Demand-planning AI · Per-SKU per-location forecasting · Multi-location retail

Your hero SKU goes OOS in Phoenix at 11am Friday. The forecast should have caught it Tuesday. Move forecasts inside the week.

You run inventory + marketing across 50-1,500 retail locations. Your demand planning runs at monthly cadence on historical velocity. The local marathon driving Friday foot traffic, the weekend weather changing the demand mix, the competitor flash sale on an adjacent SKU — none of those signals are in the forecast. The stockout lands on the day. Sub-week per-SKU per-location forecasts that ingest 12 demand signals open the pre-order or rebalance window before the stockout happens.

Audit your forecast-to-action loop (Tier 1)Explore Fractional CMO operation

Published May 30, 2026

Why historical-velocity-only forecasting misses what matters

Historical velocity assumes next week looks like the average of recent weeks. That holds when nothing changes. Multi-location operators live in a world where things change weekly. A local marathon adds foot traffic at three of your Phoenix-area stores. A weather front shifts demand mix toward weather-sensitive SKUs. A competitor runs a flash sale on an adjacent SKU and your substitution traffic spikes. Your active promo raises demand at promoted stores beyond the baseline.

The historical-velocity baseline is wrong in the only way that matters: when conditions diverge from average. Stockouts cluster on the weeks where they diverge. The forecast that only knows the baseline misses the divergence — which is exactly when the operator needs the forecast.

The 12 demand signals beyond historical velocity are not exotic data sources. Most operators already have them somewhere in the stack — promo pressure in marketing automation, local events in a content engine, weather in a per-location weather API, news in a per-location news feed. The gap is integrating them into one forecast shape so the inventory + marketing teams see one per-SKU per-location prediction per cycle.

We’ve built the forecast-to-action loop for retail operators. Here’s what we know.

You probably already have most of the data. Your ERP tracks stock on hand. Your marketing automation knows the active campaigns. A weather API can be wired up in a week. The gap is the integration layer that joins 12 demand-signal feeds + per-supplier lead-time distributions + a per-location bias-correction model into one forecast emission cycle, plus the downstream action handoff to the per-state action-decisioning layer.

We have built this for multi-location retail operators across verticals. We know which signal feeds are load-bearing per vertical (promo pressure dominates in consumer goods; weather dominates in apparel; events dominate in food + beverage). We know how to encode per-vertical signal weights so the first 12 forecast cycles produce calibrated output. We bring the runbook and the per-vertical starter weights.

How we get from monthly cadence to sub-week per-location forecasts

Step 1 — Tier 1 AI Readiness Assessment (2-3 weeks). We audit your current forecasting cadence + granularity + demand-signal coverage + per-supplier lead-time tracking. We sample your last 90-180 days of stockouts and back-test which demand signals would have flagged them. Output: the per-SKU per-location forecast spec, the per-vertical signal-weight starter, and a per- supplier lead-time distribution baseline.

Step 2 — Tier 2 AI Swarm Setup Sprint (4-8 weeks). We build the forecast layer end-to-end: per-SKU per-location demand-signal ingestion (12 sources), per-supplier lead-time modeling, sub-week cadence emission, per-horizon confidence intervals, per-location bias correction, downstream action-decisioning handoff, per-cycle forecast-quality reporting. Your engineering team receives the running system, all source code, all credentials.

Step 3 — Tier 3 Fractional CMO with AI Swarm ( 6-month minimum, 1-2 days/wk). We operate the forecast loop in production. Extend signal coverage as new sources come online. Tune per-location bias correction. Coordinate per-supplier lead-time-SLA renegotiations with your operations leadership. Roll up monthly per-location forecast- effectiveness reports.

What changes for you

You stop finding out about Friday stockouts on Friday. The Tuesday forecast surfaced the per-SKU per-location risk; the pre-order or rebalance fired Wednesday.

You stop arguing with your VP of Operations about whether the historical-velocity forecast was good enough. The forecast quality is trended per location per cycle; the demand-signal contributions are visible per forecast.

You can answer the question your CFO asks every quarterly review: how many stockouts did the forecast catch in advance last quarter, and what was the avoided-revenue- loss estimate per state. The per-action attribution rolls up per state per week.

You can onboard a new product category or a new state without re-training the forecast model from scratch. The per-vertical signal-weight starter applies; the per- location calibration converges over the first 12 cycles.

Frequently asked

How is sub-week per-SKU per-location forecasting different from the demand planning my ERP already runs?

