4. Time-Series Forecasting Models

Introduction and Context

Demand forecasting is a central planning activity in logistics and supply chain management. Decisions about procurement, production, inventory, transportation capacity, warehouse labour, and customer service all depend on an estimate of future demand. A forecast is not a promise about what will happen; it is a systematic estimate based on information available at the time the forecast is prepared.

Time-series forecasting uses historical observations of a variable, arranged in chronological order, to estimate its future values. In supply chains, the variable may be daily orders, weekly dispatches, monthly sales, spare-parts consumption, or the number of deliveries to a distribution centre. The distinctive feature of a time series is that observations are not independent of one another. Demand in one period may be related to demand in previous periods, seasonal events, promotions, weather, or broader economic conditions.

This chapter develops a practical progression from simple forecasting rules to more advanced models. Naïve methods establish useful benchmarks. Moving averages smooth random fluctuations. Exponential smoothing gives greater weight to recent observations. Holt’s method incorporates trend, while Holt–Winters models address both trend and seasonality. Decomposition separates a series into interpretable components, and ARIMA provides a structured framework for modelling autocorrelation and differencing. The emphasis throughout is on selecting, validating, interpreting, and using models rather than treating forecasting as a purely mathematical exercise.

Learning Objectives

After studying this chapter, students should be able to:

1. Explain the structure and terminology of time-series data.

2. Apply naïve, seasonal naïve, moving-average, and exponential-smoothing methods.

3. Distinguish level, trend, seasonality, cyclical movement, and irregular variation.

4. Use Holt and Holt–Winters models for demand series with trend and seasonal patterns.

5. Explain decomposition and the basic logic of ARIMA models.

6. Compare forecasting models using appropriate error measures and time-based validation.

7. Diagnose common forecasting problems and recommend suitable models for logistics decisions.

8. Evaluate the managerial, technological, and ethical implications of automated forecasts.

Reading a Demand Series: Components and Patterns

A time series is commonly represented as observations \(Y_t\), where \(t\) denotes a time period. A useful starting point is to ask what structure appears to be present in the data.

The level is the underlying average magnitude of the series. A product may have a normal weekly demand of approximately 1,000 units. A trend is a persistent upward or downward movement, such as increasing demand for electric-vehicle components. Seasonality is a regularly repeating pattern associated with the calendar or operating cycle—for example, higher beverage demand in summer or increased parcel volumes before a major festival. Cyclical movement refers to longer and less regular fluctuations, often linked to economic or industry conditions. Irregular variation consists of random shocks, data errors, disruptions, and other movements that are difficult to predict.

Two common representations are the additive and multiplicative forms. In an additive model,

\[Y_t = T_t + S_t + E_t\]

where the effects of trend, seasonality, and error are added. This is appropriate when seasonal fluctuations are approximately constant in absolute size. In a multiplicative model,

\[Y_t = T_t \times S_t \times E_t\]

seasonal variation grows or declines with the level of demand. Multiplicative seasonality is often suitable when a series with a larger overall volume also has larger seasonal peaks. Logarithmic transformation can sometimes convert multiplicative relationships into an additive form.

Before modelling, the analyst should confirm the time interval, check for missing periods, investigate stockout distortion, identify abnormal promotions, and distinguish demand from observed sales. If an item was unavailable for part of a month, recorded sales may understate actual customer demand. Forecasting poor-quality observations with sophisticated algorithms does not produce a reliable plan.

Benchmark Models: Why Simple Methods Matter

A model should earn its complexity. Simple methods provide transparent forecasts and essential benchmarks against which advanced models must be evaluated.

Naïve and Seasonal Naïve Forecasts

The naïve forecast assumes that the next observation will equal the most recent observation:

\[\hat{Y}_{t+1}=Y_t\]

For example, if weekly demand was 420 units, the forecast for the following week is 420. Although basic, this method can be effective for stable series and is difficult to outperform consistently when demand is highly volatile.

The seasonal naïve method uses the observation from the corresponding previous season:

\[\hat{Y}_{t+h}=Y_{t+h-m}\]

where \(m\) is the seasonal period. For monthly data with annual seasonality, \(m=12\). If demand in December last year was 2,400 units, the seasonal naïve forecast for the coming December is 2,400 units. This is a particularly important benchmark for seasonal supply-chain demand.

