In March 2020, global supply chains broke. Not because demand disappeared. Because demand shifted, and the systems managing supply chains could not see where it shifted to or what the downstream effects would be.
Consumer demand for hand sanitizer spiked 600% in a week. That spike required more ethanol, more plastic bottles, more pump dispensers, more corrugated packaging, more trucking capacity to retailers. Each upstream supplier saw a different demand signal at a different time, with no visibility into the network-level picture.
Accenture estimated that supply chain disruptions during 2020-2022 cost Fortune 1000 companies an average of $184 million per company annually. The total cost across global industries exceeded $4 trillion. COVID was extreme. But the structural vulnerability it exposed was always there: most supply chain forecasting treats each node in the network as independent, ignoring the graph structure that determines how disruptions propagate.
supply_chain_network — sample product bill of materials
| product_id | component | supplier | supplier_region | lead_time_days | alt_supplier |
|---|---|---|---|---|---|
| PROD-100 | Display Panel | Shenzhen Optics | China | 45 | Korea Display Co. |
| PROD-100 | Main PCB | TechBoard Mfg | Taiwan | 30 | None |
| PROD-100 | Battery Cell | PowerCell Ltd | China | 25 | Japan Energy |
| PROD-200 | Main PCB | TechBoard Mfg | Taiwan | 30 | None |
| PROD-200 | Sensor Module | SensorTech | Germany | 20 | US Sensors Inc |
| PROD-300 | Battery Cell | PowerCell Ltd | China | 25 | Japan Energy |
Highlighted: TechBoard Mfg supplies the Main PCB to both PROD-100 and PROD-200 with no alternative supplier. A disruption at this single node affects two product lines simultaneously.
demand_forecast_comparison — independent vs network-aware
| SKU-Location | Independent Forecast | Network-Aware Forecast | Actual Demand | Error Reduction |
|---|---|---|---|---|
| PROD-100 / Warehouse-East | 1,200 units | 980 units | 1,010 units | 82% reduction |
| PROD-200 / Warehouse-West | 850 units | 920 units | 940 units | 71% reduction |
| PROD-100 / Warehouse-West | 600 units | 740 units | 760 units | 77% reduction |
| PROD-300 / Warehouse-East | 2,100 units | 1,850 units | 1,820 units | 63% reduction |
Network-aware forecasting accounts for substitution effects, supplier constraints, and geographic spillover. Error reduction of 63-82% for products with strong cross-network dependencies.
Why supply chains are graph problems
A simplified supply chain for a consumer electronics company includes: 200+ component suppliers, 15 contract manufacturers, 8 regional distribution centers, 50,000 retail points of sale, and 2 million end customers. Each entity connects to multiple others through bills of materials, purchase orders, shipments, inventory transfers, and sales transactions.
The data lives in tables: suppliers, components, products, warehouses, shipments, orders, customers, and the relationships between them (supplier-provides-component, component-goes-into-product, product-stored-at-warehouse, warehouse-ships-to-customer). This is a relational database. It is also a graph.
The predictive signals that matter most in supply chain management flow through this graph. They are not properties of individual nodes. They are properties of the network.
Demand propagation
When a retailer sees 20% higher demand for a product this week, that signal needs to propagate upstream: the distribution center needs to ship more, the manufacturer needs to produce more, the component suppliers need to deliver more. But the propagation is not linear. The manufacturer makes 50 other products using some of the same components. Increased demand for one product competes for component capacity that affects supply of the other 49.
A flat forecasting model at each node cannot see this competition. A graph model that represents the full bill-of-materials network can predict the capacity constraints before they cause stockouts.
Risk propagation
When a Tier-2 supplier in Shenzhen has a factory fire, which of your 200 finished products are affected? The answer depends on the graph. The burned factory makes a specific capacitor. That capacitor goes into 3 circuit boards. Those circuit boards go into 12 products. But 2 of the circuit boards have alternative suppliers with 4-week lead times, while the third has no alternative.
Understanding this requires traversing the supplier-component-product graph. Companies with graph-level visibility identified their COVID exposure in days. Companies without it took months.
Demand forecasting: the network effect
Traditional demand forecasting treats each product-location pair as an independent time series. The model for "Product A at Warehouse B" uses historical sales of Product A at Warehouse B, plus seasonality, promotions, and maybe weather. It does not consider what is happening to Product C at Warehouse D, even if Products A and C are substitutes that share components and customers.
Graph-based demand forecasting models the full product-location- customer-supplier network. It captures four types of network effects that independent models miss.
1. Substitution effects
When Product A goes out of stock, demand shifts to substitutes. The substitution relationships are in the product graph (same category, similar price, similar attributes). A network model predicts the demand shift at the same time it predicts the stockout, giving supply planners days of lead time.
bullwhip_effect — demand amplification through the network
| node | actual_demand_change | order_placed | amplification |
|---|---|---|---|
| Retailer | +10% (100 -> 110 units) | +15% (orders 115) | 1.5x |
| Distributor | +15% signal received | +25% (orders 125) | 2.5x |
| Manufacturer | +25% signal received | +35% (produces 135) | 3.5x |
| Raw material supplier | +35% signal received | +50% (ships 150) | 5.0x |
A 10% demand increase at the retailer becomes a 50% order spike at the raw material supplier. Each node adds safety margin. A graph model that sees the full chain detects this is amplification, not genuine 50% demand growth.
2. Complementary effects
A promotion on printers drives demand for ink cartridges. A new phone launch drives demand for cases and screen protectors. These cross-product dependencies are captured in the co-purchase graph and the bill-of-materials graph.
