Time-Series Forecasting

Introduction and context

Time-series forecasting estimates future observations from data recorded sequentially over time. In marketing management, the observations may be weekly sales, monthly demand, daily website visits, quarterly subscription cancellations, or hourly transactions. Unlike cross-sectional prediction, time-series forecasting must respect temporal order: information available at time t may be used to forecast the future, but future information must not be used to explain the past.

Forecasts support decisions about inventory, staffing, advertising budgets, capacity, procurement, cash flow, and pricing. A demand forecast is not a prediction of an inevitable future. It is a conditional estimate based on historical patterns, assumptions, and an information set. A good forecast is therefore not merely statistically accurate; it must also be timely, interpretable, operationally feasible, and aligned with the decision being made.

A useful representation is:

where y_t is the observed value at time t, T_t is trend, S_t is seasonality, C_t is a cycle, and I_t is an irregular or random component. This additive representation is appropriate when seasonal fluctuations are approximately constant in absolute size. When fluctuations increase with the level of the series, a multiplicative representation is often more suitable:

The decomposition is conceptual rather than a guarantee that each component can be perfectly isolated.

Learning objectives

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

1. Distinguish trend, seasonality, cycles, irregular variation, stationarity, and autocorrelation.
2. Apply naïve, seasonal naïve, moving-average, weighted moving-average, and exponential-smoothing methods.
3. Explain and apply Holt, Holt–Winters, decomposition, and ARIMA models.
4. Diagnose time-series structure using plots, autocorrelation, residual analysis, and forecast-error measures.
5. Select a forecasting method appropriate to a business context, forecast horizon, data frequency, and decision cost.
6. Recognize limitations involving structural breaks, causal changes, data leakage, uncertainty, and ethical use of forecasts.

1. The structure of a time series

Trend

Trend is the persistent long-run direction of a series. A trend may be linear, nonlinear, increasing, decreasing, or interrupted. For example, a subscription service may experience sustained growth as market penetration increases. Trend should not be confused with a short run of increases; a formal assessment requires a sufficiently long time series and consideration of other components.

Seasonality

Seasonality is a systematic pattern that repeats at a known and fixed calendar frequency. Examples include higher retail demand in December, weekend differences in restaurant traffic, and monthly differences in utility use. If monthly data have an annual seasonal cycle, the seasonal period is m = 12; for daily data with a weekly pattern, m = 7, although multiple seasonalities may coexist.

Cycles

Cycles are longer, often irregular fluctuations associated with economic conditions, industry dynamics, or product life cycles. Unlike seasonality, cycles do not necessarily have a fixed length or stable phase. A recession-related decline in demand is cyclical, whereas a December sales peak is seasonal.

Irregular variation and shocks

Irregular variation includes unpredictable events, measurement error, and temporary shocks. A product recall, viral campaign, competitor failure, or supply disruption may produce an outlier. An outlier should not automatically be removed: it may represent either an error or an economically meaningful event.

Stationarity

A weakly stationary series has a constant mean, constant variance, and autocovariance that depends on lag rather than calendar time. In practice, many business series are not stationary because they contain trend, seasonality, changing variance, or structural breaks. ARIMA models generally require a stationary series after suitable differencing and seasonal adjustment.

Autocorrelation

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

Here, n is the number of observations, ȳ is the sample mean, and k is the lag. The statistic is dimensionless. Positive autocorrelation means high values tend to follow high values and low values tend to follow low values. It is evidence that observations are not independent, and it can make ordinary random-sampling assumptions inappropriate.

A time plot should normally be examined before any model is fitted. Autocorrelation and partial autocorrelation plots are useful diagnostics, but they should be interpreted with subject-matter knowledge and not treated as automatic model selectors.

2. Baseline forecasting methods

Simple methods are essential because they are transparent, inexpensive, and provide benchmark forecasts. A sophisticated model that cannot outperform a suitable baseline has little practical justification.

Naïve forecasting

The naïve forecast for the next period equals the most recent observation:

where ŷ is the forecast and the subscript indicates a forecast made at time t for period t + 1. This method is appropriate for a series resembling a random walk and is a strong benchmark for short-horizon forecasting.

For a seasonal naïve method, the forecast equals the observation from the same season in the previous cycle:

where h is the forecast horizon and m is the seasonal period. This is useful when seasonal patterns repeat but trend is limited.

Drift method

A naïve forecast can be extended by estimating average change over the historical period. The drift forecast is:

The method assumes that the average historical slope is a reasonable estimate of future change. It is vulnerable to structural breaks and should not be extrapolated mechanically over very long horizons.