Moving Averages

A simple moving average forecasts the next period using the mean of the most recent \(k\) observations:

\[\hat{Y}_{t+1}=\frac{Y_t+Y_{t-1}+\cdots+Y_{t-k+1}}{k}\]

Suppose the last three weekly demands are 480, 510, and 450 units. The three-period moving-average forecast is 480 units. A larger window produces greater smoothing but responds more slowly to genuine changes. A smaller window responds more quickly but retains more random variation.

A weighted moving average assigns different weights to recent observations, with weights usually declining as observations become older:

\[\hat{Y}_{t+1}=w_1Y_t+w_2Y_{t-1}+\cdots+w_kY_{t-k+1}\]

where the weights sum to one. Weight selection should be justified through validation rather than convenience.

Moving averages are easy to explain and useful for stable demand, but they have important weaknesses. They lag behind turning points, do not naturally forecast long horizons, and require special treatment when trend or seasonality is present. A moving average should therefore be viewed as a smoothing and benchmark technique, not a universal solution.

Exponential Smoothing: Updating the Forecast as New Data Arrive

Exponential smoothing produces forecasts by combining the previous forecast with the latest observation. Simple exponential smoothing is suitable for data with a relatively stable level and no systematic trend or seasonality:

\[F_{t+1}=\alpha Y_t+(1-\alpha)F_t\]

where \(F_{t+1}\) is the next forecast and \(0<\alpha<1\). A high value of \(\alpha\) gives greater emphasis to recent demand and makes the forecast more responsive. A low value creates a smoother forecast that is less affected by individual shocks.

For example, if the previous forecast was 500 units, actual demand was 560, and \(\alpha=0.30\), the updated forecast is:

\[F_{t+1}=0.30(560)+0.70(500)=518\]

The initial forecast may be set equal to the first observation, an average of early observations, or an estimate obtained through optimization. In operational systems, the smoothing parameter can be selected by minimizing an error measure on historical data, but the result should still be reviewed for operational plausibility.

Simple exponential smoothing is attractive because it requires limited data storage, is computationally efficient, and can be updated immediately as new demand arrives. However, it is not designed to represent a changing trend or recurring seasonal pattern. Using it for strongly seasonal demand can produce systematic under- and over-forecasting.

Holt’s Method for Level and Trend

Holt’s linear method extends exponential smoothing by maintaining separate estimates of the current level and trend. One common formulation is:

\[L_t=\alpha Y_t+(1-\alpha)(L_{t-1}+B_{t-1})\]

\[B_t=\beta(L_t-L_{t-1})+(1-\beta)B_{t-1}\]

\[\hat{Y}_{t+h}=L_t+hB_t\]

Here, \(L_t\) is the estimated level, \(B_t\) is the estimated trend, \(\alpha\) controls level updating, and \(\beta\) controls trend updating. The trend may be damped when it is unlikely to continue indefinitely. A damped trend gradually reduces the contribution of the trend as the forecast horizon increases, which is often more realistic for mature products or capacity-constrained markets.

Holt’s method is useful when demand is rising or falling but does not exhibit a stable seasonal pattern. It can nevertheless overreact to temporary changes in slope and may generate implausibly high long-range forecasts if an observed trend is extrapolated without managerial review.

Holt–Winters Models: Combining Level, Trend, and Seasonality

Holt–Winters exponential smoothing extends Holt’s method by including a seasonal component. It is commonly used for monthly or weekly demand with a recurring pattern. The method has additive and multiplicative forms.

In the additive form, seasonal effects are measured in units. A seasonal increase of 200 units is approximately similar at low and high levels of demand. In the multiplicative form, seasonal effects are expressed as proportions. A seasonal index of 1.20 indicates demand approximately 20 percent above the underlying level.

The additive Holt–Winters model can be represented conceptually by updating:

• the level, using current demand adjusted for trend and seasonality;
• the trend, using changes in the estimated level; and
• the seasonal index, using the latest seasonally adjusted observation.

