SKU-Level Demand Forecasting
“What will SKU-level demand be at each store next week?”
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 SKU-level demand be at each store next week?
Retailers lose $1.1T globally to overstock and stockouts each year (IHL Group). A grocer with 500 stores and 40,000 SKUs must make 20M replenishment decisions weekly. Traditional time-series models forecast each SKU-store pair independently, missing cross-product cannibalization (when the organic brand goes on sale, the conventional brand drops 30%), weather-driven demand shifts (ice cream sales spike 2x when temperatures exceed 90F), and promotional halo effects (discounting chips lifts salsa sales by 18%). These blind spots cause 8-12% forecast error that compounds into millions in waste and lost sales.
Quick answer
The best retail demand forecasting models connect products, stores, promotions, and weather into a single graph rather than forecasting each SKU-store pair in isolation. On the RelBench product demand task, Kumo's relational foundation model (KumoRFM) scores 76.71 vs 62.44 for the next-best method. In SAP SALT benchmarks, relational approaches hit 91% accuracy vs 75% for XGBoost and 63% for manual rules. The key insight: cross-product signals like cannibalization and halo effects account for 15-30% of demand variation that single-table models miss entirely.
Approaches compared
4 ways to solve this problem
1. Spreadsheets and moving averages
Planners use trailing 4-8 week averages with manual adjustments for promotions and holidays. Common in small-to-mid grocers.
Best for
Stable-demand staples with minimal promotional activity and fewer than 500 SKUs.
Watch out for
Completely blind to cross-product effects, weather shifts, and promotional halo. Error rates typically run 15-25%.
2. Time-series models (ARIMA, Prophet, ETS)
Each SKU-store pair gets its own univariate model. Facebook Prophet and exponential smoothing are the most common choices.
Best for
Single-store or single-category forecasting where external signals are minimal.
Watch out for
Each model trains independently, so cannibalization between organic and conventional milk is invisible. Forecast error stays at 10-15% for promotional periods.
3. Gradient-boosted trees (XGBoost, LightGBM)
Feature-engineered tabular models trained on flattened transaction data with hand-built lag features and promotional flags.
Best for
Teams with strong ML engineering who can invest weeks building and maintaining feature pipelines.
Watch out for
Feature engineering is the bottleneck. SAP SALT benchmarks show 75% accuracy, a clear ceiling from missing relational structure.
4. KumoRFM (relational foundation model)
Connects products, stores, promotions, suppliers, and weather into a relational graph and learns cross-entity demand signals automatically.
Best for
Retailers with 1,000+ SKUs across multiple stores where cross-product and promotional interactions dominate demand patterns.
Watch out for
Requires relational data (multiple connected tables). If your data is genuinely a single flat CSV, you will not see the full benefit.
Key metric: RelBench product demand: KumoRFM 76.71 vs next-best 62.44. SAP SALT: relational 91% vs XGBoost 75% vs rules 63%.
Why relational data changes the answer
Single-table demand models treat each SKU-store combination as if it exists in a vacuum. They cannot see that when you run a BOGO promotion on organic milk, conventional milk drops 30% at the same store, or that a competitor store closing 2 miles away shifts 12% of foot traffic to your location. These cross-entity signals live in the joins between your products, stores, promotions, and external data tables.
A relational model ingests these tables as a connected graph and learns the substitution, complementarity, and cannibalization patterns automatically. On the RelBench product demand benchmark, this approach scores 76.71 vs 62.44 for the next-best method because it captures promotional halo effects (chips discount lifts salsa sales 18%), weather-driven category shifts, and supplier disruption cascades that ripple across the product catalog.
Forecasting SKU demand with a single-table model is like predicting grocery store checkout times by only looking at the number of items in each cart. You miss that the express lane just opened, three cashiers are on break, and a busload of tourists just walked in. The checkout time for your cart depends on everything happening around it. The same is true for SKU demand: the organic milk forecast depends on the conventional milk promotion, the weather, and the competitor across the street.
How KumoRFM solves this
Relational intelligence built for retail and e-commerce data
Kumo connects products, transactions, stores, promotions, weather, and supplier data into a single relational graph. The model learns that SKU-4201 (organic milk) at Store S-14 will see 340 units next week because a competitor store nearby closed, a BOGO promotion starts Wednesday, and regional temperatures are forecast to stay cool (higher dairy consumption). Cross-product signals, like the substitution relationship between organic and conventional milk, emerge automatically from the graph without manual feature engineering.