Simple moving average

A moving average of order k averages the most recent k observations:

The forecast for the next period is often set equal to M_t. All observations receive equal weight. A larger k produces greater smoothing but responds more slowly to changes; a smaller k responds faster but retains more noise.

For example, if the last three weekly demands are 120, 132, and 138 units, the three-period moving-average forecast is:

The calculation is simple, but the method does not explicitly model trend or seasonality.

Weighted moving average

A weighted moving average assigns larger weights to observations considered more informative:

with w_i ≥ 0 and ∑w_i = 1. The weights must sum to one so that the forecast retains the same units and is a convex combination of past observations. A business may choose weights in advance, or estimate them using a validation sample. Arbitrary weights can create the appearance of sophistication without improving accuracy.

3. Exponential smoothing

Exponential smoothing updates a forecast by combining the latest observation with the previous forecast. Simple exponential smoothing is suitable for level data without systematic trend or seasonality:

where ℓ_t is the estimated level and 0 ≤ α ≤ 1 is the smoothing parameter. The one-step forecast is:

A high α gives more weight to recent observations and reacts quickly; a low α produces smoother forecasts. Initial level selection matters, especially for short series. Parameters are commonly selected by minimizing a loss function on the training data, but validation data should be used to assess out-of-sample performance.

Holt’s linear trend method

Holt’s method extends exponential smoothing to a series with trend but no seasonality:

Here, b_t is the estimated trend, 0 ≤ α ≤ 1 is the level parameter, 0 ≤ β ≤ 1 is the trend parameter, and h is the horizon. Additive trend assumes that the absolute rate of change is reasonably stable. Damped-trend Holt forecasting replaces the indefinitely linear extrapolation with a trend that gradually flattens:

where 0 < ϕ < 1 is the damping parameter. Damping is often more realistic for mature markets.

Holt–Winters seasonal methods

Holt–Winters methods add seasonality. For additive seasonality with period m:

The parameters α, β, and γ lie between 0 and 1. Additive seasonality is appropriate when seasonal effects are approximately constant in units. For multiplicative seasonality, deviations are expressed as ratios rather than differences; this is suitable when seasonal amplitude grows with the level and all relevant values are positive. Log transformation can sometimes convert multiplicative behavior into approximately additive behavior, but back-transformation requires care because the mean of exponentiated forecasts is not generally the exponent of the mean log forecast.

4. Decomposition and diagnosis

Decomposition separates a series into interpretable components. Classical decomposition may use moving averages to estimate trend-cycle, then calculate seasonal indices. For an additive model, seasonal indices typically sum to approximately zero across one complete cycle. For a multiplicative model, indices typically average approximately one.

A practical decomposition procedure is:

1. Plot the raw series and identify possible trend, changing variance, outliers, and breaks.
2. Estimate the trend-cycle using a moving average or a model-based smoother.
3. Estimate seasonal effects from deviations from the trend-cycle.
4. Remove estimated seasonality and inspect the remainder.
5. Forecast the components or use them to guide a model choice.

Decomposition is useful for communication and diagnosis, but estimated components can be sensitive to endpoints, unusual observations, and changing seasonal behavior. Modern methods such as STL use locally estimated regression and can be more flexible, particularly when seasonality changes gradually. Decomposition is not causal analysis: it describes patterns without proving why they occur.

Residual diagnostics

Let the forecast error be:

A satisfactory residual series should have a mean near zero, no substantial remaining autocorrelation, approximately stable variance, and no systematic pattern against time or fitted values. Normality is not essential for point forecasts, but it matters for some interval procedures and statistical tests.

Common accuracy measures include:

MAE is in the units of demand and is easy to interpret. RMSE penalizes large errors more heavily. MAPE is scale-free but is undefined when actual demand is zero and unstable when actual values are close to zero. For intermittent or low-volume demand, MAE, RMSE, weighted absolute percentage error, or scaled measures are often preferable.

Forecast evaluation must preserve time order. A rolling-origin evaluation repeatedly trains on an earlier period and tests on a subsequent period. This better reflects deployment than random train-test splitting, which can leak future information into the training set.

5. ARIMA models

ARIMA models represent dependence in a stationary series and use differencing to address nonstationarity. An ARIMA(p,d,q) model contains:

• p: autoregressive order;
• d: number of nonseasonal differences;
• q: moving-average error order.

A first difference is:

A second difference is ∆²y_t = ∆y_t − ∆y_{t−1}. Differencing should be used only as needed; excessive differencing can increase noise and produce difficult-to-interpret forecasts.