The forecast then combines these components. The exact equations vary according to the chosen initialization and notation, but the underlying logic is consistent: recent observations update level, trend, and seasonal behaviour at controlled rates.

The model is appropriate when the seasonal cycle is reasonably stable and sufficient historical data exist to estimate it. It is less suitable when seasonal timing changes, when demand is intermittent, or when the series is dominated by one-off promotions. A seasonal model should also use the correct seasonal period: 7 for daily data with a weekly pattern, 12 for monthly annual seasonality, and 52 for weekly annual seasonality, subject to the data and business context.

Decomposition as an Interpretive Framework

Decomposition separates a time series into components so that each can be examined independently. Classical decomposition estimates trend-cycle movement, seasonal indices, and residual variation. A moving average is often used to estimate the trend-cycle component, after which seasonal effects are calculated from deviations around the trend.

Decomposition is valuable for diagnosis as well as forecasting. A warehouse manager may discover that apparently unstable demand is actually a stable seasonal pattern combined with a gradual trend. A planner may identify a sudden residual spike caused by a promotion or supply disruption rather than a permanent change in demand.

Decomposition may be additive or multiplicative. The choice should be guided by plots and business interpretation. Seasonal indices should be estimated without allowing unusual events to distort the normal pattern. When interventions are frequent, the analyst may need regression with seasonal indicators, event variables, or a model that explicitly represents promotions.

Decomposition does not automatically provide a superior forecast. Its strength lies in interpretability and in helping the analyst decide which forecasting method is reasonable. It is also useful for communicating forecasts to managers who need to understand why demand is expected to change.

ARIMA Concepts: Autocorrelation and Differencing

Exponential-smoothing models focus primarily on evolving level, trend, and seasonality. ARIMA models provide another perspective by modelling dependence among successive observations and the effect of shocks.

ARIMA stands for AutoRegressive Integrated Moving Average. The three elements are represented as \((p,d,q)\):

• \(p\), the autoregressive order, indicates how many prior observations or transformed observations influence the current value;
• \(d\), the degree of differencing, indicates how many times the series is differenced to achieve stationarity; and
• \(q\), the moving-average order, indicates how many previous forecast errors influence the current value.

A stationary series has statistical properties such as mean and variance that are reasonably stable over time. A trending series may be made more stationary through first differencing:

\[\Delta Y_t=Y_t-Y_{t-1}\]

Seasonal differencing may compare an observation with the one from the previous season. Autocorrelation measures the relationship between observations separated by a particular lag. Autocorrelation and partial autocorrelation plots can assist in identifying plausible model orders, although automated selection should not replace judgement.

A seasonal ARIMA model adds seasonal autoregressive, differencing, and moving-average terms. ARIMA models can be powerful when residual dependence remains after accounting for trend and seasonality. They require careful diagnostics, adequate historical data, and attention to stationarity and parameter estimation. Their statistical sophistication does not protect them from poor data, structural breaks, or an inappropriate business specification.

Selecting and Validating a Forecasting Model

Model selection should be based on out-of-sample performance and operational usefulness. Randomly splitting time-series observations into training and test sets is usually inappropriate because it allows future information to influence model development. A better approach is rolling-origin or time-series cross-validation: train on an initial period, forecast the next period, expand the training window, and repeat.

Common error measures include:

\[MAE=\frac{1}{n}\sum |Y_t-\hat{Y}_t|\]

\[RMSE=\sqrt{\frac{1}{n}\sum(Y_t-\hat{Y}_t)^2}\]

\[MAPE=\frac{100}{n}\sum\left|\frac{Y_t-\hat{Y}_t}{Y_t}\right|\]

MAE is expressed in the original units and is easy to interpret. RMSE penalizes large errors more heavily and is useful when major misses are costly. MAPE is intuitive as a percentage but becomes unstable or undefined when actual demand is zero or close to zero. Scaled measures such as MASE, or weighted measures based on business cost, may be more appropriate for intermittent demand or portfolios with different volumes.

