Demand Sensing
“What will demand be at each node in the next 7 days?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
What will demand be at each node in the next 7 days?
Traditional demand planning uses monthly or weekly forecasts that miss short-term demand shifts caused by promotions, weather, competitor actions, and upstream disruptions. These misses cascade through the supply chain: stockouts at high-demand nodes, overstock at low-demand nodes. For a retailer with 500 locations and $5B in inventory, a 15% improvement in 7-day demand accuracy reduces carrying costs by $75M and stockout losses by $120M annually.
Quick answer
Graph neural networks predict short-term demand at every supply chain node by learning how demand propagates across warehouses, stores, and distribution centers. Unlike monthly forecasts that miss short-term shifts from promotions, weather, and competitor actions, graph-based demand sensing captures cross-node signal propagation, improving 7-day accuracy by 15% and saving $195M annually for a 500-location retailer.
Approaches compared
4 ways to solve this problem
1. Statistical forecasting (ARIMA, exponential smoothing)
Model demand as a time series per SKU-location with trend, seasonality, and noise components. The traditional approach for demand planning.
Best for
Stable, seasonal demand patterns. Well-understood, easy to maintain, and interpretable for planners.
Watch out for
Treats each SKU-location independently. Cannot detect when a surge at a distribution center will cascade to downstream stores 2-3 days later. Also blind to external signals like weather and competitor actions.
2. ML forecasting (LightGBM, Prophet on tabular features)
Engineer demand features (lagged sales, promotional flags, weather indicators) and train a tree-based or additive model per SKU-location or cluster.
Best for
Captures promotional and external signal effects when features are well-engineered. Handles non-linear demand patterns better than pure statistical methods.
Watch out for
Feature engineering bottleneck. Cross-node demand propagation requires manually creating 'upstream order velocity' features, and you need a different lag for each node-to-node relationship. Scales poorly as the network grows.
3. Deep learning time series (DeepAR, Temporal Fusion Transformer)
Train neural network forecasters on multiple related time series simultaneously. Can share patterns across SKU-locations.
Best for
Learns shared temporal patterns across many SKU-locations. Handles intermittent demand better than traditional methods.
Watch out for
Still treats each time series as a separate input. Cannot model the network topology -- it doesn't know that WH001 feeds WH003 and that a surge at WH001 predicts a surge at WH003 with a 3-day lag.
4. KumoRFM (relational graph ML)
Connect warehouses, products, orders, shipments, and external signals into a supply chain graph. The GNN learns how demand propagates across the network and how external events modulate demand at specific nodes.
Best for
Captures cross-node demand propagation, external signal effects, and network topology. A surge at an upstream DC automatically informs predictions at downstream stores without manual feature engineering.
Watch out for
Requires supply chain data in normalized tables with clear node-to-node relationships. Adds most value for multi-echelon networks with 10+ nodes.
Key metric: 15% improvement in 7-day demand accuracy saves $195M annually ($75M carrying costs + $120M stockout losses) for a 500-location retailer.
Why relational data changes the answer
Demand does not happen in isolation at each node. When orders spike at a distribution center, the downstream stores will feel it 2-3 days later. When a competitor goes out of stock, demand shifts to nearby locations in your network. When a heat wave hits the West Coast, electronics demand rises at WH001 (a West Coast DC), and that signal should propagate to every store WH001 supplies. These cross-node, cross-signal patterns live in the relationships between your tables -- warehouses connect to orders, orders connect to products, products connect to external signals, and warehouses connect to each other through shipments.
Relational models read this supply chain graph directly. They learn the propagation delays (WH001 to WH003 takes 3 days via SHP201), the amplification patterns (a 10% upstream surge becomes a 15% downstream surge for electronics), and the external triggers (heat waves drive 51% demand increases for SKU101 at US-West nodes). On the RelBench benchmark, relational models score 76.71 vs 62.44 for single-table approaches. In supply chain, that accuracy gap translates to $195M in reduced carrying costs and stockout losses for a major retailer.
Forecasting demand at each store independently is like predicting traffic at each intersection without knowing the road network. You miss that a highway closure upstream will flood your intersection with cars in 20 minutes. Graph-based demand sensing reads the full road map of your supply chain, predicting where demand surges will propagate before they arrive.
