.

In PlanSieveHierarchies are master data: multiple, alternate, ragged, shared and effective-dated, and history rolls up along the hierarchy as it stood at the time.

L

Lag

The distance between when a forecast was made and the period it forecast. A lag-3 accuracy figure judges the forecast that was made three periods ahead. Accuracy without a stated lag (and a stated level) is meaningless: a plan is easy to get right the day before and hard a quarter out, and the two must never be averaged together.

In PlanSieveLag is computed by formula over the cycle axis, cycle anchor minus fact bucket, so weighted error, bias and forecast value added are all reported at a stated lag over frozen snapshots.

Level

A tier within a dimension's hierarchy (company, region, store; category, brand, item) down to the leaf, the finest tier at which facts are recorded. A level is the unit a grain is built from, and the altitude at which a policy or a plan can be set and inherited downward.

In PlanSieveLevels are named generically underneath and wear your display names and tooltips on top, so one engine reads a retailer's departments and a manufacturer's plants without changing.

Lineage

The record of how a number came to be: whether it was loaded, edited, computed by a formula or produced by a model run; which version of that formula or model; which inputs it read; and who or what triggered it. Lineage is what lets a planner ask "why is this number what it is?" and get an answer instead of a shrug, and what makes a published cycle reproducible.

In PlanSieveEvery fact value carries lineage, and Explain is a door on it with the same access walls as every other door, so an explanation is never a side channel around security.

M

MAPE vs WMAPE

MAPE averages each item's own percentage error, so a tiny item that missed by 300% counts as much as a top seller that missed by 3%, and an item with zero actuals divides by zero. WMAPE sums the errors before dividing, which weights them by actual volume, so the number describes the portfolio in proportion to what was really sold. That is why accuracy at a category, a region or the whole business is reported as WMAPE, and why a per-item MAPE is at best a diagnostic, never a scorecard.

In PlanSieveThe metric dictionary defines the weighted form once, sums components before it ratios, and refuses the ratio-of-aggregates shortcut (|Σ forecast − Σ actual| ÷ Σ actual), which hides offsetting errors and reads deceptively small.

N

Named set

A saved, reusable definition of which members you mean, "top 100 by revenue in the North region, excluding discontinued", written as attribute rules, hierarchy operations, explicit lists, set algebra over other sets, or conditions on a measure. A dynamic set re-evaluates as the data changes; a snapshot set freezes its membership on the day it was taken.

In PlanSieveOne filter grammar: the same set filters a grid, scopes a forecast run, targets a policy, drives an alert queue and defines who may see the data.

O

OTB (open-to-buy)

The retail money plan's control on buying. Planned receipts for a period are the planned sales plus planned reductions plus the planned closing stock, less the opening stock; what is already on order is subtracted, and the remainder is open to buy. When it turns negative the buyer cancels or marks down. The opening and closing stock are often abbreviated BOM and EOM, beginning and end of month, which has nothing to do with a bill of materials.

In PlanSieveMerchandise-financial measures such as open-to-buy are designed as a measure pack on the same engine: configuration, not a second product; the retail suite is carried by the data model by design and is not listed as available until it ships.

P

Placeholder item

A stand-in product that carries newness in the plan before the real item exists: a launch that has a budget, a launch date and a like-item profile but not yet a code in the master data. When the real items arrive, the placeholder's plan is split across them, so the forecast, the buy and the money were never missing while the product was being finalised.

In PlanSievePlaceholder is the first stage of the lifecycle policy family: placeholder, pre-launch, launch, ramp, mature, phase-out, end of life. And the placeholder split is a switch with a basis, set at the level you choose.

Policy

A named setting that tells the engine how to behave, how to treat outliers, whether to censor stockouts, which launch curve to apply, with an on/off switch and typed options, set once at a chosen level and inherited by everything beneath it, with a single item allowed to differ on the record with a reason. The alternative is behaviour hard-coded in the engine and changed by a support ticket.

In PlanSieveEvery engine behaviour is a policy with layered precedence: package default, tenant default, segment, series pin with a reason. And turning one off measurably skips it.

Q

Quantile forecast (P10 / P50 / P90)

A forecast expressed as a range rather than a single number: the P50 is the median outcome; the P10 is a low case that actual demand should fall below about one time in ten; the P90 a high case it should exceed about one time in ten. Planning consumes the range directly, plan the P50, hedge the receipt at the P90, and the width of the range says how much the model actually knows.

In PlanSieveP10, P50 and P90 come from the model that won the contest and are built by it directly, never as a proportional share applied to a point forecast afterwards.

S

Scenario

A what-if version of the plan you can change without disturbing the plan itself, a promotion pulled forward, a supplier lost, a price moved, then compare, and promote back if you like what you see. A scenario is a mutable exploration; a version, by contrast, is an immutable published snapshot of the plan as it stood.

In PlanSieveA scenario stores only the cells you changed, so creating or deleting one is instant; branch a scenario off a scenario, compare side by side, promote the parts you want, and your permissions follow you inside.

Scope

A saved, multi-dimensional region of the plan, "what am I running on": one selection per dimension (a named set, a list of members or a rule), optionally a time window and conditions on measures. A scope is the argument a run, a report, a rule or an export takes, so the same region can be named once and used by all of them.

In PlanSieveEffective data is always grants ∩ scope: a scope can narrow what a person sees and never widen it, on every door including Explain.

T

Time fence

A boundary in the planning horizon inside which the rules change. A demand time fence marks the near window where only real orders count and the forecast is set aside; a planning time fence marks the window where the system may no longer create new planned orders and a person must decide; a frozen window is locked outright. Fences are what keep a plan from thrashing where it is closest to execution.

In PlanSieveBy design a fence is a per-measure or per-item attribute rather than a system constant, set at the level you choose and inherited beneath it, like every other policy.

Tournament (champion / challenger)

A contest in which several forecasting methods are fitted to the same series' history and scored on a held-out window; the best becomes the champion, the rest remain challengers, and the contest is re-run each cycle so a champion has to keep earning its place as the data changes. It is run per series rather than once for the whole business, because one product's demand is smooth and another's is intermittent.

In PlanSieveSimple methods, classical statistics, intermittent-demand methods and machine learning compete on each series, scored on the one accuracy definition; the winner is named on the grid and the scoreboard is filterable data.

V

Vintage

The forecast as it stood at a particular planning cycle: the promise made in the March cycle, distinct from the promise made in April for the same weeks. Keeping vintages is what makes accuracy honest: a forecast is judged as it was frozen, not as it was later revised, and a planner can see how the picture of one week changed cycle by cycle.

In PlanSieveEach cycle freezes its forecast as a governed procedure, and the accuracy staircase shows every vintage's frozen promise against what happened, by lag.

W

WMAPE

Weighted mean absolute percentage error: the total absolute forecast error as a share of the total actual demand, the sum of |forecast − actual| divided by the sum of |actual| over the series and periods being judged. Because it is a ratio of sums, a high-volume item weighs more than a low-volume one, which is what a portfolio needs: it reads as "the misses added up to this fraction of the demand that was really there". Lower is better, and it is always stated with its lag and its level, or it means nothing.

In PlanSieveDefined once in the metric dictionary and re-derived from summed components at whatever level you pivot to, never an average of the errors underneath, never a ratio of aggregates.

See the vocabulary on your data

Bring your history and your hierarchies. We show every term above as a number on your own plan, with its lag, its level and its framing.

Request a demo

a working demo on your data · no commitment