Timedora

Demand Forecasting: From Data to Decisions

A practical guide to defining demand forecasts, understanding demand data, choosing and evaluating methods, and accounting for uncertainty in planning.

Define the Forecast and Understand the Data

Define the forecast before choosing a method

A estimates how much of a product or service customers will require at a particular place and over a particular period. Organizations use forecasts to plan inventory, staffing, capacity, and purchasing. But a forecast is an estimate—not a sales target or a promise.

Start by specifying what is being forecast (for example, a product at a particular location), how often the estimate is needed, and how far ahead it must extend. These choices depend on the decision: a method and horizon useful for one decision may not be useful for another. A forecast is most helpful when its scope, method, and time horizon match the decision it will inform.

Read the patterns in demand

A time series can show several patterns, sometimes at once:

  • Level is typical demand when there is no clear trend or seasonality.

  • Trend is a sustained increase or decrease over time.

  • Seasonality repeats at a known, fixed interval, such as a weekly or yearly pattern.

  • Cycles rise and fall without a fixed repeating interval and are often associated with broader economic conditions.

  • Irregular variation consists of unpredictable changes, including random fluctuations and unusual events.

The distinction between seasonality and cycles is the regularity of repetition: seasonality has a fixed calendar frequency; cycles do not. For example, a business might have rising overall demand, a weekly peak on weekends, and occasional irregular spikes. Seeing one pattern does not rule out the others.

Check whether sales represent demand

Recorded sales are not always the same as what customers wanted. If an item was out of stock, for example, sales may understate demand because some customers could not buy it. Promotions, price changes, holidays, and other unusual events can also affect observed sales. Before forecasting, check for missing periods, stockouts, one-time events, and errors. Otherwise, a method may treat a gap or an unusual observation as a normal feature of demand.

Choose a Method That Fits the Demand

Match the method to the pattern and decision

No single forecasting method is best for every demand pattern or forecast horizon. First define the forecast’s scope and horizon, then examine the data and compare suitable methods on historical periods that were not used to fit them. Consider both the patterns in demand and what information will be available when the forecast is made.

Two simple benchmarks are useful starting points:

  • A uses the latest observed demand as the forecast for the next period. It can work well when demand changes little.

  • A uses demand from the corresponding season in the previous cycle. For example, it could use the previous Monday’s demand to forecast this Monday’s demand. It is a useful benchmark when seasonality is strong.

These benchmarks are not automatically suitable for every situation. They provide straightforward comparisons for evaluating more elaborate methods.

Methods that smooth or explain demand

A moving average averages demand over a recent number of periods. The window length involves a trade-off: a short window responds quickly to changes but may be noisy, while a longer window smooths fluctuations but responds more slowly.

Exponential smoothing gives more weight to recent observations while retaining information from older ones. Extensions can represent changes in level, trend, and seasonality. The appropriate version depends on which patterns are relevant; smoothing should not be mistaken for a guarantee that demand will follow its past pattern.

Regression with explanatory variables relates demand to factors such as price, promotions, holidays, weather, or economic conditions. It can help when those factors matter, but useful forecasts depend on having reliable future values—or explicit assumptions—for the explanatory factors. If a future promotion is uncertain, for example, its assumed presence or absence should be clear.

Use judgment where the data are limited or conditions are new

Judgmental forecasting uses structured expert knowledge when historical data are limited or conditions are new, such as at a product launch. Record the assumptions behind the judgment and evaluate the resulting forecasts afterward.

The examples show why identifying the pattern matters. If weekly demand is consistently higher in December, a seasonal method can use earlier Decembers rather than treating each December increase as evidence of a new long-term trend. By contrast, sporadic demand with many zero-demand periods may be better handled by a method designed for intermittent demand than by ordinary smoothing.

In practice, method choice depends on the demand pattern, the horizon, and the available information. Compare plausible candidates rather than assuming that one method will suit every product or decision.

Test Forecast Accuracy and

Interpret errors before comparing methods

For period or forecast case tt, define as

et=yt−y^t,e_t = y_t - \hat{y}_t,

where yty_t is actual demand and y^t\hat{y}_t is forecast demand. A positive error means actual demand exceeded the forecast; a negative error means the forecast was too high.

Accuracy and describe different things. Accuracy concerns the size of errors. concerns whether errors tend to be consistently positive or negative. A method can have errors in both directions, while a persistent direction of error is a reason to examine rather than looking only at average error size.

Choose an error measure for the question

Three common measures summarize error in different ways. Across nn forecast cases, they can be written as

MAE=1n∑t=1n∣et∣,RMSE=1n∑t=1net2.\mathrm{MAE}=\frac{1}{n}\sum_{t=1}^{n}|e_t|, \qquad \mathrm{RMSE}=\sqrt{\frac{1}{n}\sum_{t=1}^{n}e_t^2}.
  • Mean absolute error (MAE) averages the absolute errors in the original units. It is easy to interpret for one product or for comparable series.

  • Root mean squared error (RMSE) squares the errors before averaging and then takes the square root. Squaring gives larger misses more weight, so RMSE is more sensitive to occasional large errors.

  • Mean absolute percentage error (MAPE) averages absolute errors as a percentage of actual demand. In formula form, it averages 100∣et∣/∣yt∣100|e_t|/|y_t| across periods. It is undefined when actual demand is zero and can be misleading when actual demand is close to zero.

No one measure suits every situation. Choose measures that fit the data and the costs of the decision.

Test forecasts on later data

Evaluate a method on data that were not used to fit it. For a time series, rolling-origin evaluation repeatedly fits a method using past observations and tests it on later periods. This resembles the practical task of using past information to forecast what comes next.

Compare candidate methods with simple benchmarks and evaluate at the horizons that matter operationally. A one-week-ahead result may not represent a twelve-week-ahead forecast. A method with a smaller error on one week may not be reliably better; results across relevant test periods provide a more useful comparison.

Worked example: Actual weekly demand is 110 units and the forecast is 100 units. Subtract the forecast from actual demand:

et=110−100=+10 units.e_t = 110 - 100 = +10\text{ units}.

The positive sign means demand exceeded the forecast by 10 units. This single result describes one ; it cannot establish that the method is generally more or less accurate than another. That requires comparing errors across relevant, unused test periods using measures suited to the decision.

Use Uncertainty in Planning

Look beyond a single forecast number

A point forecast gives one central estimate of future demand. A gives a range of plausible future values at a stated coverage level, such as 80% or 95%. A 95% interval is intended to contain the future value in about 95% of comparable forecasting situations, assuming the model and its uncertainty estimates are appropriate. It does not guarantee that the value in any particular period will fall inside the range.

Intervals generally widen as the forecast horizon increases, because more can change over a longer period. Uncertainty may also be high when there is little history, demand is volatile, or future conditions differ from the past.

Make assumptions part of the plan

If a forecast depends on uncertain inputs—such as a planned promotion or future price—make the assumptions explicit and consider alternative scenarios. A scenario helps decision-makers see how a forecast could differ under different assumed conditions; it does not remove the uncertainty about which conditions will occur.

Use uncertainty in planning rather than hiding it behind one number. For inventory, a decision may consider expected demand and the consequences of running short versus holding excess stock. The forecast informs the choice, but the desired stock level also depends on costs, service objectives, and risk tolerance.

  • : A sequence of demand observations recorded at regular intervals.