Time-Series Structure and Classical Models

Introduction

Demand in logistics is observed over time: orders arrive daily, weekly, monthly, or annually, while inventory, transport capacity, labour, and procurement decisions must be made before future demand is known. Time-series analysis provides a disciplined way to extract predictable structure from historical observations and to generate forecasts.

A time series is an ordered sequence of observations indexed by time. Let y_t denote demand observed at period t, where the measurement unit may be pallets, tonnes, orders, or vehicle loads. Unlike cross-sectional data, time-series observations are often dependent: demand this week may be related to demand last week, and demand in December may resemble demand in previous Decembers. This dependence is both a source of forecasting information and a source of statistical difficulty.

Classical forecasting begins by identifying the structure of a series and then selecting a model whose assumptions match that structure. The principal elements are level, trend, seasonality, cycles, irregular variation, autocorrelation, and stationarity. The methods covered in this chapter range from simple benchmarks to exponential smoothing, decomposition, and the ARIMA family.

Learning Objectives

By the end of this chapter, a learner should be able to:

1. Distinguish level, trend, seasonality, cycles, and irregular variation.
2. Explain autocorrelation and stationarity and diagnose them using plots and statistics.
3. Apply naïve, seasonal naïve, moving-average, and exponential-smoothing methods.
4. Explain and apply Holt, Holt–Winters, and classical decomposition models.
5. Describe AR, MA, ARMA, ARIMA, and seasonal ARIMA models.
6. Select models using data characteristics, validation, and forecast accuracy.
7. Interpret forecasting results for logistics decisions and recognise model limitations.

1. The Structure of a Demand Series

A useful conceptual representation is

where T_t is trend, S_t is seasonal variation, C_t is cyclical variation, and I_t is irregular or random variation. This is an additive representation, appropriate when seasonal fluctuations are approximately constant in absolute size.

An alternative is the multiplicative representation:

It is appropriate when seasonal variation grows in proportion to the level. For example, a warehouse may experience a seasonal increase of approximately 20% of normal demand rather than a fixed 2,000 units. Taking logarithms converts the multiplicative relationship into an additive one:

The logarithmic transformation requires positive observations and changes the scale of interpretation. Zero or negative observations require another transformation or a different modelling strategy.

Level

Level is the underlying magnitude around which observations fluctuate. A stable product with demand near 500 cartons per week has a level of approximately 500. Level is not necessarily the first observation, nor is it simply the arithmetic mean when trend or seasonality exists.

Trend

Trend is a persistent long-term increase or decrease. It may arise from market growth, product adoption, network expansion, inflation, or declining demand caused by substitution. A trend need not be linear. A linear trend can be written as

where a is the intercept, b is the change per period, and e_t is the residual. The equation is descriptive; it does not guarantee that growth will continue.

Seasonality

Seasonality is a systematic pattern that repeats at a fixed and known frequency. Weekly demand may vary by day, monthly demand may vary by month, and retail logistics often show annual peaks. If observations are monthly, an annual seasonal period is m = 12; for daily data with a weekly pattern, m = 7, subject to operating calendars and holiday effects.

Seasonality must be distinguished from calendar effects. A public holiday may shift demand but does not necessarily create a repeating pattern of equal length. Promotions and one-off disruptions are interventions or events, not seasonality.

Cycles

Cycles are medium- or long-term movements without a fixed, known period. Economic expansions and recessions can alter freight volumes, but cycle length varies. Cycles are harder to forecast than seasonality because their timing is uncertain.

Irregular variation and outliers

Irregular variation includes random noise, measurement error, strikes, weather shocks, port closures, and exceptional promotions. An outlier is an observation unusually distant from the expected pattern. It should not be deleted automatically: a genuine disruption may be important for resilience planning, while a data-entry error should be corrected.

2. Autocorrelation and Stationarity

Autocorrelation

Autocorrelation measures association between observations separated by a lag. The lag-k autocorrelation is commonly estimated as

where n is the number of observations and ȳ is the sample mean. The numerator compares each observation with the observation k periods earlier. The statistic is dimensionless.