An AR(p) model expresses the stationary series x_t as:

where c is a constant, ϕ coefficients measure lag effects, and ε_t is an innovation with mean zero and constant variance under the standard model assumptions. An MA(q) model expresses the series through current and previous shocks:

An ARIMA model applies AR and MA components to a differenced series. Seasonal ARIMA, written SARIMA(p,d,q)(P,D,Q)_m, adds seasonal autoregressive, differencing, and moving-average terms at period m. It is useful when autocorrelation remains at seasonal lags after ordinary differencing.

Model identification and estimation

A disciplined ARIMA workflow is:

1. Plot the series and identify transformations, breaks, trend, and seasonality.
2. Stabilize variance if necessary, for example with a logarithmic transformation for positive data.
3. Difference nonstationary components cautiously.
4. Examine ACF and PACF plots to propose candidate orders.
5. Estimate several plausible models using maximum likelihood or an equivalent appropriate method.
6. Compare models using out-of-sample accuracy and, as secondary evidence, information criteria such as AIC or BIC.
7. Check residuals for remaining autocorrelation and instability.
8. Produce point forecasts and prediction intervals.

AIC and BIC reward fit but penalize complexity. They are not substitutes for business-relevant validation. A model with the lowest information criterion may forecast poorly during a market transition.

6. Choosing a method in practice

Method selection depends on the data-generating pattern, forecast horizon, data volume, interpretability requirements, and cost of errors.

Forecasting should be separated from pricing causality. A time-series model can forecast demand under a continuation of historical pricing, but it cannot automatically estimate how demand will change after a price alteration. Price elasticity requires a causal design, such as experiments, quasi-experiments, or a carefully specified demand model.

Worked application: regional beverage demand

This illustrative case concerns a beverage distributor forecasting monthly demand for a mature product. The firm has 36 months of sales data, with a recurring summer peak and modest long-run growth. Management must set production for the next three months.

Diagnosis

A time plot shows increasing demand and a repeating annual pattern. Seasonal amplitude is approximately constant in cases rather than proportional to the level, suggesting additive seasonality. A seasonal naïve benchmark is established first. Because demand has trend and seasonality, simple exponential smoothing is unsuitable, while Holt–Winters additive and SARIMA models are plausible candidates.

Evaluation design

The final six months are held out as a test period. Models are fitted only to the first 30 months. Within the training period, rolling-origin validation compares the seasonal naïve, additive Holt–Winters, and SARIMA candidates using MAE. This prevents the test period from influencing model selection.

Suppose the resulting test errors are:

The additive Holt–Winters model is selected because it has the lowest test MAE and produces stable residuals. Its three-month forecasts are 1,240, 1,310, and 1,460 cases. The commercial team receives both point forecasts and prediction intervals, not only the central estimates. Production is planned near the expected demand, while safety stock reflects forecast uncertainty and the cost of stockouts.

Managerial interpretation

The model does not imply that demand will continue growing indefinitely. The distributor therefore also prepares a damped-trend sensitivity forecast. If a planned price increase is introduced, the historical model is not treated as a price-response model. Instead, the forecast is adjusted using an independently estimated elasticity or a controlled market test. Forecast accuracy is monitored monthly, and the model is re-evaluated if a competitor enters or distribution changes.

Learning points

The case illustrates five principles: benchmark before optimizing; preserve temporal order; diagnose rather than assume seasonality; evaluate with business-relevant errors; and separate descriptive forecasting from causal pricing analysis.

Limitations, governance, and contemporary practice

Forecasts can fail because of structural breaks, missing data, inconsistent definitions, stockouts, promotions, channel changes, and endogeneity. Observed sales may understate true demand when inventory was unavailable. A model trained on sales alone may therefore learn supply constraints rather than customer demand.

Forecast systems should document data sources, transformations, excluded periods, model versions, parameter choices, validation windows, and override policies. Human overrides may incorporate information unavailable to the model, but they should be recorded and evaluated rather than silently applied. Governance should include access controls, audit trails, monitoring for drift, and clear accountability for consequential decisions.

AI and machine-learning methods can model nonlinearities, large numbers of predictors, and multiple seasonalities. They do not eliminate the need for time-aware validation, leakage prevention, uncertainty quantification, or interpretability. A complex model may be inappropriate when the series is short, the decision requires explanation, or the underlying process has recently changed. Hybrid systems can combine decomposition or ARIMA structure with machine-learning models for residual patterns, but the added complexity must produce measurable decision value.

Prediction intervals widen with the forecast horizon because uncertainty accumulates. They should not be confused with confidence intervals for an estimated parameter. A prediction interval concerns a future observation and incorporates both parameter uncertainty and future innovation uncertainty, depending on the method used.

