Demand Planning
“What will order volume be by product line next quarter?”
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 order volume be by product line next quarter?
Manufacturing demand plans drive capacity allocation, material procurement, and workforce scheduling. A 10% forecast error means either excess capacity ($2-5M wasted) or insufficient capacity (lost orders worth $5-10M). Traditional statistical forecasts miss the demand network: how customer ordering patterns shift based on their end-market conditions, competitor moves, and macroeconomic signals. For a $1B manufacturer, improving quarterly forecast accuracy by 15% saves $8-12M in misallocated capacity.
Quick answer
Demand planning AI for manufacturing uses graph-based models to forecast order volumes by product line, incorporating customer network signals, end-market conditions, and competitive dynamics. Unlike time-series models that treat each product line independently, graph-based approaches capture how demand propagates through the customer network. A $1B manufacturer typically saves $8-12M annually by improving quarterly forecast accuracy 15%, reducing both excess capacity waste and missed orders.
Approaches compared
4 ways to solve this problem
1. Statistical Time-Series (ARIMA/Exponential Smoothing)
Forecast each product line's demand independently using historical order patterns, seasonality, and trend decomposition. The traditional approach used in most S&OP processes.
Best for
Stable product lines with strong seasonal patterns and limited exposure to market disruptions.
Watch out for
Treats each product line as independent. Cannot capture how a downturn in one customer's industry affects orders across multiple product lines. Slow to react to demand shifts because it relies on historical patterns that may not repeat.
2. Collaborative Planning (Customer Forecasts)
Collect demand forecasts directly from key customers and aggregate them into a production plan. Incorporates customer intelligence that purely statistical models lack.
Best for
Make-to-order environments with a concentrated customer base where a few customers drive most volume.
Watch out for
Customer forecasts are notoriously biased (typically over-forecast by 10-20% to ensure supply). Aggregating biased inputs does not produce accurate plans. Also fails to capture the network effect: what happens when multiple customers face the same end-market shift simultaneously.
3. ML on Flat Features (XGBoost with Exogenous Variables)
Train gradient-boosted models on flattened tables combining historical orders, market indicators, and customer data. More flexible than time-series and can incorporate external signals.
Best for
Environments with rich external data (market indices, economic indicators) that correlate with demand but where relational complexity is moderate.
Watch out for
Flattening customer-product-market relationships into features loses the network structure. Cannot represent how demand cascades through the customer network when end-market conditions change. Feature engineering requires constant manual maintenance.
4. Graph Neural Networks (Kumo's Approach)
Connect customers, orders, products, market data, and historical forecasts into a demand graph. GNNs learn how demand signals propagate through the customer-product network.
Best for
Complex manufacturing with diverse customer bases, multiple product lines, and exposure to interconnected end-markets.
Watch out for
Requires relational data linking customers to products to market conditions. Less value-add for commodity manufacturing with few products and homogeneous demand patterns.
Key metric: Graph-based demand models improve quarterly forecast accuracy 15% by capturing customer-product-market network effects. SAP SALT shows 91% accuracy for graph models vs 63% for gradient-boosted trees, with the gap driven by cross-customer demand propagation signals.
Why relational data changes the answer
Demand in manufacturing does not happen in isolation. When the automotive industry ramps EV production, it does not just affect your Precision Parts line. It also affects Specialty Alloys (battery housings), Micro Components (sensors), and creates capacity competition that shifts your other customers' ordering behavior. Time-series models see each product line as a separate forecast. They cannot represent the fact that AutoMakers Inc, AeroSpace Corp, and ConsumerTech Ltd are all connected through shared supply chains, overlapping end-markets, and competitive dynamics.
Graph-based demand models represent these connections directly. The GNN learns that when AutoMakers' EV ramp signal strengthens, Precision Parts demand increases +12%, but Micro Components may soften because ConsumerTech diverts its own supply chain capacity. These cross-product, cross-customer demand correlations are invisible to flat models. SAP's SALT benchmark shows the accuracy advantage: 91% for graph-based models vs 63% for gradient-boosted trees. RelBench confirms at 76.71 vs 62.44 for relational prediction tasks. In demand planning, that accuracy gap translates to 15%+ improvement in quarterly forecast accuracy, which saves $8-12M annually for a $1B manufacturer in reduced capacity misallocation.
Time-series demand planning is like predicting traffic on a single highway by looking only at that highway's historical patterns. You would miss that a stadium event, a construction detour on a parallel road, and a weather front are all about to hit simultaneously. Graph-based demand planning monitors the entire transportation network: when traffic shifts on one route, the model predicts how it cascades to connected routes. Manufacturing demand works the same way. Your customers are connected through shared markets, and demand shifts propagate through the network.
How KumoRFM solves this
Graph-powered intelligence for manufacturing
Kumo connects customers, orders, products, historical forecasts, and market data into a demand graph. The GNN learns how demand propagates through the customer network: when a major customer's end-market shifts, which products and product lines will see correlated demand changes. PQL forecasts quarterly order volume per product line, incorporating real-time customer signals that traditional time-series models treat as external variables.