Positive autocorrelation means high observations tend to follow high observations and low observations tend to follow low observations. It indicates persistence. Negative autocorrelation indicates alternating behaviour. Strong autocorrelation at lag 7 in daily data may indicate weekly seasonality.

The autocorrelation function (ACF) displays correlations across multiple lags. The partial autocorrelation function (PACF) measures the association at a lag after accounting for shorter lags. ACF and PACF are useful diagnostics, but they do not replace substantive knowledge or out-of-sample validation.

Stationarity

A weakly stationary series has a constant mean, constant finite variance, and autocovariance that depends on the lag rather than the calendar time. In practical terms, the statistical behaviour of the series is stable over time. Many ARIMA identification and inference procedures assume stationarity after appropriate differencing.

A series with trend is generally non-stationary because its mean changes over time. Seasonal means also violate ordinary stationarity unless seasonality is modelled or removed. Differencing can remove stochastic trend:

Seasonal differencing with period m is

Second-order differencing is Δ²y_t = Δy_t − Δy_t−1. Differencing should be used sparingly: excessive differencing can create unnecessary noise and induce negative autocorrelation.

Stationarity can be assessed through time plots, rolling means and variances, ACF behaviour, and formal tests such as the augmented Dickey–Fuller test. Formal tests have assumptions and can have low power in short samples; visual and operational evidence remain essential.

3. Forecasting Benchmarks

A benchmark establishes a minimum standard against which complex models should be compared.

Naïve method

The one-step naïve forecast is

The most recent observation is used as the forecast. It is difficult to improve upon when a series follows a random walk and is highly useful as a control method.

Seasonal naïve method

For a seasonal period m, the seasonal naïve forecast is

when the required historical observation exists. Thus, next Monday's forecast may equal demand on the previous Monday. This method is often a strong benchmark for seasonal logistics demand.

Drift method

The drift method extrapolates the average historical change:

where h is the forecast horizon and y₁ is the first observation. It assumes the average change over the sample continues.

4. Moving Averages

A simple moving-average forecast with window length k is

Each of the latest k observations receives equal weight. A larger k produces more smoothing but responds more slowly to genuine changes in level. A smaller window reacts faster but retains more noise.

For example, if weekly demand in the latest three weeks is 120, 135, and 150 cartons, the three-week moving-average forecast is

The method is simple, transparent, and useful for stable demand. It is less suitable when strong trend or seasonality exists unless seasonal adjustments are made.

A weighted moving average assigns weights w_i to observations, with

The most recent observation may receive the greatest weight. Weight selection should be based on validation rather than convenience.

5. Exponential Smoothing

Exponential smoothing assigns declining weights to older observations. It is computationally efficient and adapts to changing demand.

Simple exponential smoothing

For a series without systematic trend or seasonality, the level update is

and the forecast for any future horizon is

where 0 < α < 1, ℓ_t is the estimated level, and h ≥ 1. A high α reacts strongly to the latest observation; a low value produces greater smoothing. The initial level ℓ₀ must be specified, often using the first observation or an initial average.

Holt’s linear method

Holt’s method extends exponential smoothing to trend:

Here b_t is the estimated trend, and 0 < β < 1. The additive trend assumes a roughly constant change per period. Damped-trend methods reduce long-horizon extrapolation by gradually weakening the trend, which is often more realistic for mature logistics networks.

Holt–Winters seasonal method

Holt–Winters adds seasonality. In additive form, with seasonal period m:

The parameters α, β, and γ lie between zero and one. The additive model is appropriate when seasonal effects have constant magnitude. In multiplicative Holt–Winters, seasonal indices multiply the level and are therefore appropriate for proportional seasonality; positive data are generally required.

Parameters may be estimated by minimizing in-sample squared errors or by maximising a likelihood under an error model. Optimisation does not guarantee good forecasts if the model structure is inappropriate.

6. Decomposition

Decomposition separates a time series into interpretable components. Classical decomposition commonly estimates trend-cycle using a moving average, calculates seasonal indices from detrended observations, and treats the remainder as irregular variation.

For additive decomposition:

where R_t is the remainder. Seasonal indices in an additive model commonly sum to approximately zero over a complete seasonal cycle. For multiplicative decomposition:

seasonal indices commonly average approximately one.