Forecast accuracy should not be the sole selection criterion. A model must also be timely, explainable, stable, computationally feasible, and compatible with planning systems. Bias is particularly important: repeated under-forecasting may create stockouts, while repeated over-forecasting increases inventory and obsolescence. Forecast error should be examined by product, location, customer, horizon, and season rather than only as one aggregate number.

Prediction intervals are as important as point forecasts. A point forecast of 1,000 units does not reveal whether plausible demand lies between 950 and 1,050 or between 600 and 1,400. Inventory policies and capacity decisions should reflect forecast uncertainty, lead time, service objectives, and the cost of errors.

Applied Case: Forecasting Regional Distribution-Centre Demand

Background

The following is an illustrative case. A distributor of packaged food supplies three regional distribution centres. One product family shows monthly demand with a recurring increase during the final quarter, while total demand has also grown gradually over two years.

Situation and Problem

The planning team currently uses the previous month’s demand as the forecast. This method is easy to operate but produces under-forecasts before the seasonal peak and excessive forecasts immediately after the peak. The resulting decisions include emergency replenishment, costly expedited transport, and surplus stock in January.

Analysis

The analyst first verifies that the data represent customer orders rather than shipments constrained by stock availability. Missing months are investigated, and an unusually large promotion is flagged. A time plot shows upward movement and annual seasonality whose amplitude increases with the level of demand. The seasonal naïve forecast improves substantially over the ordinary naïve benchmark, indicating that seasonality is real. However, it does not capture the growth trend.

The team compares a three-month moving average, simple exponential smoothing, Holt’s method, additive Holt–Winters, multiplicative Holt–Winters, and a seasonal ARIMA model. Rolling-origin validation is used, with the final six months reserved for evaluation. MAPE is reported cautiously because some smaller regional series contain low-volume months; MAE and weighted service-impact measures are also considered.

Approach and Solution

The multiplicative Holt–Winters model performs best for the main distribution centre because seasonal variation rises with the overall level. For a smaller centre with frequent zero observations, the model is unstable, and an intermittent-demand procedure combined with planner review is preferred. The promotional month is not treated as a normal seasonal observation; an event variable is used, and the forecast is prepared both with and without a planned promotion.

Forecasts are generated at the monthly level and reconciled with the annual operating plan. Planners receive a point forecast, prediction interval, assumptions concerning promotions, and an exception flag when the forecast differs materially from the previous cycle. Inventory parameters are then reviewed rather than being changed automatically from the forecast alone.

Outcome

The illustrative evaluation indicates lower error and less systematic bias than the previous-month method. More importantly, the planning team gains earlier visibility of the seasonal peak and can reserve transport capacity and supplier production slots. The result is not a claim that one model will remain best permanently; monthly monitoring is established so that model performance is reassessed when customer behaviour or supply conditions change.

Learning Points

The case demonstrates that model choice depends on data characteristics, not on statistical sophistication alone. Seasonal benchmarks should always be included, exceptional events should be identified explicitly, validation must respect time order, and forecasts should be integrated with inventory, capacity, and managerial decisions.

Professional Use, Technology, and Governance

Modern supply-chain planning platforms can automate data extraction, model fitting, forecast generation, exception reporting, and dashboard presentation. Machine-learning systems may augment classical time-series models with promotions, prices, weather, holidays, web traffic, and macroeconomic variables. Nevertheless, external variables should be included only when they are available at the forecast date and have a defensible causal or predictive relationship.

Automation creates governance responsibilities. Organisations should record model versions, parameters, training periods, overrides, data transformations, and forecast changes. Planners should be able to explain material decisions and challenge implausible outputs. Access controls are necessary where forecasts reveal commercially sensitive information. Bias may arise when historical demand reflects unequal service, discriminatory allocation, or systematic stockouts. A forecast should not reinforce such distortions without review.

Forecast overrides can add valuable market intelligence, but uncontrolled overrides can introduce optimism, political pressure, or double counting of information already in the model. A sound process records the reason, owner, expected effect, and subsequent accuracy of each override. Human judgement should complement, not conceal, model weakness.

