Demand Sensing for Manufacturers: Shortening the Forecast Horizon
Demand sensing is short-horizon forecasting that uses recent, granular signals - daily orders, channel sell-through, open quotes, point-of-sale data, and customer schedule changes - to correct the statistical forecast inside the next one to thirteen weeks. It does not replace the mid-range forecast that drives S&OP and capacity planning; it corrects the near term where the statistical model is weakest and where inventory and expedite decisions are actually made. Manufacturers that implement it well typically reduce forecast error 20 to 40 percent in the zero to four week window against a statistical baseline.
Why the Near Horizon Needs Different Math
Traditional statistical forecasting - exponential smoothing, Holt-Winters, ARIMA - fits monthly history and projects seasonality and trend. It is the right tool for a six to eighteen month horizon and the wrong tool for next week, because it is structurally blind to information that arrived after the last period close. Demand sensing models operate on daily or weekly buckets and consume signals with much shorter latency: order intake in the last ten days, customer released schedules against blanket orders, quote conversion rate, backlog aging, and distributor sell-through versus sell-in. The mathematical approach is usually gradient-boosted trees or a regularized regression that blends the statistical baseline with these features, rather than a single time series model trying to do both jobs.
The Signals That Actually Move Accuracy
In discrete manufacturing, ranked by contribution, the useful signals are customer released schedules against long-term agreements, recent order intake velocity by item and customer, open quote pipeline with historical conversion rates by customer segment, and distributor or channel inventory position where you sell through partners. Weather and macroeconomic indices help in a narrow set of categories and are frequently oversold. Be ruthless about signal latency - a POS feed that arrives eleven days after the fact contributes almost nothing to a two-week horizon. Also separate demand from shipments: shipment history is censored by your own stockouts and capacity limits, so a model fit on shipments learns your constraints rather than your customers' actual demand.
- Customer released schedules against blanket orders, the single strongest near-term signal in build-to-order
- Order intake velocity in the last 7-14 days by item-customer versus the same window prior periods
- Quote pipeline weighted by segment-specific historical win rates and typical quote-to-order lag
- Channel sell-through and partner inventory position where you sell through distribution
Measuring Demand Sensing Honestly
Measure at the horizon and granularity where the decision is made, which is usually item-site at weekly buckets for lags of one to four weeks. Use weighted MAPE for high-volume items and mean absolute scaled error for intermittent ones, since MAPE is unstable and misleading when actuals are near zero. Always report bias alongside error, because a model that is consistently low will quietly drain inventory. The essential discipline is forecast value add: compare the sensed forecast against both a naive forecast and the existing statistical baseline over a holdout window. If demand sensing does not beat the naive forecast at lag one, the problem is data latency or leakage, and no amount of model tuning will fix it.
- Evaluate at item-site weekly buckets across lags 1 through 4, not at aggregate monthly level
- Report weighted MAPE plus bias, and use MASE for intermittent items rather than MAPE
- Benchmark against both naive and statistical baselines on a strict out-of-time holdout
- Audit for leakage: any feature unavailable at true forecast time invalidates the whole result
Connecting Sensed Demand to MRP Without Chaos
The value of a better near-term forecast is realized only if it changes replenishment, and the risk is that it changes replenishment too often. Inside the frozen zone defined by cumulative lead time, sensed demand should raise exceptions for a planner rather than automatically re-drive MRP. Outside the frozen zone it can update forecast records directly. In SyteLine and CloudSuite Industrial that means writing to forecast records that feed MRP or APS with a change threshold so small deltas do not churn planned orders; in Infor LN it is the demand input to Enterprise Planning; in M3 it feeds MPS and DRP. Set a minimum change threshold, often 10 to 15 percent or a minimum unit quantity, and log every automated adjustment so planners can see what moved and why.
How Netray AI Agents Deliver Demand Sensing on Infor ERP
Netray builds demand sensing agents that train on your own SyteLine, LN, or M3 history rather than a generic industry model. The agents assemble features from order intake, released customer schedules, quote pipeline, and backlog directly out of ERP tables, fit and revalidate models on a rolling basis, and publish forecast adjustments through an approval workflow with the projected inventory and service impact attached. Because they run inside your environment or on-premises, customer order data - which is frequently competition-sensitive and, in defense work, export controlled - never leaves your boundary. Clients typically see 20 to 35 percent reduction in one-to-four week forecast error and a corresponding drop in expedite freight within two quarters.
