What this is
The demand forecasting model: the components of a retail demand forecast, the parameter structure that drives accuracy, and the process for reviewing, overriding, and feeding forecast output into replenishment and financial planning.
Purpose
Forecasting exists to enable continuous product availability by projecting demand before it occurs, not by reacting to stockouts after the fact. An accurate forecast drives replenishment quantities, sizes purchase commitments against actual expected demand, and feeds the financial planning process with a demand basis for OTB. A poor forecast — one that ignores seasonality, promotional impact, or new-item dynamics — produces either excess inventory or out-of-stocks, both of which destroy margin.
Structure
Forecast components — A retail demand forecast is built from four independently computed and selectively combined components:
| Component | What it captures |
|---|---|
| Basic demand | Baseline sales velocity from historical sell-through, adjusted for trend |
| Seasonal demand | Cyclical sales pattern: peak weeks, off-peak weeks, seasonal uplift factors |
| Promotional demand | Expected sales lift during defined promotional events; isolated so it does not corrupt baseline |
| Shrink factor | Expected loss rate applied as a multiplier to demand to size orders at the gross level |
Forecasts can be computed using any user-selected combination of these components. For most items, all four are relevant; for new items with no history, like-item forecasting uses a reference item's pattern.
Forecast parameters — Forecasting parameters are maintained at the item × site level. Key parameters include: the forecasting model in use (demand-based, inventory-based, order-up-to), the historical look-back window (in weeks), the seasonal profile assigned to the item, the promotional event calendar, and the minimum and maximum order constraints. Parameters are mass-maintained across merchandise and site hierarchies with effective dating — a single parameter change can be applied to a class across all stores in a climate region.
Forecast schedule — For most items, forecasting runs weekly. The forecast cycle: calculate demand forecast → review exception list (items where forecast variance from actuals exceeds threshold) → buyer or replenishment planner overrides where necessary → approved forecast feeds the replenishment process. Exception-first review keeps the process scalable across tens of thousands of SKUs.
Promotional demand isolation — Promotional demand must be isolated from baseline history. A week with a major promotional event should not inflate the baseline demand estimate for future non-promotional weeks. The forecasting system tracks promotional vs. non-promotional sales separately and excludes promotional periods from baseline calculations unless explicitly included.
Collaborative forecasting — In a more advanced configuration, vendors receive buyer-level sales forecasts and provide their own forward-looking demand estimates. Collaborative forecasting improves vendor fill rates by giving suppliers earlier visibility into order quantities. EDI 852 (Product Activity Data) enables automated data sharing. Collaborative forecasting is a vendor compliance dimension: vendors who commit to collaborative forecasting and miss it by more than the agreed tolerance are scored accordingly.
Simulation capability — Replenishment simulations must be runnable without creating actual purchase orders or deliveries. Simulations use production parameters but produce simulated suggested order quantities (SOQs) only, enabling buyers to test parameter changes before committing.
Consumers
The Replenishment module reads the approved forecast as the primary input to suggested order quantity calculation. The Merchandise Financial Planning process reads forecast by item and store to build the sales assumption in the financial plan. The Vendor Agent shares forecasts with collaborative-forecasting-enrolled vendors. The Operations Agent monitors forecast accuracy (MAPE — Mean Absolute Percentage Error) by category and site, and alerts when accuracy degrades below a defined threshold.
Invariants
- Promotional demand must be isolated from baseline forecasting. Allowing promotional events to inflate the baseline is the most common cause of post-promotional inventory excess.
- Forecast simulations must not affect production data. Simulation is a planning tool, not a production process.
- New item forecasting must use a designated like-item profile until sufficient history accumulates (typically 13+ weeks). Forecasting a new item with zero history and no reference produces a zero or noise forecast.
Platform (2030)
Agent mandate: Business Agent runs and reviews the demand forecast. Operations Agent monitors forecast accuracy (MAPE) by category and site continuously, and alerts when accuracy degrades below threshold. Business Agent manages like-item assignments for new items and coordinates collaborative forecast data exchange with enrolled vendors.
AI-native forecasting vs. rule-based statistics. Traditional retail forecasting uses statistical time-series models (exponential smoothing, moving averages) with manual seasonal profiles and promotional uplift tables maintained by replenishment analysts. Canary Go Business Agent runs AI-native demand forecasting: the model ingests historical sell-through, promotional history, seasonality signals, weather zone data, and market-level competitive signals to produce item-site-week demand estimates. The model updates continuously as new sales data arrives — weekly batch recalculation is replaced by incremental updates. Promotional demand isolation is automatic: the model identifies promotional vs. non-promotional periods from the event calendar without manual flagging.
Collaborative forecast via MCP. Vendors enrolled in collaborative forecasting receive a machine-readable forecast via an MCP endpoint — not a spreadsheet. The vendor submits their own forward availability estimate via a corresponding endpoint. Collaborative forecasting becomes a real-time data exchange, not a quarterly meeting. Vendor forecast submission is a scored compliance dimension; a vendor who commits to collaborative forecasting and misses the submission window is recorded as a compliance event.
MCP surface. demand_forecast(sku, site, horizon) returns the forward demand estimate by week. forecast_accuracy(dept_or_sku, period) returns MAPE and bias metrics. new_item_profile(like_item_id) returns the reference demand profile for new item setup. promo_lift(event_id, item_id) returns expected promotional demand uplift for a planned event. Single-call, agent-readable.
RaaS. Forecast inputs derive from receipt events — sales history from validated, audited POS data. Forecast accuracy is only meaningful when built on a complete and correctly sequenced receipt record; forecasts built on unaudited or out-of-order data inherit those errors. demand_forecast(sku, site, horizon) is pre-computed and cached (sub-200ms per SKU/site/horizon tuple). Ad hoc forecast queries across all SKUs in a department must be batch-friendly — not real-time joins over raw transactions. Forecast outputs in SQL indexed on (item_id, site_id, forecast_date). Model inputs must be portable for retraining; the same input sequence must produce the same forecast.
Related
- retail-replenishment-model — consumes forecast as the primary demand input for SOQ calculation
- retail-merchandise-financial-planning — forecast provides the sales assumption for OTB and plan components
- retail-event-management — promotional event calendar drives promotional demand component
- retail-merchandise-hierarchy — hierarchy groupings enable mass parameter maintenance and forecast aggregation