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
CUSTOMERS
| customer_id | name | industry | annual_spend |
|---|---|---|---|
| CUST01 | AutoMakers Inc | Automotive | $45M |
| CUST02 | AeroSpace Corp | Aerospace | $28M |
| CUST03 | ConsumerTech Ltd | Electronics | $18M |
ORDERS
| order_id | customer_id | product_line | qty | timestamp |
|---|---|---|---|---|
| ORD7001 | CUST01 | Precision Parts | 12,000 | 2025-02-15 |
| ORD7002 | CUST02 | Specialty Alloys | 3,500 | 2025-02-20 |
| ORD7003 | CUST03 | Micro Components | 45,000 | 2025-02-28 |
PRODUCTS
| product_line | margin_pct | capacity_util | lead_time_weeks |
|---|---|---|---|
| Precision Parts | 32% | 78% | 6 |
| Specialty Alloys | 45% | 85% | 8 |
| Micro Components | 28% | 92% | 4 |
FORECASTS
| product_line | quarter | forecast_qty | actual_qty | error_pct |
|---|---|---|---|---|
| Precision Parts | Q4-2024 | 48,000 | 52,300 | -8.2% |
| Specialty Alloys | Q4-2024 | 14,000 | 12,800 | +9.4% |
| Micro Components | Q4-2024 | 180,000 | 195,000 | -7.7% |
MARKET_DATA
| industry | indicator | trend | confidence |
|---|---|---|---|
| Automotive | EV production ramp | Strong growth | High |
| Aerospace | Defense spend increase | Moderate growth | Medium |
| Electronics | Consumer demand soft | Flat | High |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT SUM(ORDERS.qty, 0, 90, days) FOR EACH PRODUCTS.product_line
Prediction output
Every entity gets a score, updated continuously
| PRODUCT_LINE | Q2_2025_FORECAST | VS_Q1_ACTUAL | CONFIDENCE |
|---|---|---|---|
| Precision Parts | 58,200 | +12% | High |
| Specialty Alloys | 14,800 | +8% | Medium |
| Micro Components | 188,000 | -2% | High |
Understand why
Every prediction includes feature attributions — no black boxes
Product line: Precision Parts -- Q2 2025 forecast
Predicted: 58,200 units (+12% vs Q1)
Top contributing features
AutoMakers EV production ramp signal
Strong
30% attribution
Customer order velocity (last 60 days)
+18% trend
25% attribution
Seasonal Q2 uptick pattern
Historically +8-10%
19% attribution
Competitor capacity constraint (supply shift)
Detected
15% attribution
Historical forecast bias correction
-8.2% under-forecast
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 planning
How much can AI improve demand forecast accuracy in manufacturing?
Graph-based models typically improve quarterly forecast accuracy by 10-20% (measured as reduction in MAPE). For a manufacturer with 8-10% forecast error, this means reducing error to 5-7%. The improvement is largest for product lines with diverse customer bases and exposure to multiple end-markets, where relational signals provide the most value. Product lines with concentrated customers and stable demand see smaller but still meaningful improvement (5-10%).
Can demand planning AI replace S&OP meetings?
No, but it makes them more productive. The model provides a data-driven baseline forecast that incorporates signals most S&OP teams discuss qualitatively (customer pipeline changes, market shifts, competitive moves). This shifts S&OP from debating what the baseline forecast should be to discussing where to deviate from the model's recommendations based on judgment and intelligence the model cannot access.
How far ahead can AI forecast manufacturing demand?
Useful accuracy extends 1-4 quarters, depending on your industry's demand visibility. For make-to-stock manufacturing with short lead times, 1-2 quarter forecasts are most actionable. For capital equipment and long-lead-time manufacturing, 3-4 quarter forecasts drive capacity and procurement decisions. Accuracy degrades roughly 3-5% per quarter of forecast horizon.
What data sources improve demand forecasting accuracy?
Beyond historical orders, the highest-value data sources are: customer end-market indicators (industry production indices, earnings reports), competitor signals (capacity changes, pricing shifts), and order velocity trends (acceleration or deceleration in customer ordering patterns). Most of this data exists in CRM notes, market research subscriptions, and public financial filings. Structuring and connecting this data is the main challenge.
How does demand planning AI handle demand shocks and unprecedented events?
Graph-based models handle shocks better than time-series because they detect demand shifts through the customer network before they show up in your order book. When a customer's end-market weakens, the model sees early signals (order velocity deceleration, smaller order sizes) before the quarterly aggregate changes. For truly unprecedented events with no historical analog, the model provides a baseline that humans then adjust. No model replaces judgment for black swan events.
Bottom line: A $1B manufacturer saves $8-12M per year by improving quarterly demand accuracy 15%. Kumo's demand graph connects customer end-market signals, order velocity trends, and competitive dynamics that time-series models treat as flat exogenous variables.
Related use cases
Explore more manufacturing 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.