- Feature engineering directly from ERP order, schedule, quote, and backlog tables with no data export
- Rolling revalidation with automatic drift detection so degraded models are flagged before they mislead
- Change-thresholded write-back to forecast records with a full audit log of every automated adjustment
- On-premises deployment for ITAR and CUI environments where customer demand data cannot leave the site
Frequently Asked Questions
What is the difference between demand sensing and demand planning?
Demand planning produces the mid-range consensus forecast that drives S&OP, capacity, and budgets, typically monthly buckets over 12 to 24 months. Demand sensing corrects the near term, usually daily or weekly buckets across one to thirteen weeks, using recent signals the statistical model cannot see. They are complementary: sensing without a sound mid-range plan leaves capacity decisions unsupported, and planning without sensing leaves near-term inventory badly positioned.
How much data do you need for demand sensing to work?
Roughly two to three years of transactional order history at daily or weekly granularity, plus at least one full seasonal cycle of the signal features you intend to use. More important than volume is latency and cleanliness: signals must be available at true forecast time, and demand must be separated from shipments, since shipment history is censored by your own stockouts and capacity limits.
Does demand sensing work for low-volume build-to-order manufacturers?
Yes, but the signals differ. For build-to-order and engineer-to-order work, statistical patterns in shipment history are weak, and the value comes from quote pipeline conversion, customer released schedules, and backlog aging rather than seasonality. Accuracy gains show up more in component and long-lead material planning than in finished goods, which is usually where the working capital sits anyway.
Key Takeaways
- 1Why the Near Horizon Needs Different Math: Traditional statistical forecasting - exponential smoothing, Holt-Winters, ARIMA - fits monthly history and projects seasonality and trend. It is the right tool for a six to eighteen month horizon and the wrong tool for next week, because it is structurally blind to information that arrived after the last period close.
- 2The Signals That Actually Move Accuracy: In discrete manufacturing, ranked by contribution, the useful signals are customer released schedules against long-term agreements, recent order intake velocity by item and customer, open quote pipeline with historical conversion rates by customer segment, and distributor or channel inventory position where you sell through partners. Weather and macroeconomic indices help in a narrow set of categories and are frequently oversold.
- 3Measuring Demand Sensing Honestly: Measure at the horizon and granularity where the decision is made, which is usually item-site at weekly buckets for lags of one to four weeks. Use weighted MAPE for high-volume items and mean absolute scaled error for intermittent ones, since MAPE is unstable and misleading when actuals are near zero.
Put this into numbers
Free interactive tools for exactly this problem. No signup to use them.
Forecast Accuracy (MAPE) Calculator
Turn your forecast and actual history into MAPE, accuracy, and bias figures, then price what that error is costing you across the year.
Free ToolCycle Time & Capacity Calculator
Translate cycle time, parallel stations, shift pattern, and realistic utilization into hourly, daily, and annual production capacity.
Free ToolS&OP Maturity Assessment
Assess your sales and operations planning process across cadence, data quality, scenario capability, and accountability, and get a staged roadmap to the next maturity level.
Terms used in this article
Your statistical forecast is weakest exactly where inventory decisions get made. Ask Netray about demand sensing agents trained on your own Infor ERP history.
Related Resources
Sales and Operations Planning (S&OP): A Practical Guide
A practical sales and operations planning guide for manufacturers: monthly S&OP cadence, demand and supply review, and executive sign-off in Infor ERP.
ERPMulti-Echelon Inventory Optimization
Multi-echelon inventory optimization explained: how MEIO sets safety stock across plants, DCs, and suppliers to cut inventory 15-30% without service loss.
ERPSupply Chain Visibility: What Actually Works
Supply chain visibility platforms compared: what multi-tier tracking, EDI/API telemetry, and exception alerting actually deliver for manufacturers.