Decomposition is valuable for explanation, capacity planning, and diagnosing unusual periods. However, classical decomposition can be sensitive to end effects, assumes stable seasonal patterns, and is not automatically the best forecasting model. Modern decomposition procedures may use robust smoothers and transformations to reduce outlier influence.

7. ARIMA-Family Models

Autoregressive models

An AR(p) model expresses the current value as a function of previous values:

where c is a constant, φ_i are autoregressive coefficients, and ε_t is an error term with mean zero. Stationarity imposes restrictions on the coefficients; these restrictions are more complex than simply requiring every coefficient to be below one in absolute value.

Moving-average error models

An MA(q) model uses current and previous shocks:

where μ is the mean and θ_j are moving-average coefficients. The term “moving average” here refers to past errors, not the smoothing method described earlier.

ARMA and ARIMA

ARMA combines AR and MA terms for stationary data. ARIMA(p,d,q) applies d ordinary differences before fitting an ARMA(p,q) model. In operator notation, with backshift operator By_t = y_t−1, a non-seasonal ARIMA model can be expressed as

The notation is compact: φ(B) represents the AR polynomial and θ(B) represents the MA polynomial. The errors are commonly assumed to be uncorrelated with constant variance, although extensions allow non-normal or changing-variance errors.

Seasonal ARIMA, written SARIMA(p,d,q(P,D,Q_m)), adds seasonal AR, differencing, and MA terms at period m. It is useful for monthly demand with annual seasonality or daily demand with weekly seasonality.

Identification and diagnostics

A disciplined ARIMA workflow is:

1. Plot the series and investigate missing values, outliers, calendar effects, and structural breaks.
2. Transform or difference only when needed to achieve approximate stationarity.
3. Use ACF and PACF patterns to propose candidate orders.
4. Estimate several plausible models.
5. Check residuals. Residual autocorrelation indicates that predictable structure remains.
6. Compare out-of-sample forecasts using rolling-origin validation.
7. Refit and monitor the selected model after deployment.

Information criteria such as AIC or BIC balance fit and complexity, but they are not substitutes for operational validation. A residual portmanteau test, such as the Ljung–Box test, can assess whether residual autocorrelations are jointly inconsistent with white noise. A non-significant result does not prove that the model is correct; it only indicates insufficient evidence of remaining autocorrelation under the test assumptions.

8. Forecast Evaluation and Model Selection

Let actual demand be y_t and forecast be ŷ_t. Common measures include:

MAE is expressed in demand units and is easy to interpret. RMSE penalises large errors more heavily. MAPE is scale-free but is undefined when actual demand is zero and unstable when actual demand is close to zero. Weighted absolute percentage error or scaled measures may be preferable for intermittent demand.

Forecast evaluation must respect time order. Randomly splitting observations can allow future information to enter the training set. Rolling-origin evaluation repeatedly trains on an earlier window and tests on the next period or horizon. The evaluation horizon should match the decision: transport scheduling may require a seven-day horizon, whereas procurement may require several months.

Accuracy is not the only criterion. A model should also be explainable, stable, computationally feasible, responsive to structural change, and compatible with inventory or capacity systems. Forecast bias is particularly important: persistent underforecasting creates shortages, while persistent overforecasting creates excess stock and avoidable handling.

9. Illustrative Case: Regional E-Commerce Fulfilment

Background

A regional fulfilment centre serves a network of retail outlets. Daily carton demand has been recorded for three years. The operation experiences a strong weekly pattern, higher volumes before major holidays, and a gradual upward trend caused by customer growth.

Problem

The centre originally used a seven-day moving average. Managers observed repeated underforecasting on Mondays and overforecasting on low-volume weekends. The resulting labour plan caused Monday overtime and unnecessary weekend capacity.

Analysis

A time plot showed trend and weekly seasonality. The ACF had notable peaks at lags 7, 14, and 21, indicating weekly dependence. A seasonal naïve benchmark was established by forecasting each day using demand on the same weekday in the previous week. A Holt–Winters model with weekly seasonality was then estimated. Holiday indicators were considered separately because holiday peaks were not regular weekly seasonality.