Limitations and Common Errors

Frequent errors include choosing a model before plotting the data, ignoring stockouts, treating promotions as normal demand, using random cross-validation, selecting parameters on the test set, interpreting correlation as causation, and reporting only an aggregate accuracy score. Analysts also sometimes assume that a more complex model must be better, or extrapolate a short-term trend far into the future.

Time-series models are fundamentally dependent on historical patterns continuing sufficiently into the future. Structural breaks—such as a new competitor, regulatory change, product redesign, pandemic disruption, or major network change—can invalidate historical relationships. Intermittent and lumpy demand may require specialized approaches rather than ordinary smoothing. Forecasts should therefore be combined with scenario analysis, market intelligence, and explicit review of assumptions.

Best-Practice Workflow

A disciplined forecasting workflow is as follows:

1. Define the decision, forecast horizon, aggregation level, and required service outcome.

2. Assemble and cleanse demand data, documenting stockouts, returns, promotions, and missing values.

3. Plot the series and inspect level, trend, seasonality, outliers, and autocorrelation.

4. Establish naïve and seasonal naïve benchmarks.

5. Fit a small set of models appropriate to the observed structure.

6. Validate through rolling-origin forecasts and use error measures aligned with business cost.

7. Examine residuals and forecast bias, not only average accuracy.

8. Produce intervals, assumptions, and exceptions alongside point forecasts.

9. Review the forecast with relevant planners and record justified overrides.

10. Monitor performance and retrain or redesign the model when conditions change.

Current Developments

Forecasting practice is increasingly moving toward probabilistic forecasts, hierarchical reconciliation, and integrated demand-and-supply planning. Probabilistic forecasts support inventory and capacity decisions by representing uncertainty rather than presenting a single number. Hierarchical methods seek coherence between forecasts for products, regions, channels, and total company demand. Ensembles combine forecasts from several models and may improve robustness when no single model is consistently dominant.

Artificial intelligence and machine learning can support feature-rich forecasting, but classical models remain valuable because they are transparent, fast, and often competitive for regular demand. The emerging professional standard is therefore not to replace statistical forecasting with AI automatically, but to compare methods rigorously, govern their use, and match model complexity to the decision problem.

Chapter Summary

Time-series forecasting estimates future demand from chronologically ordered observations. Naïve and seasonal naïve methods provide essential benchmarks. Moving averages smooth data but may lag, while simple exponential smoothing updates a stable level. Holt’s method adds trend, and Holt–Winters models add seasonality through additive or multiplicative structures. Decomposition helps interpret level, trend, seasonal, and irregular components. ARIMA models represent autoregression, differencing, and moving-average error behaviour and are useful when autocorrelation and stationarity require explicit treatment.

Reliable model selection requires time-ordered validation, suitable error measures, residual and bias analysis, and attention to forecast intervals. Data quality, stockouts, promotions, structural breaks, and intermittent demand can matter more than algorithmic complexity. In professional practice, forecasts must be connected to inventory, transport, capacity, technology, governance, and managerial judgement.

Key Terms

Time series: Chronologically ordered observations of a variable.

Level: The underlying average magnitude of a time series.

Trend: A persistent long-term upward or downward movement.

Seasonality: A repeating pattern associated with a fixed calendar or operating cycle.

Cyclical movement: A relatively long and less regular fluctuation, often related to economic conditions.

Irregular component: Unsystematic variation or random disturbance in a series.

Naïve forecast: A forecast equal to the most recent observed value.

Seasonal naïve forecast: A forecast using the observation from the corresponding prior season.

Moving average: An average of observations within a rolling time window.

Exponential smoothing: A method that updates forecasts by weighting recent observations and prior forecasts.

Holt’s method: Exponential smoothing that estimates both level and trend.

Holt–Winters method: Exponential smoothing that estimates level, trend, and seasonality.

Decomposition: Separation of a time series into interpretable components.

Stationarity: Stability of key statistical properties of a series over time.

Differencing: Subtracting an earlier observation from a later one to remove changing level or trend.

Autocorrelation: Correlation between observations separated by specified time lags.

ARIMA: A model combining autoregression, differencing, and moving-average error terms.

MAE: The average absolute difference between actual and forecast values.

RMSE: The square root of average squared forecast errors.