Traditional demand planning runs at aggregate cadence (monthly or weekly best-case) at the SKU-and-warehouse level, with historical velocity as the primary signal. NetStock, Inventoro, Lokad, NetSuite, SAP IBP, Oracle SCM, Anaplan, and SAS Forecast Server are good at the forecasting primitive inside that granularity assumption. The gap at multi-location scale is sub-week per-SKU per-location cadence + demand-signal ingestion beyond historical velocity (promo pressure, local events, weather, news sentiment, competitive pressure, substitution cannibalization) + per-supplier lead-time variance modeled at per-location granularity. The forecast triggers per-SKU per-location pre-order, replenishment-from-overstock-location, or per-SKU per-location demand-modulation actions before the stockout happens. Building that composition is operator-side wiring.

What demand signals beyond historical velocity matter for per-SKU per-location forecast accuracy?

Twelve recur across multi-location retail operators. Historical velocity (the baseline). Current-week velocity vs baseline. Promo pressure (active campaigns + active offers at the location). Creative-rotation uplift (campaign-asset effects). Local event uplift (per-location event ingestion stream). Weather uplift (per-location weather forecast joined to per-SKU weather sensitivity). News-sentiment uplift. Per-cohort customer presence at the location. Per-channel mix at the location. Competitive pressure (competitor promo activity per geography). Per-vertical seasonality. Substitution-product availability shifts. The combination produces forecasts that catch demand spikes the historical-velocity baseline misses; the operational unlock is the action layer downstream that converts the forecast into per-state ad-spend + GBP-post + save-flow + attribution actions.

Why does per-supplier lead-time variance at per-location granularity matter?

Per-supplier lead-time is not a single number per supplier. Each supplier delivers to each location at a different average + a different variance depending on per-supplier-per-location routing, per-shipping-mode, per-customs-delay-risk, per-supplier on-time-rate, and per-supplier SLA terms. The forecasting layer maintains per-SKU per-location per-supplier historical-lead-time distribution + per-supplier variance + per-supplier SLA-breach pattern. The pre-order trigger consumes the distribution. If forecast-stockout-date is 14 days out and per-supplier-per-location 90th-percentile lead-time is 17 days, the trigger fires immediately rather than waiting for the central-warehouse average to indicate sufficient runway. The per-location specificity catches stockouts the aggregate average masks.

What does Completions commit to on Tier 3 if we run this layer in production?

Tier 3 process commitments include: weekly per-SKU per-location forecast emission cycle on a documented schedule; per-location bias-correction model update on every cycle close; per-supplier lead-time distribution refresh quarterly as supplier performance evolves; weekly false-positive review of pre-order triggers that fired without producing inventory benefit; per-vertical signal-weight tuning quarterly with your inventory + marketing leadership. We commit to the operating discipline. Per-location MAPE + per-horizon prediction-interval coverage are tuned per stack and recorded as engagement KPIs.

Who owns the forecasts, the signal-weight config, and the per-supplier data post-engagement?

Your team owns the forecast records, the per-supplier metadata, the per-location demand-signal source feeds, the per-vertical signal-weight config, the inventory system credentials, and the ad-platform credentials. Completions owns the orchestration knowledge: the per-location bias-correction tuning history, the demand-signal-source integration runbook, the per-supplier SLA-renegotiation playbook for operations leadership. At engagement end we transition operational ownership back to your team over 30-60 days with documented handover.

How does the forecast layer connect to the rest of the inventory + marketing stack?

The forecast layer subscribes upstream to your inventory system (stock events), the change-event-emission stream (per-customer + per-SKU state), and 12+ per-location demand-signal feeds. It publishes downstream: per-SKU per-location stockout-probability across multiple horizons. The per-state action-decisioning layer subscribes to high-probability stockout events and fires the canonical downstream actions (pause ad spend, throttle PMax, rewrite GBP, reroute save-flow, emit attribution). The save-offer library coordinates per-substitution offer updates. Three layers, one forecast contract.

Start with the audit

Tier 1 AI Readiness Assessment (2-3 weeks): we audit your forecasting cadence + granularity + demand- signal coverage, sample 90-180 days of stockouts against the signals that fired, and produce the per-SKU per- location forecast spec + per-vertical signal-weight starter. If you decide to build, Tier 2 ships the forecast layer. If you decide to operate it with us, Tier 3 runs it in production. You choose the next step at each gate.