The data were divided chronologically into training and validation periods. Models were evaluated with MAE and bias over a four-week rolling-origin procedure. A model using a multiplicative seasonal form performed better during high-volume weeks because seasonal variation increased with the level. However, extreme promotion days produced large errors, so the organisation recorded promotions as explanatory events rather than allowing them to distort the baseline seasonal pattern.

Solution and outcome

The fulfilment centre adopted the seasonal benchmark for low-volume products and Holt–Winters for stable, high-volume products. Promotion-adjusted forecasts were used for event periods. Forecasts were reconciled with labour and vehicle-capacity constraints, and errors were monitored weekly. Monday staffing improved because the model represented weekday structure explicitly. The organisation did not claim that the model predicted exceptional disruptions; instead, managers used scenario adjustments for strikes, weather, and major campaigns.

Learning points

The case illustrates that a sophisticated model is not automatically superior. A seasonal naïve method can be highly competitive, while Holt–Winters is useful when level, trend, and seasonality evolve gradually. Data quality, event information, validation design, and operational integration were as important as model estimation.

10. Limitations, Governance, and Contemporary Practice

Classical models assume that historical structure contains useful information about the future. Structural breaks, new products, network redesign, geopolitical disruption, and changes in customer behaviour can invalidate that assumption. Intermittent demand may require Croston-type methods or probabilistic inventory models rather than ordinary smoothing. Hierarchical supply chains may require forecast reconciliation so that item, facility, region, and national forecasts add consistently.

Forecasts should be accompanied by prediction intervals, not only point estimates. An interval represents uncertainty under model and data assumptions; it is not a guarantee. Wider intervals are generally appropriate for longer horizons and unstable series.

Automation and AI can support anomaly detection, feature construction, model selection, and forecast monitoring. Nevertheless, automated systems require governance. Analysts should document data sources, transformations, parameter changes, overrides, and model performance. Forecasts may influence allocation of scarce inventory among customers, so organisations should examine whether systematic data gaps or service priorities create unfair outcomes. Human overrides should be logged and evaluated rather than treated as unquestionable expertise.

Best practice is to maintain a portfolio of models, compare against simple benchmarks, use rolling validation, monitor residuals and forecast bias, incorporate known events, and retrain when performance or demand structure changes. The objective is not to identify a universally best algorithm; it is to produce reliable decision support under clearly understood assumptions.

Chapter Summary

Time-series forecasting begins by identifying level, trend, seasonality, cycles, and irregular variation. Additive models assume constant seasonal magnitude, whereas multiplicative models assume proportional seasonal effects. Autocorrelation reveals dependence across lags, and stationarity describes stable statistical behaviour over time. Differencing can remove trend or seasonal persistence but should not be excessive.

Naïve and seasonal naïve methods provide essential benchmarks. Moving averages smooth observations but may respond slowly to changes. Simple exponential smoothing estimates level; Holt’s method adds trend; Holt–Winters adds seasonality. Classical decomposition separates components for interpretation and planning. AR, MA, ARMA, ARIMA, and SARIMA model dependence and differencing more explicitly.

Model choice should be based on data structure, operational horizon, residual diagnostics, rolling-origin validation, accuracy, interpretability, and stability. Logistics professionals must also address outliers, promotions, structural breaks, intermittent demand, uncertainty intervals, governance, and forecast bias. A sound forecasting process combines statistical models with domain knowledge and continuous monitoring.

Key Terms

Level: The underlying magnitude around which a series varies.

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

Seasonality: A pattern repeating at a fixed, known frequency.

Cycle: A longer movement with variable duration and timing.

Irregular variation: Unpredictable residual movement or noise.

Autocorrelation: Correlation between observations separated by a time lag.

ACF: The autocorrelation function across multiple lags.

PACF: Partial autocorrelation after accounting for shorter lags.

Stationarity: Stable mean, variance, and lag-based dependence over time.

Differencing: Subtracting a previous observation to remove persistence or trend.

Naïve forecast: A forecast equal to the latest observed value.

