Inventory Planning & Stock Optimization
“How many units of each SKU should we stock at each warehouse over the next 30 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
How many units of each SKU should we stock at each warehouse over the next 30 days?
Inventory carrying costs consume 20–30% of product value annually. Overstocking ties up working capital; understocking means lost sales and expedited shipping at 3–5x standard cost. With thousands of SKUs across multiple warehouses, getting the right quantity at the right location requires understanding cross-product demand dependencies, regional consumption patterns, and supply chain lead times that spreadsheets and simple rules cannot capture.
Quick answer
Inventory planning predicts how many units of each SKU should be stocked at each warehouse over the next 30 days. Graph-based models connect products to orders, warehouses, suppliers, and seasonal patterns, producing stock-level predictions that account for lead times, cross-product demand dependencies, and regional consumption patterns that spreadsheets miss.
Approaches compared
4 ways to solve this problem
1. Safety Stock Formulas
Calculate reorder points and safety stock using statistical formulas (average demand + Z-score * standard deviation * lead time). The textbook approach taught in supply chain courses.
Best for
Stable-demand products with predictable lead times and well-understood demand variability.
Watch out for
Assumes demand follows a normal distribution, which is rarely true for seasonal or promotional products. Cannot account for cross-product demand shifts or supplier disruptions.
2. ERP-Based Planning (SAP APO, Oracle ASCP)
Enterprise planning systems that combine demand forecasts with supply constraints, warehouse capacity, and lead times. The standard for large enterprises.
Best for
Large enterprises with established ERP systems and supply chain planning teams.
Watch out for
Demand inputs are often stale (monthly updates). The models treat each SKU-warehouse independently. Integration with external signals (weather, events, competitor actions) requires custom development.
3. ML Demand Forecasting + Optimization
Train ML models (XGBoost, LightGBM) to forecast demand, then feed forecasts into an optimization layer that accounts for lead times, costs, and capacity constraints.
Best for
Teams with ML infrastructure who can maintain two systems (forecasting + optimization). Good accuracy when features are well-engineered.
Watch out for
Two-system architecture means errors compound: forecast errors flow into suboptimal stocking decisions. Feature engineering is a manual bottleneck. Cross-product demand dependencies require explicit feature creation.
4. KumoRFM (Graph Neural Networks on Relational Data)
Connects products, orders, warehouses, and suppliers into a unified relational graph. Predicts demand per SKU per warehouse directly, capturing cross-product and cross-warehouse signals that feed into stocking decisions with zero feature engineering.
Best for
Multi-warehouse operations with thousands of SKUs, shared suppliers, and regional demand patterns.
Watch out for
Best when products and warehouses have meaningful relational connections (shared suppliers, overlapping geographies). A single-warehouse operation with independent products benefits less from the graph structure.
Key metric: Graph-based inventory models reduce carrying costs by 20% while cutting stockout rates by 30-50%, capturing cross-product and cross-warehouse signals that independent SKU models structurally miss.
Why relational data changes the answer
Product P-100 (Wireless Mouse) needs 4,820 units stocked in the next 30 days. A simple model arrives at this number by extrapolating recent order velocity. But the relational graph reveals why: P-100 shares supplier SUP-12 with P-200 (USB-C Hub), and that supplier has a 21-day lead time, meaning any reorder placed today arrives just in time. Warehouse WH-East is in the Northeast region, which historically sees 40% higher volume in Q4. The Electronics category is growing 18% year-over-year, and the current 14-day order velocity trend shows an additional 22% acceleration.
The critical insight is that these signals live in different tables: supplier lead times in PRODUCTS, regional patterns in WAREHOUSES, category trends in ORDER_LINES aggregated by product category. A flat model would require an engineer to pre-compute 'supplier lead time buffer,' 'regional seasonal factor,' and 'category growth rate' as separate features. The graph neural network discovers these cross-table patterns automatically and learns how they interact. When supplier lead times increase and regional demand peaks simultaneously, the required safety stock is not additive but multiplicative. Only models that see the full relational structure capture these compound effects.
