Overview
Demand forecasting is the estimation of future customer demand using historical data, assumptions, statistical models, or expert judgment. It supports decisions ranging from near-term replenishment to long-term capacity investment. Forecasting methods vary with the available data, the planning horizon, and the decision being supported.
Demand sensing is a near-term forecasting practice that uses recent information—such as current orders, point-of-sale activity, inventory availability, and promotional changes—to revise an existing forecast or produce a short-horizon estimate. It is generally treated as a complement to, rather than a replacement for, conventional forecasting. The related statistical idea of using timely data to estimate current conditions is often called nowcasting; a later published version of that work appeared in 2012 as "Predicting the Present with Google Trends".
The distinction matters in logistics because a monthly forecast may support procurement and capacity planning, while a daily update may help prioritize warehouse replenishment, labor, and allocation. Neither approach eliminates uncertainty, and a revised estimate is operationally useful only when stock, transport, or labor can still be adjusted.
Etymology and Origins
Forecasting combines the idea of estimating with anticipation of a future event. The verb forecast means to calculate or predict a future event or condition. In business planning, the term describes systematic estimates based on observations and assumptions. Modern forecasting terminology and methods cover both judgmental and quantitative approaches.
Sensing refers to detecting signals or changes. In supply-chain usage, demand sensing emphasizes the rapid observation and interpretation of demand-related signals. The compound term does not have a universally established first documented use or a single agreed technical definition; vendors and practitioners apply it to different combinations of data, models, and update frequencies. Its relationship to nowcasting is useful, although the terms are not interchangeable in every application.
Historical Development
Demand planning predates computer systems: businesses estimated future requirements from sales records, seasonal patterns, and managerial judgment. In 1957, Charles C. Holt described methods for updating level and trend estimates by exponential smoothing, later reprinted and widely cited as a foundation of Holt's linear trend method. The 2004 reprint identifies the work as a 1957 Office of Naval Research memorandum. Computerized methods later made it practical to update and compare models across many products.
In 1970, George E. P. Box and Gwilym M. Jenkins published Time Series Analysis: Forecasting and Control, which popularized a systematic approach to autoregressive integrated moving average models. ARIMA methods remain part of the forecasting toolkit alongside exponential smoothing and newer machine-learning methods. In 1972, J. D. Croston addressed forecasting intermittent demand for stock control, a problem especially relevant to slow-moving spare parts.
By the 1990s, enterprise planning systems and electronic data interchange helped organizations consolidate demand and supply information. Lee, Padmanabhan, and Whang's 1997 account of the bullwhip effect showed how ordering behavior and limited visibility could amplify demand fluctuations upstream. Barcode standards such as the UPC were introduced decades earlier, but in the 2000s and 2010s more frequent electronic point-of-sale feeds and greater computing capacity made shorter forecast cycles more practical. Hyunyoung Choi and Hal Varian's work on predicting the present illustrated how timely search data could inform near-term estimates. Cloud systems and machine learning have since made frequent signal processing more accessible, though neither is required to practice demand sensing.
Core Concepts
Conventional demand forecasting
A forecast starts with a defined target, such as units sold per product and region per week. A model may use historical demand, seasonality, price, promotions, and other explanatory variables. Depending on the decision, it may produce a point estimate, a range, or a probability distribution. Time-series methods differ in how they represent trends and seasonal patterns.
Forecasts are usually generated on a recurring cycle and may be reviewed by sales, finance, and operations. A six-month forecast, for example, can inform supplier commitments even when its daily detail is uncertain.
Demand sensing
Demand sensing typically begins with a baseline forecast and incorporates newly available signals. A system might compare this week's actual orders and store sales with the baseline, adjust the remaining short-term estimate, and flag a material deviation. The method can use simple rules, statistical models, or machine learning; rapid updates alone do not guarantee accuracy.
The distinction between forecasting and nowcasting with current information helps explain why a near-term estimate can change as fresh data arrive. In logistics, the estimate still has to be translated into an actionable order, transfer, allocation, or staffing decision.
