Demand Planning is where ERP discipline either begins or breaks.
Statistical Demand Forecast Generation looks operational from far away. In a real finance team, it is a chain of assertions: the right actor started the work, the required records existed, the control policy was applied, the state change was preserved, and the outcome can be explained later without rebuilding the transaction from emails and spreadsheets.
The expected business outcome is specific: Forecast MAPE below 15% for A-class SKUs; planner override rate below 20%; supply plan generated within 1 hour of forecast publish
The control flow a finance team actually needs.
Step 1
At Minimum 6 Statistical Models With...
Step 2
Allow Manual Override With Reason Code...
Step 3
Publish Forecast At Configurable Bucket...
Step 4
Maintain 24-Month Frozen Actuals For...
Step 5
Track Forecast Accuracy Per SKU Per Period
The ERP surface involved.
Module
Demand Planning
Actors
Demand Planner, Forecast Engine (ML service), ERP Scheduler
Tier
Tier 3
Finance area
Supply Chain, Demand Planning & Logistics
Region lens
US and UK finance teams
Publication date
May 12, 2026
Support at minimum 6 statistical models with automatic model selection by lowest MAPE; allow manual override with reason code and audit trail; publish forecast at configurable bucket granularity (daily/weekly/monthly); maintain 24-month frozen actuals for backtesting; track forecast accuracy (MAPE, WMAPE, bias) per SKU per period; processing of 500,000 SKU-location combinations must complete within 4 hours; overrides must be version-controlled
US and UK teams have different compliance hooks, but the same control problem.
US teams usually care about clean evidence for audit support, vendor records, payment controls, tax reporting, and management review. UK teams usually care about VAT-ready records, approval evidence, digital-record discipline, and traceable postings. The country-specific details differ, but the operating pattern is the same: the ERP needs controlled records, explicit ownership, defensible state changes, and evidence that survives beyond the person who completed the task.
The control matrix.
| Control area | Requirement | Acceptance proof |
|---|---|---|
| Control 1 | Support at minimum 6 statistical models with automatic model selection by lowest MAPE | Given 24 months of cleaned sales history for SKU-location combinations |
| Control 2 | allow manual override with reason code and audit trail | when the forecast engine runs nightly |
| Control 3 | publish forecast at configurable bucket granularity (daily/weekly/monthly | then forecast quantities at weekly buckets for a 13-week horizon are written per SKU-location with model selection by lowest MAPE, seasonality indices applied, and outliers flagged in an exception report |
| Control 4 | maintain 24-month frozen actuals for backtesting | negative) when fewer than 6 statistical models are available or MAPE exceeds 15% for A-class SKUs without planner override then the system must flag exceptions and block consensus publication. |
| Control 5 | track forecast accuracy (MAPE, WMAPE, bias) per SKU per period | Forecast MAPE below 15% for A-class SKUs; planner override rate below 20%; supply plan generated within 1 hour of forecast publish |
| Control 6 | processing of 500,000 SKU-location combinations must complete within 4 hours | Forecast MAPE below 15% for A-class SKUs; planner override rate below 20%; supply plan generated within 1 hour of forecast publish |
Audit evidence is a chain, not a folder.
| Evidence layer | What should be preserved |
|---|---|
| Business event | The ERP scheduler triggers the forecast engine nightly using 24 months of cleaned sales history per SKU-location combination. The engine selects the best-fit statistical model (Holt-Winters, ARIMA, or Croston for intermittent demand) based on error minimization, applies seasonality indices and promotional lift factors, and writes consensus forecast quantities at weekly buckets for a 13-week rolling horizon. Demand planners review outliers flagged by the system via a waterfall exception report and override individual SKU forecasts before publishing. The approved forecast is published to the supply planning module and made available for S&OP review. |
| Control rules | Support at minimum 6 statistical models with automatic model selection by lowest MAPE; allow manual override with reason code and audit trail; publish forecast at configurable bucket granularity (daily/weekly/monthly); maintain 24-month frozen actuals for backtesting; track forecast accuracy (MAPE, WMAPE, bias) per SKU per period; processing of 500,000 SKU-location combinations must complete within 4 hours; overrides must be version-controlled |
| Acceptance proof | Given 24 months of cleaned sales history for SKU-location combinations; when the forecast engine runs nightly; then forecast quantities at weekly buckets for a 13-week horizon are written per SKU-location with model selection by lowest MAPE, seasonality indices applied, and outliers flagged in an exception report; (negative) when fewer than 6 statistical models are available or MAPE exceeds 15% for A-class SKUs without planner override then the system must flag exceptions and block consensus publication. |
| Data record | |
| System event | |
| Lifecycle state | |
The useful version of this workflow is not only fast. It is inspectable. A controller, auditor, or operator should be able to move from source event to system record to state transition to final business outcome without guessing.
Implementation contracts.
Reference data model
`demand_forecast` { forecast_id: string, sku_id: string, location_id: string, bucket_start: date, bucket_granularity: enum(DAILY/WEEKLY/MONTHLY), quantity: int64, model_used: enum, mape: numeric, override_quantity: int64, override_reason_code: string, override_by: string, status: enum(DRAFT/OVERRIDDEN/PUBLISHED), external_id: string }; `forecast_accuracy` { period: date, sku_id: string, mape: numeric, wmape: numeric, bias: numeric }; (reference, product may differ).API and events
`POST /v1/demand-forecasts/runs` { sku_ids, location_ids, horizon_weeks, bucket_granularity } -> 202 { run_id, status: RUNNING }; `GET /v1/demand-forecasts/runs/{run_id}`; `PUT /v1/demand-forecasts/{forecast_id}/override` { override_quantity: int64, reason_code: string } -> 200 { forecast_id, status: OVERRIDDEN }; `POST /v1/demand-forecasts/{run_id}/publish` -> 200; emits `demand.forecast.published` event; idempotent via `external_id`.State transitions
`RUNNING -> COMPLETED -> REVIEW`; planner overrides move lines to `OVERRIDDEN`; `REVIEW -> PUBLISHED`; terminal `ARCHIVED`; guard: publish blocked if any A-class SKU MAPE exceeds threshold without explicit override and reason code.Common implementation traps.
Treating the workflow as data entry
If the ERP only stores the final record, the team loses the decision trail that explains how the record became valid.
Hiding exception logic
Exceptions need owners, reason codes, and time stamps. A vague pending state is not a control.
Posting without recovery design
Retries, duplicate submissions, and partial failures must be explicit so the system does not create inconsistent records.
Skipping evidence design
A workflow that cannot produce evidence on demand will eventually push finance teams back into manual screenshots and spreadsheets.
Where Rivane fits.
Rivane is built for finance workflows where automation must stay tied to source documents, approvals, state transitions, ledger impact, reporting, and audit evidence. Use this guide as a checklist for evaluating whether an ERP workflow is merely digitized or actually controlled.
References and source basis.
These sources provide the standards, regulatory, or government context around the flow. They are included so the guide is useful to finance operators, auditors, and implementation teams, not only buyers reading software copy.