Prediction interval: A range intended to contain a future observation with stated probability.

Review Questions

Short Answer Questions

1. What is the difference between a naïve forecast and a seasonal naïve forecast?

2. How does the smoothing parameter affect simple exponential smoothing?

3. Distinguish additive from multiplicative seasonality.

4. What does differencing accomplish in an ARIMA model?

5. Why is random train-test splitting inappropriate for most time-series data?

Long Answer Questions

1. Explain the progression from moving averages to simple exponential smoothing and discuss the circumstances in which each is appropriate.

2. Describe Holt’s and Holt–Winters methods, including their components, assumptions, and limitations.

3. Explain how decomposition can support both diagnosis and forecasting of logistics demand.

4. Discuss the meaning of the three ARIMA orders, \(p\), \(d\), and \(q\), and explain the role of stationarity.

5. Develop a complete model-selection and validation procedure for a seasonal distribution-centre demand series.

Analytical / Application Questions

1. Weekly demand for the last three weeks was 480, 510, and 450 units. Calculate the three-week moving-average forecast. Explain how the forecast would change if a one-week window were used.

2. A product has rising monthly demand and strong peaks every December. Compare the likely performance of simple exponential smoothing, Holt’s method, seasonal naïve forecasting, and multiplicative Holt–Winters forecasting.

3. A forecasting system reports a low MAPE but repeatedly under-forecasts during stockout periods. Diagnose the problem and recommend data, modelling, validation, and governance actions.

22. Multiple Choice Questions with Answers

Question 1

A company’s monthly demand shows a stable annual pattern whose absolute seasonal increase remains approximately constant as volume changes. Which model structure is most appropriate?

A. Additive seasonality

B. Multiplicative seasonality

C. First-order autoregression only

D. Simple naïve forecasting

Correct Answer: A. Additive seasonality

Question 2

The last three weekly demands are 480, 510, and 450 units. What is the three-week moving-average forecast?

A. 450 units

B. 470 units

C. 480 units

D. 510 units

Correct Answer: C. 480 units

Question 3

A high value of alpha in simple exponential smoothing causes the forecast to:

A. Ignore the latest observation

B. Respond more quickly to recent changes

C. Remove all seasonal variation

D. Estimate a longer seasonal cycle

Correct Answer: B. Respond more quickly to recent changes

Question 4

Which method is specifically designed to estimate level, trend, and seasonality together?

A. Simple exponential smoothing

B. Ordinary naïve forecasting

C. Holt–Winters exponential smoothing

D. A simple moving average

Correct Answer: C. Holt–Winters exponential smoothing

Question 5

For monthly demand with annual seasonality, the seasonal naïve forecast for next December should normally use:

A. The previous November demand

B. The average of all months

C. The previous December demand

D. The most recent week’s demand

Correct Answer: C. The previous December demand

Question 6

In an ARIMA(p,d,q) model, the parameter d represents:

A. The number of seasonal cycles

B. The degree of differencing

C. The number of forecast horizons

D. The number of external variables

Correct Answer: B. The degree of differencing

Question 7

Why is rolling-origin validation preferable to random splitting for time-series forecasting?

A. It guarantees zero forecast error

B. It preserves the chronological forecasting process

C. It eliminates the need for test data

D. It always favours complex models

Correct Answer: B. It preserves the chronological forecasting process

Question 8

Which error measure becomes problematic when actual demand is zero or close to zero?

A. MAE

B. RMSE

C. MAPE

D. Bias

Correct Answer: C. MAPE

Question 9

A series has a strong upward trend but no stable seasonal pattern. Which method is generally more suitable than simple exponential smoothing?

A. Holt’s method

B. Seasonal naïve forecasting only

C. A zero-demand model

D. A fixed seasonal index without trend

Correct Answer: A. Holt’s method

Question 10

A planner observes low forecast error overall but systematic under-forecasting during stockouts. What is the most appropriate first response?

A. Immediately increase model complexity

B. Ignore the issue because aggregate error is low

C. Investigate whether recorded sales understate true demand

D. Delete all observations from the database

Correct Answer: C. Investigate whether recorded sales understate true demand