3. Geographic spillover
A stockout at one retail location drives customers to nearby locations. A new warehouse opening shifts demand patterns across the distribution network. The location graph captures these spatial dependencies.
4. Supplier-driven constraints
A supplier delay on one component affects the availability (and therefore the sales) of all products using that component. The bill-of-materials graph connects supplier performance to finished goods demand in a way that no product-level model can.
Independent forecasting
- Each product-location forecasted in isolation
- Substitution and complementary effects missed
- Supplier disruptions detected after stockout
- Bullwhip effect amplified at each node
- 20-40% forecast error for network-dependent products
Network-aware forecasting
- Full product-supplier-warehouse-customer graph modeled
- Cross-product dependencies captured automatically
- Supplier disruptions propagated through the network
- Bullwhip effect dampened through network visibility
- 20-30% improvement in forecast accuracy
PQL Query
PREDICT stockout_risk_7d FOR EACH inventory.sku_location_id
One query predicts stockout risk across the entire network, accounting for supplier lead times, demand propagation, substitution effects, and geographic spillover.
Output
| sku_location | stockout_risk | root_cause | recommended_action |
|---|---|---|---|
| PROD-100 / WH-East | 0.82 | TechBoard Mfg delay (PCB) | Expedite from WH-West |
| PROD-200 / WH-West | 0.74 | Same PCB supplier constraint | Alert procurement |
| PROD-300 / WH-East | 0.15 | Battery alt supplier available | Monitor |
| PROD-100 / WH-West | 0.41 | Spillover from WH-East stockout | Pre-position inventory |
Inventory optimization across the network
Inventory placement is the most expensive decision in supply chain management. Too much inventory ties up capital ($1.43 trillion in US business inventory as of 2024, per Census Bureau data). Too little causes stockouts, which cost retailers an estimated $1 trillion annually in lost sales globally, according to IHL Group.
inventory_independent — each warehouse optimized in isolation
| warehouse | product | avg_demand/week | demand_variability | safety_stock | total_stock |
|---|---|---|---|---|---|
| WH-East | PROD-100 | 200 units | High (+/- 80) | 160 units | 360 units |
| WH-West | PROD-100 | 180 units | High (+/- 70) | 140 units | 320 units |
| WH-Central | PROD-100 | 150 units | Moderate (+/- 40) | 80 units | 230 units |
Independent optimization: each warehouse carries its own safety stock. Total network safety stock: 380 units (160 + 140 + 80). Total inventory: 910 units.
inventory_network_aware — warehouses optimized as a network
| warehouse | product | avg_demand/week | safety_stock | total_stock | transfer_buffer |
|---|---|---|---|---|---|
| WH-East | PROD-100 | 200 units | 90 units | 290 units | Can receive from Central in 18h |
| WH-West | PROD-100 | 180 units | 80 units | 260 units | Can receive from Central in 24h |
| WH-Central | PROD-100 | 150 units | 100 units | 250 units | Hub: ships to East or West |
Network optimization: total safety stock drops from 380 to 270 units (29% reduction). Total inventory drops from 910 to 800 units. WH-Central acts as a buffer hub, and its demand variability is uncorrelated with East and West.
financial_impact — independent vs network inventory
| metric | Independent | Network-Aware | Savings |
|---|---|---|---|
| Total safety stock (units) | 380 | 270 | 110 units (29%) |
| Inventory carrying cost/year | $684K | $486K | $198K/year |
| Stockout rate | 4.2% | 2.8% | 33% fewer stockouts |
| Emergency expedite costs/year | $120K | $35K | $85K/year |
Network-aware optimization reduces safety stock by 29% while simultaneously reducing stockouts by 33%. The risk pooling effect means less inventory AND better service.
This is the risk pooling effect, and it compounds across the network. Amazon, Walmart, and Zara have built proprietary systems to exploit network-level inventory optimization. Their inventory turns (6-10x annually) far exceed the industry average (4-6x) because they model the network, not the nodes.
Supplier risk assessment
Most supplier risk assessment is reactive: a supplier misses a delivery, you flag them. Graph-based risk assessment is predictive: it identifies suppliers likely to fail before they do.
The signals are relational. A supplier whose on-time delivery rate has declined from 95% to 88% over 3 months, while their peer suppliers maintain 96%, is showing stress. If that supplier also serves 3 of your critical-path products with no alternative source, the network risk is high even though the individual performance metric has not crossed an alert threshold.
Graph-based models also detect concentration risks that node-level analysis misses. If your top 5 products all depend on components from suppliers in the same geographic region, a regional disruption (earthquake, flooding, political instability) affects all five simultaneously. This correlated risk is only visible in the supplier-component-product graph.
The foundation model opportunity
Supply chain data is among the most naturally relational in enterprise IT. The entities (suppliers, products, warehouses, customers) and the relationships between them (provides, contains, stored-at, ordered-by) map directly to a graph that a relational foundation model can learn from.
KumoRFM connects to the supply chain data warehouse, understands the relational schema, and serves predictions across demand forecasting, inventory optimization, supplier risk, and logistics planning. The same model that forecasts demand at the SKU-location level also predicts supplier lead times, identifies stockout risks, and recommends inventory transfers.
Building separate models for each of these tasks requires 4-6 dedicated data science resources per task and 12-18 months of development. A foundation model approach serves all tasks from a single platform, with time to first prediction measured in days.
In an industry where every day of inventory costs money and every stockout loses sales, the speed advantage alone justifies the approach. The accuracy advantage, driven by network-level signals that independent models cannot access, makes the case overwhelming.