Upload data

All intake paths — CSV upload, ERP connector, or SFTP — normalize into one canonical schema. The forecasting engine only sees this schema.

Required table: orders

One row per order line. Do not pre-aggregate; the engine builds monthly (customer_id, product_id) series.

Column	Required	Notes
customer_id	Yes	Stable over time — series key
product_id	Yes	SKU — stable over time
order_date	Yes	Any parseable date; normalised to month
ordered_qty	Yes	Forecast target (units)
revenue	Recommended	Enables value-based analysis
unit_price	Recommended	Demand driver; derivable from revenue ÷ qty
currency	Optional	Defaults to workspace currency
location_id	Optional	Site or region dimension

Minimum to run: the four required columns.

Minimum to be useful: required columns plus revenue or unit_price.

Optional enrichment tables

product_master

Unlocks hierarchy, readable labels, and cold-start detection.

Column	Unlocks
product_id	Join key
product_name	Dashboards and reports
brand, category, sub_category	Hierarchical reconciliation
launch_date	Cold-start handling
price_tier	Attribute-based priors

customer_master

Column	Unlocks
customer_id	Join key
customer_name	Readability
segment	Segment-aware modelling (Growth+)
region, country	Geographic seasonality

events

Promotion, stock-out, or shutdown calendars (FC-008). Masks or down-weights affected periods during training.

Column	Notes
event_type	promotion, stockout, price_change, shutdown, other
start_date, end_date	Inclusive month range
scope	all, product_group:…, or series:customer_product
mask_strategy	exclude or downweight

Data readiness scorecard

After upload, AccuPredix surfaces a readiness score (also on the dashboard as data quality score):

Check	Blocked	Limited	Good
History per series	cold-start only	< 12 months	≥ 24 months
Overall history	< 3 months	3–11 months	≥ 24 months
Recency	> 6 months old	2–6 months	within 1 month
ID stability	IDs in disjoint windows	some churn	stable keys
Required columns	any missing	no revenue/price	complete

Rules of thumb

• 24+ months of history gives the best seasonality learning.
• 12 months is the practical minimum for trend.
• Keep customer_id and product_id stable — re-coded IDs silently break history.
• Sparse / intermittent demand is expected in B2B; specialist models handle it.

Upload workflow in the app

1. Workspaces → select workspace → Upload.
2. Drop CSV/XLSX or use a connector sync (Growth+).
3. Review validation messages and readiness badges.
4. Optional: upload product_master, customer_master, events in the same session.
5. Proceed to Run forecast when readiness is not blocked.

ERP and connector mapping

For SAP, NetSuite, Dynamics, and similar systems, map ERP field names to the canonical columns in the connector setup wizard. The engine is identical whether data came from CSV or live sync.

See Integrations for connector-specific steps.

Template and synthetic data

Developers can generate reference datasets with the engine CLI:

cd model && uv run accupredix generate-data

This emits tables matching the canonical schema and serves as a downloadable template.