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
| sku_id | product_name | category | brand | unit_cost |
|---|---|---|---|---|
| SKU-4201 | Organic Whole Milk 1gal | Dairy | Valley Fresh | $3.20 |
| SKU-4202 | Conv. Whole Milk 1gal | Dairy | DairyPure | $2.80 |
| SKU-4310 | Tortilla Chips 12oz | Snacks | Casa Crunch | $2.10 |
SALES_HISTORY
| sku_id | store_id | date | units_sold | revenue | promo_active |
|---|---|---|---|---|---|
| SKU-4201 | S-14 | 2025-09-22 | 48 | $287.52 | False |
| SKU-4201 | S-14 | 2025-09-23 | 52 | $311.48 | False |
| SKU-4310 | S-14 | 2025-09-22 | 85 | $339.15 | True |
STORES
| store_id | name | format | region | sqft |
|---|---|---|---|---|
| S-14 | Union Square Market | Urban | West | 42,000 |
| S-22 | Midtown Grocery | Suburban | Northeast | 55,000 |
| S-37 | Lakeside Fresh | Rural | Midwest | 35,000 |
PROMOTIONS
| promo_id | sku_id | store_id | type | start_date | end_date |
|---|---|---|---|---|---|
| PR-801 | SKU-4201 | S-14 | BOGO | 2025-10-01 | 2025-10-07 |
| PR-802 | SKU-4310 | S-14 | 20% Off | 2025-09-20 | 2025-09-26 |
| PR-803 | SKU-4202 | S-22 | None | N/A | N/A |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT SUM(SALES_HISTORY.UNITS_SOLD, 0, 7, days) FOR EACH PRODUCTS.SKU_ID, STORES.STORE_ID WHERE STORES.REGION = 'West'
Prediction output
Every entity gets a score, updated continuously
| SKU_ID | STORE_ID | PREDICTED_UNITS | CURRENT_STOCK | REORDER_QTY |
|---|---|---|---|---|
| SKU-4201 | S-14 | 412 | 180 | 232 |
| SKU-4202 | S-14 | 195 | 220 | 0 |
| SKU-4310 | S-14 | 520 | 300 | 220 |
Understand why
Every prediction includes feature attributions — no black boxes
SKU-4201 (Organic Whole Milk) at Store S-14
Predicted: 412 units predicted for next week
Top contributing features
BOGO promotion starting Oct 1
+65% lift
30% attribution
Competitor store closure nearby
+12% traffic
22% attribution
Weekly seasonal trend
Peak dairy week
19% attribution
Cross-product substitution (conv. milk)
-15% conv.
16% attribution
Temperature forecast (cool week)
62F avg
13% 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 sku-level demand forecasting
How accurate is AI demand forecasting for retail?
Accuracy depends on the model architecture and the richness of connected data. On the SAP SALT retail benchmarks, relational approaches achieve 91% accuracy, XGBoost reaches 75%, and manual rule-based methods hit 63%. On the RelBench product demand task, KumoRFM scores 76.71 vs 62.44 for the next-best model. The gap widens during promotional periods and seasonal transitions where cross-product signals matter most.
What data do I need for SKU-level demand forecasting?
At minimum: a products table, a transactions/sales history table, and a stores table. Adding promotions, supplier lead times, weather data, and competitor pricing progressively improves accuracy. The critical factor is not volume of data but the relational connections between tables. A retailer with 12 months of transactional data across 4-5 connected tables will outperform one with 5 years of flat-file history.
How does demand forecasting handle new stores or new SKUs?
This is the cold-start problem. Traditional time-series models cannot forecast without historical data. Relational models solve this by transferring learned patterns from similar stores or similar products through the graph. A new urban store inherits demand patterns from existing urban stores with similar demographics, while a new organic SKU borrows signals from the organic product cluster.
Bottom line: Reduce stockouts by 40% and overstock by 25%, freeing $2-5M in working capital per quarter for a 500-store grocery chain.
Related use cases
Explore more retail & e-commerce 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.