Planning inventory with a flat model is like a restaurant chef deciding how much to prep by only looking at last week's order counts. A relational model also sees that there is a football game tomorrow (event), the fish supplier is delayed by 2 days (lead time), the neighboring restaurant just closed (demand shift), and it is the first warm weekend of spring (seasonal). Each signal alone changes the prep estimate; together they transform it.
How KumoRFM solves this
Relational intelligence for every forecast
Kumo connects products to orders, warehouses, suppliers, and seasonal patterns in a single relational graph. Instead of forecasting each SKU-warehouse pair independently, Kumo learns that Product P-100 and P-200 share a supplier with a 21-day lead time, that Warehouse WH-East sees 40% higher volume in Q4, and that recent order velocity for the Electronics category is accelerating. These cross-table signals produce stock-level predictions that account for the full supply chain context.
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
PRODUCTS
| product_id | product_name | category | lead_time_days |
|---|---|---|---|
| P-100 | Wireless Mouse | Electronics | 21 |
| P-200 | USB-C Hub | Electronics | 14 |
| P-300 | Desk Lamp | Home Office | 7 |
ORDER_LINES
| order_id | product_id | warehouse_id | quantity | timestamp |
|---|---|---|---|---|
| ORD-5001 | P-100 | WH-East | 120 | 2025-09-10 |
| ORD-5002 | P-200 | WH-East | 45 | 2025-09-11 |
| ORD-5003 | P-300 | WH-West | 310 | 2025-09-12 |
WAREHOUSES
| warehouse_id | warehouse_name | region | capacity |
|---|---|---|---|
| WH-East | Newark Distribution | Northeast | 500,000 |
| WH-West | Reno Fulfillment | West | 350,000 |
| WH-Central | Dallas Hub | South | 420,000 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT SUM(ORDER_LINES.QUANTITY, 0, 30, days) FOR EACH PRODUCTS.PRODUCT_ID
Prediction output
Every entity gets a score, updated continuously
| PRODUCT_ID | TIMESTAMP | TARGET_PRED |
|---|---|---|
| P-100 | 2025-10-01 | 4,820 |
| P-200 | 2025-10-01 | 1,250 |
| P-300 | 2025-10-01 | 8,340 |
Understand why
Every prediction includes feature attributions — no black boxes
Product P-100 (Wireless Mouse)
Predicted: 4,820 units needed in next 30 days
Top contributing features
Order velocity (14d trend)
+22%
29% attribution
Seasonal pattern (Q4 ramp)
Strong
25% attribution
Warehouse region demand
Northeast peak
20% attribution
Lead time buffer required
21 days
15% attribution
Category growth rate
+18% YoY
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 inventory planning & stock optimization
How much can AI reduce inventory carrying costs?
Companies using graph-based inventory planning typically reduce carrying costs by 15-25% while simultaneously reducing stockout rates by 30-50%. The savings come from right-sizing safety stock at each location rather than applying blanket buffers across all SKUs.
What is the difference between demand forecasting and inventory planning?
Demand forecasting predicts how much customers will buy. Inventory planning determines how much to stock, accounting for lead times, carrying costs, service level targets, and warehouse capacity. Demand forecasts are an input to inventory planning, but the planning layer adds supply-side constraints.
How do supplier lead times affect inventory optimization?
Lead times directly determine how far ahead you need to forecast and how much safety stock to hold. Products with 21-day lead times need 3-4 weeks of forecasting accuracy, while 7-day lead time products can be restocked more responsively. Graph models capture lead time data from the supplier table and factor it into stocking recommendations automatically.
Can inventory planning AI handle seasonal demand spikes?
Yes. Graph models learn seasonal patterns from historical data across products, warehouses, and regions simultaneously. They detect that Q4 demand at Northeast warehouses is 40% higher than Q2, and that promotional events create demand spikes 3-5 days before the promotion starts (pre-purchase behavior). This pre-positioning of inventory before the spike is where the largest savings occur.
Bottom line: Cut carrying costs by 20% while reducing stockouts — right-size every SKU at every warehouse with zero manual forecasting.
Related use cases
Explore more forecasting 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.
