3 Demand Planning and Forecasting

From demand estimates to coordinated plans

Demand planning estimates what customers are likely to buy and turns that estimate into a coordinated plan for supply, inventory, capacity, and service. A forecast is an informed estimate rather than a promise or target. Effective planning combines statistical forecasts with business knowledge and compares forecasts with actual results.

Recognizing demand patterns and data limitations

Demand is observed over time, often by product, location, and customer segment. Before choosing a forecasting method, examine the data and identify its main features.

• Level and variability: Level is the typical quantity; variability describes how much it fluctuates. Stable items with regular sales are usually easier to forecast than items with large, unpredictable swings.

• Trend: A sustained increase or decrease, such as rising sales as a product gains customers.

• Seasonality: A pattern that repeats at a known interval, such as higher weekly sales on weekends or annual demand for winter clothing. A series can have more than one seasonal pattern.

• Cycles and irregular events: Longer, less regular movements may reflect economic conditions. Promotions, price changes, product launches, competitor actions, and supply disruptions can also cause sudden changes. A planned event should generally be modeled as an explanatory factor rather than mistaken for normal baseline demand.

• Intermittency: Some items have many periods with zero demand. If the few nonzero orders are also large or irregular, demand is often called lumpy. Such patterns can occur in service parts and specialized products, and ordinary methods may perform poorly. Croston’s method separately smooths the size of nonzero demands and the time between them, then combines those estimates.

Interpreting demand data

Point-of-sale (POS) sales are closer to consumer purchases than shipments to a retailer, which may include inventory buildup. However, sales can understate true demand when a product is out of stock. Flag stockouts, incomplete records, one-time orders, and unusual events rather than treating every observation as ordinary demand.

Store-level event forecasts can use POS history and event details such as price, dates, and displays.

Choosing forecasting methods

No single forecasting method is best for every item or horizon. Start with a simple benchmark, test suitable alternatives using historical data, and select a method that matches both the demand pattern and the decision being supported.

Qualitative and judgmental methods

Expert judgment is useful when history is limited or the future is changing, such as for a new product, a major promotion, or a market disruption. Sales, marketing, customers, and operations may contribute relevant information. Record the assumptions behind adjustments so they can be checked afterward; unsupported optimism or duplicated adjustments can introduce bias.

Time-series methods

Time-series methods use an item’s own past demand:

• Naïve forecast: Use the most recent actual value as the next forecast. This simple benchmark can be surprisingly useful.

• Seasonal naïve forecast: Use demand from the corresponding season in the latest cycle. For example, last December’s sales can be used for this December’s initial forecast.

• Moving average: Average demand across a recent window to smooth short-term noise. A short window responds quickly to changes but is more sensitive to random variation; a longer window is smoother but slower to react.

• Exponential smoothing: Give recent observations more weight, with older observations receiving progressively less. Simple smoothing suits relatively level demand; extensions can model trend and seasonality.

Causal and event-based methods

Regression and related methods estimate how demand relates to predictors such as price, advertising, weather, or day of the week. A predictor must be available, or itself forecastable, for the period ahead. For example, a retailer can estimate promotional lift from past events with similar discounts and display activity. Correlation alone does not establish that a predictor causes demand.

Methods can be combined: create a statistical baseline, add documented effects for known events, and then review the result with informed business partners. Forecasts may be prepared at product, location, and time levels and reconciled with higher-level plans so that detail and totals remain consistent.

Evaluating forecast performance

Evaluate forecasts on data that were not used to fit the method. Rolling-origin evaluation repeatedly trains on earlier periods and tests forecasts on later ones. Match the evaluation horizon to the real decision: a one-week-ahead measure may not describe a forecast used to plan three months ahead.

Let actual demand be $A_t$, forecast demand be $F_t$, and signed error be defined as:

With this convention, a positive error means the forecast was too low, while a negative error means it was too high.

• Mean absolute error (MAE): $\text{mean}(|e_t|)$. It reports the typical miss in the original units.

• Root mean squared error (RMSE): $\sqrt{\text{mean}(e_t^2)}$. Squaring gives greater influence to large misses.

• Mean absolute percentage error (MAPE): The average of $100|e_t|/A_t$. It is easy to interpret as a percentage, but is undefined when actual demand is zero and can become extreme when actuals are near zero. It is therefore a poor sole measure for intermittent demand.

• Bias: $\text{mean}(e_t)$, the average signed error. A persistent positive value indicates underforecasting; a persistent negative value indicates overforecasting. Bias can be near zero even when individual errors are large, so review it alongside an error measure.

A useful score depends on the decision and the data. Compare methods using consistent definitions, product and location groupings, units, and forecast horizons. Check both aggregate results and important item-level misses: a large forecast error may create excess inventory, a stockout, or a capacity problem even if an overall score looks acceptable. Avoid relying on MAPE where actual values can be zero or close to zero.

Worked example

If a product’s forecast is 100 units and actual demand is 120, the signed error under this convention is $120-100=20$ units, so the forecast underpredicted demand. Its absolute error is 20 units. The percentage error is about $16.7\%$ when actual demand is the denominator.

Collaborating across the supply chain

Collaborative Planning, Forecasting, and Replenishment (CPFR) is a process in which supply-chain partners share plans and demand information, agree on forecasts where appropriate, and coordinate replenishment. Collaboration is especially valuable when one partner’s promotion, ordering, or inventory decision affects another partner’s production and delivery plans.

A workable process typically:

1. Sets shared objectives, product and location scope, data definitions, responsibilities, and review frequency.

2. Shares relevant information, such as POS movement, inventory, event calendars, prices, promotions, and supply constraints.

3. Builds or reviews the forecast, identifies material differences, and resolves them using evidence and stated assumptions.

4. Translates the agreed demand plan into orders, replenishment, production, and distribution plans.

5. Monitors exceptions and compares event results with the plan, then applies findings to future cycles.

For example, a retailer and supplier planning a promotion can share its dates, price, display plan, store-level sales history, and inventory. They can estimate consumer demand together, aggregate store needs into distribution-center requirements, and adjust replenishment as the event unfolds. A shared calculation method matters because a reported “forecast accuracy” percentage can mean different things when partners use different formulas or forecast horizons.

Connecting demand plans to supply decisions

Demand management coordinates demand with the organization’s ability to supply it. It includes four connected activities:

• Plan: Develop and update estimates of expected demand, including known orders and relevant market information.

• Communicate: Share the plan and its assumptions with sales, marketing, procurement, production, inventory, and logistics teams.

• Influence: Where appropriate, use pricing, promotion, product substitution, or other actions to shape demand. Such actions can shift purchases between periods, but do not automatically increase total underlying demand.

• Prioritize: If supply cannot satisfy every request, allocate available product according to agreed business rules, such as customer commitments, service needs, or strategic importance, and communicate the decision.

Revisit the demand plan as new orders, sales, event details, and constraints become known. Time fences and approval rules can help manage near-term changes: within a defined horizon, schedule changes may require higher-level approval, while later periods remain more flexible.

Together, demand planning and demand management create a shared, revisable estimate for business decisions. Identify demand patterns and data limitations, choose appropriate methods, evaluate forecasts at the relevant horizon, and track both error and bias. Collaboration improves visibility and alignment, while demand management connects estimates to communication, demand-shaping choices, supply constraints, and customer priorities.