How KumoRFM solves this
Graph-powered intelligence for supply chains
Kumo connects warehouses, products, orders, shipments, and external signals (weather, events, promotions) into a supply chain graph. The GNN learns how demand propagates across nodes: when a distribution center sees a surge, which downstream stores will feel it 2-3 days later. PQL predicts demand at each node for the next 7 days, incorporating real-time signals that monthly forecasts ignore.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
WAREHOUSES
| warehouse_id | region | type | capacity |
|---|---|---|---|
| WH001 | US-West | Distribution Center | 50,000 units |
| WH002 | US-East | Regional Hub | 25,000 units |
| WH003 | US-Central | Fulfillment | 15,000 units |
PRODUCTS
| product_id | category | avg_daily_demand | lead_time_days |
|---|---|---|---|
| SKU101 | Electronics | 450 | 3 |
| SKU102 | Apparel | 1,200 | 5 |
| SKU103 | Grocery | 3,500 | 1 |
ORDERS
| order_id | warehouse_id | product_id | qty | timestamp |
|---|---|---|---|---|
| ORD5001 | WH001 | SKU101 | 85 | 2025-03-01 |
| ORD5002 | WH002 | SKU102 | 320 | 2025-03-01 |
| ORD5003 | WH003 | SKU103 | 1,450 | 2025-03-01 |
SHIPMENTS
| shipment_id | from_warehouse | to_warehouse | status | eta |
|---|---|---|---|---|
| SHP201 | WH001 | WH003 | In Transit | 2025-03-03 |
| SHP202 | WH002 | WH003 | Delivered | 2025-03-01 |
EXTERNAL_SIGNALS
| signal_id | region | type | value | date |
|---|---|---|---|---|
| SIG01 | US-West | Weather | Heat wave | 2025-03-05 |
| SIG02 | US-East | Promotion | Flash sale | 2025-03-04 |
| SIG03 | US-Central | Event | Sports final | 2025-03-06 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT SUM(ORDERS.qty, 0, 7, days) FOR EACH WAREHOUSES.warehouse_id, PRODUCTS.product_id
Prediction output
Every entity gets a score, updated continuously
| WAREHOUSE_ID | PRODUCT_ID | PREDICTED_DEMAND_7D | VS_BASELINE |
|---|---|---|---|
| WH001 | SKU101 | 680 | +51% |
| WH002 | SKU102 | 2,850 | +18% |
| WH003 | SKU103 | 28,400 | +16% |
Understand why
Every prediction includes feature attributions — no black boxes
WH001 x SKU101 (Electronics at US-West DC)
Predicted: 680 units in 7 days (+51% vs baseline)
Top contributing features
Upcoming heat wave driving electronics demand
Heat wave Mar 5
30% attribution
Upstream order velocity increase
+38% WoW
25% attribution
Historical seasonal pattern
Spring uptick
19% attribution
In-transit shipment to downstream nodes
SHP201 in transit
15% attribution
Competitor stockout signal
Detected
11% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about demand sensing
What is demand sensing vs demand forecasting?
Demand forecasting predicts monthly or quarterly volumes for planning. Demand sensing predicts daily or weekly volumes at each node using real-time signals like orders, shipments, weather, and promotions. The 7-day horizon is where graph-based models add the most value because they capture cross-node signal propagation that monthly models miss.
How do you improve short-term demand forecasting?
Move from per-node statistical models to graph-based approaches that model your supply chain as a connected network. The biggest accuracy gains come from cross-node demand propagation signals and real-time external triggers (weather, promotions, competitor actions). Graph models capture these automatically without manual feature engineering.
What data do you need for demand sensing?
Order history by warehouse/store and SKU with timestamps, shipment data linking upstream and downstream nodes, and external signals (weather forecasts, promotional calendars, event schedules). Product metadata and inventory positions help the model understand substitution effects and capacity constraints.
How does weather affect supply chain demand?
Weather effects are location-specific, category-specific, and propagate through the network. A heat wave drives immediate demand for beverages and electronics cooling at impacted stores, then cascades to upstream DCs 1-3 days later as replenishment orders flow. Graph models learn these propagation patterns automatically from historical order-weather correlations.
What is the ROI of better demand sensing?
A retailer with 500 locations and $5B in inventory saves $195M annually from a 15% improvement in 7-day demand accuracy: $75M in reduced carrying costs (less overstock) and $120M in reduced stockout losses (fewer missed sales). The savings compound because accurate demand signals improve every downstream decision from replenishment to staffing.
Bottom line: A retailer with 500 locations and $5B in inventory saves $195M annually by improving 7-day demand sensing by 15%. Kumo's supply chain graph propagates demand signals across nodes, catching short-term shifts that monthly forecasts miss entirely.
Related use cases
Explore more supply chain use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