Seasonal naïve forecast: A forecast equal to the observation from the corresponding previous season.

Moving average: An average of a fixed window of recent observations.

Exponential smoothing: A method assigning declining weights to older observations.

Holt’s method: Exponential smoothing with a trend component.

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

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

Residual: The difference between an observed value and its fitted or forecast value.

Rolling-origin validation: Time-ordered repeated evaluation using successive forecast origins.

Forecast bias: Systematic tendency to overforecast or underforecast.

Review Questions

Short Answer Questions

1. What distinguishes seasonality from a cycle?
2. Define weak stationarity.
3. Why is a seasonal naïve forecast a useful benchmark?
4. What does the smoothing parameter α control in simple exponential smoothing?
5. Why can MAPE be problematic when actual demand is zero?

Long Answer Questions

1. Explain the additive and multiplicative decomposition models and state when each is appropriate.
2. Describe how Holt–Winters extends simple exponential smoothing and Holt’s method.
3. Explain the roles of ACF, PACF, differencing, and residual diagnostics in ARIMA modelling.
4. Compare moving averages, exponential smoothing, and ARIMA models for logistics demand.
5. Discuss how forecast validation should be designed for a warehouse with weekly and annual planning horizons.

Analytical / Application Questions

1. A product has weekly demand of 120, 135, and 150 cartons in its latest three weeks. Calculate the three-week moving-average forecast and explain one limitation of this forecast.
2. Daily demand has strong ACF peaks at lags 7 and 14, but no visible long-term trend. Which structural feature is most likely present, and which benchmark should be tested first?
3. A forecasting model has low RMSE but persistent underforecasting during promotions. Recommend a modelling and governance response, distinguishing baseline seasonality from event effects.

Multiple Choice Questions with Answers

Question 1

Which feature is defined as a pattern that repeats at a fixed and known frequency?

A. Cycle

B. Seasonality

C. Irregular variation

D. Structural break

Correct Answer: B. Seasonality

Question 2

If the latest three weekly demands are 120, 135, and 150 cartons, what is the three-week moving-average forecast?

A. 125 cartons

B. 130 cartons

C. 135 cartons

D. 140 cartons

Correct Answer: C. 135 cartons

Question 3

The seasonal naïve forecast for a monthly series with annual seasonality normally uses which observation?

A. The immediately preceding month only

B. The average of all historical months

C. The observation from the same month in the previous year

D. The first observation in the series

Correct Answer: C. The observation from the same month in the previous year

Question 4

What does a high value of α in simple exponential smoothing generally imply?

A. The forecast reacts more strongly to recent observations

B. The forecast ignores the latest observation

C. The seasonal period becomes longer

D. The forecast becomes automatically stationary

Correct Answer: A. The forecast reacts more strongly to recent observations

Question 5

Which transformation is commonly used to remove a stochastic trend?

A. Seasonal indexing only

B. Ordinary differencing

C. Increasing the forecast horizon

D. Randomly shuffling observations

Correct Answer: B. Ordinary differencing

Question 6

Which model explicitly includes level, trend, and seasonality?

A. Simple exponential smoothing

B. Holt’s linear method

C. Holt–Winters method

D. A single moving average

Correct Answer: C. Holt–Winters method

Question 7

In an AR model, current values are modelled primarily as a function of what?

A. Previous values of the series

B. Only future observations

C. Seasonal indices alone

D. The forecast error percentage

Correct Answer: A. Previous values of the series

Question 8

Why is random train-test splitting inappropriate for many time-series forecasting tasks?

A. It always increases sample size

B. It can allow future information to enter model training

C. It prevents calculation of MAE

D. It removes all seasonality

Correct Answer: B. It can allow future information to enter model training

Question 9

Which accuracy measure is expressed in the original demand units?

A. MAE

B. MAPE

C. ACF

D. BIC

Correct Answer: A. MAE

Question 10

Persistent underforecasting during promotions is best addressed initially by which action?

A. Delete all promotion observations

B. Treat promotions as possible event variables and evaluate them separately

C. Randomly reorder the data

D. Use a larger error metric without changing the model

Correct Answer: B. Treat promotions as possible event variables and evaluate them separately