Comparison and Planning Roles
| Dimension | Demand forecasting | Demand sensing |
|---|---|---|
| Typical horizon | Days to years, depending on the decision | Usually days to weeks |
| Update frequency | Often weekly or monthly, but can be more frequent | Often daily or more frequently |
| Main inputs | Historical demand, seasonality, plans, and causal factors | Recent orders, sales, availability, and other timely signals, often alongside a baseline |
| Common decisions | Purchasing, budgets, capacity, and longer-term inventory | Allocation, replenishment, labor, and near-term exceptions |
| Main risk | Missing a recent change | Overreacting to noise or incomplete data |
These are tendencies rather than rigid categories: a forecast can be updated daily, and sensing is itself a forecasting activity. The relevant comparison is whether the additional signals improve a specific decision at a specific horizon. Planners should also consider the bullwhip effect: reacting to every order fluctuation can amplify variability rather than reduce it.
For example, a distributor forecasts 1,000 units for a coming week. Midweek sales and open orders suggest stronger demand, so its sensed estimate rises to 1,150 units. If 200 units can be transferred from a nearby warehouse before the weekend, planners might authorize a partial transfer. The revised estimate is not proof that 1,150 units will sell; availability, transport capacity, and the cost of being wrong still govern the decision. A stock replenishment workflow is one example of the execution process that can follow such a decision.
Data, Systems, and Applications
Common inputs include point-of-sale transactions, ecommerce orders, shipments, inventory balances, returns, promotions, and calendar events. These signals measure different things: shipments can reflect allocation decisions, while sales can understate demand when an item is out of stock. Planners should distinguish observed sales from estimated unconstrained demand and document any adjustments.
A planning system may receive inventory and order information from a warehouse management system, calculate a forecast externally, and return approved actions for execution. Inventory synchronization and purchase-order management are examples of operational data and purchasing workflows; their presence alone does not establish that a business uses a demand-sensing model. A documented cycle-counting process can help teams check the inventory records used to assess whether a proposed response is feasible.
Applications differ by sector. Retailers may use recent sales and promotion activity to allocate stock among stores; consumer-goods suppliers may combine retailer sales with incoming orders; and distributors may use recent order patterns to schedule picking labor. For products with sparse or intermittent sales, aggregation across locations or longer periods may be more useful than frequent SKU-level adjustments. A method designed for intermittent demand should be assessed against an appropriate baseline. The suitable method depends on the forecasting problem and available data.
Measurement and Limitations
Evaluate both approaches using forecasts recorded before the outcome was known. Compare the sensed estimate with the baseline at the same product, location, and horizon, using rolling historical tests and, where feasible, a controlled operational pilot. Forecast accuracy measures must be chosen for the use case; no single metric captures every consequence.
Useful measures include:
- Mean absolute error (MAE): the average absolute difference between actual and forecast demand, expressed in units.
- Weighted absolute percentage error (WAPE): the sum of absolute errors divided by total actual demand over the evaluation set. It is undefined when total actual demand is zero. It is also called the weighted mean absolute percentage error.
- Bias: the average signed forecast error, with the sign convention stated explicitly; persistent bias can indicate systematic over- or under-forecasting.
- Operational outcomes: service level, estimated lost sales, waste, expedites, and inventory costs, assessed alongside forecast error.
For a fair test, compare forecasts for the same target and decision point. For example, compare a baseline available on Wednesday with a sensed forecast made on Wednesday for remaining Thursday–Sunday demand, not with a forecast of the full week made earlier. Out-of-sample evaluation helps prevent a model from being judged only on the data used to fit it. If observed sales are constrained by stockouts, a comparison against sales alone does not establish accuracy against unobserved demand.
Frequent signals can mislead when transactions arrive late, promotions are misclassified, or stockouts suppress sales. Adjustments can also create unnecessary transfers or orders if planners chase noise. More accurate near-term forecasts may have limited value when supplier lead times exceed the sensing horizon. Sharing downstream information can help, but the bullwhip effect shows that incentives, batching, and ordering policies also matter.
Best Practices
- Start with the decision. Define the product-location level, horizon, lead time, and action that the estimate will support.
- Establish a baseline. Compare sensing with a seasonal or other suitable forecast, not merely with the absence of a forecast. Use out-of-sample accuracy testing at matching horizons.
- Validate and timestamp inputs. Reconcile units of measure, returns, cancellations, inventory records, and data arrival times. Exclude information that would not have been available when a historical forecast was made. Where appropriate, check records through a documented cycle-counting process.
- Control operational responses. Set thresholds, review exceptions, and account for transfer costs and supplier constraints before changing orders. Connect approved decisions to documented replenishment processes.
- Monitor outcomes and ownership. Assign responsibility for model performance and overrides. Review forecast bias, service, and inventory effects by segment, and retire signals or rules that do not improve decisions.