Methodology

The model that reads your supply chain before it breaks.

Three-signal architecture. Planner-readable outputs. No black box.

3-Signal Forecast Architecture v2.4
Signal 01
POS Velocity
hourly ingestion
Signal 02
Weather Patterns
ZIP-level mapping
Signal 03
Supplier Lead Times
live variance feed
→ Ensemble model → SKU-level forecast + confidence interval + reorder flag

Three inputs. Each one your ERP ignores.

POS Velocity

We ingest hourly sell-through from your existing POS system. The model learns SKU-level demand curves from your actual transaction history — not industry averages, not category benchmarks. Your data.

Data format: JSON or CSV push / REST pull
Update cadence: every 1 hour
History required: 12 months minimum

Weather Patterns

Local 7-14 day weather forecasts mapped to your store locations at ZIP code level. The model learns which SKUs in your catalog are weather-elastic — so a cold front in Seattle triggers the right reorder, not a generic seasonal bump.

Data format: structured weather API (managed)
Update cadence: every 6 hours
Geographic granularity: ZIP code

Lead Time Variance

Live lead-time variance from your supplier confirmations. When a supplier slips from a 14-day lead time to 21 days, reorder triggers for dependent SKUs adjust automatically — no manual spreadsheet update required.

Data format: PO confirmation CSV or ERP webhook
Update cadence: as confirmed
Variance window: trailing 90 days per supplier
Mid-shot of warehouse shelving rows with barcode scan activity, representing live POS and inventory data capture

How the model votes on each SKU

Automcore uses an ensemble approach that blends three model families: a gradient-boosted regression layer trained on your POS history, an ARIMA-based seasonality component calibrated to your SKU categories, and a rule-based override layer that applies hard constraints from your lead-time feed.

The three models vote on each SKU’s forecast. When models agree, confidence intervals narrow. When models diverge — typically on high-velocity SKUs during unusual weather events — the confidence interval widens, flagging that prediction for human planner review.

Recency weighting means the last 30 days of POS data carry more influence than the same period from a year ago. This makes the model responsive to emerging trend shifts without ignoring long-run seasonality.

The model retrains incrementally every 24 hours as new POS data arrives. Major retrains (full model weight refresh) run weekly. You always see which model version produced each forecast.

Model Training Cycle v2.4 — retrained 6h ago
Gradient boost layer weight 0.52
ARIMA seasonality weight 0.31
Lead-time override weight 0.17
MAPE last 30d 12.4%
Weather radar-like visualization overlaid on a grid map, representing regional weather signals mapped to demand patterns

Outputs: what your planner sees

Reorder Signal Dashboard — All Locations Updated 2h ago
SKU / Product Location 6-wk Forecast Confidence Signal
Rain Jacket M Seattle NW +34% demand High — 88% Order Now
Fleece Pullover L Phoenix SW −12% demand Med — 71% Overstock
Base Layer S All Stable ±3% High — 92% On Track
Puffer Vest XL Chicago MW +18% demand Low — 54% Review

Forecast accuracy benchmarks

58%

Average MAPE reduction from baseline forecast on new customers, measured over first 90 days

±12%

6-week horizon accuracy on 80% of tracked SKUs — within planning tolerance for most reorder decisions

94%

Supplier delay capture rate — delays flagged 3+ weeks ahead on 94% of confirmed late shipments

Benchmarks measured over first 90 days across retail and manufacturing accounts. Results vary by SKU count, POS data quality, and seasonal complexity. We report actuals — not cherry-picked best-case outcomes.

See it run on your SKU data.

Request access and we’ll run a live demo on a sample of your real inventory data within 48 hours.

Request Access View pricing