Best-practice checklist

1. Define the forecast target, horizon, frequency, and decision owner.
2. Verify that the measured series represents demand rather than merely fulfilled sales.
3. Plot the data before modeling and document anomalies.
4. Establish naïve and seasonal-naïve benchmarks.
5. Match the method to trend, seasonality, autocorrelation, and data availability.
6. Use rolling or expanding time-based validation.
7. Evaluate point accuracy and interval calibration where relevant.
8. Inspect residuals and monitor forecast drift after deployment.
9. Incorporate promotions, prices, holidays, and interventions only when their future values are known or forecast separately.
10. Communicate uncertainty and avoid presenting a forecast as a fact.

Chapter Summary

Time-series forecasting uses the historical ordering of observations to estimate future demand. A series may contain trend, fixed seasonality, irregular cycles, shocks, and autocorrelation. Naïve and seasonal-naïve forecasts provide essential benchmarks. Moving averages smooth noise, while exponential-smoothing methods estimate level, trend, and seasonality. Holt–Winters is especially useful for evolving level, trend, and seasonal patterns. Decomposition supports diagnosis and communication but does not establish causality. ARIMA and SARIMA model dependence after appropriate differencing and are valuable when autocorrelation remains.

Reliable forecasting requires residual diagnosis, time-respecting validation, appropriate error metrics, prediction intervals, and monitoring. Method choice should reflect the business decision, not only statistical fit. Forecasting historical demand is distinct from estimating price elasticity or other causal effects. Ethical and professional practice requires transparent documentation, protection against leakage, recognition of structural breaks, and responsible communication of uncertainty.

Key Terms

Trend: Persistent long-run movement in a time series.

Seasonality: A pattern that repeats at a known calendar frequency.

Cycle: A relatively long fluctuation without a fixed, known period.

Irregular component: Unsystematic variation or shocks not explained by the main components.

Stationarity: Stability of the series’ distributional properties over time.

Autocorrelation: Dependence between observations separated by a specified lag.

Lag: The number of time periods between two observations.

Naïve forecast: A forecast equal to the most recent observation.

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

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

Exponential smoothing: Recursive forecasting that assigns declining influence to older observations.

Holt’s method: Exponential smoothing with level and trend components.

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

Decomposition: Separation of a series into trend, seasonal, cyclical, and irregular components.

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

ARIMA: An autoregressive integrated moving-average model.

SARIMA: ARIMA extended with seasonal autoregressive, differencing, and moving-average terms.

Rolling-origin validation: Repeated time-ordered training and testing for forecast evaluation.

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

Structural break: A change in the data-generating process or relationship over time.

Review Questions

Short Answer Questions

1. What distinguishes seasonality from a business cycle?
2. Why is a seasonal-naïve forecast a useful benchmark?
3. What does the smoothing parameter α control in simple exponential smoothing?
4. Why can MAPE be inappropriate when actual demand is zero or near zero?
5. What does autocorrelation in residuals indicate?

Long Answer Questions

1. Compare moving averages, simple exponential smoothing, Holt’s method, and Holt–Winters in terms of assumptions and business suitability.
2. Explain how decomposition can be used to diagnose a demand series and discuss its limitations.
3. Describe a complete workflow for identifying, estimating, validating, and diagnosing an ARIMA model.
4. Discuss why time-series forecasting and price-elasticity estimation should not be treated as the same analytical task.
5. Evaluate the managerial risks of using a statistically accurate forecast without prediction intervals or structural-break monitoring.

Analytical / Application Questions

1. A retailer has weekly demand with a strong annual pattern, moderate growth, and increasing seasonal amplitude. Which benchmark and model family would you test first, and why?
2. A forecast model has lower training RMSE than a seasonal-naïve model but higher rolling-origin test MAE. Which model should management prefer, and what further diagnosis is required?
3. A product experienced six months of stockouts during the training period. Explain how this affects demand forecasting and propose a method for improving the forecast inputs or model design.

Multiple Choice Questions with Answers

Question 1

Which method forecasts the next observation using the most recently observed value?

A. Naïve method

B. Holt–Winters method

C. Seasonal decomposition

D. SARIMA method

Correct Answer: A. Naïve method

Question 2

A monthly series has a recurring annual pattern. What is its seasonal period?

A. 4

B. 7

C. 12

D. 52

Correct Answer: C. 12

Question 3

If the last three demands are 120, 132, and 138 units, what is the three-period moving-average forecast?

A. 126 units

B. 130 units

C. 132 units

D. 390 units

Correct Answer: B. 130 units

Question 4

What is the principal limitation of MAPE when actual demand is zero?

A. It has demand units

B. It cannot be calculated

C. It always penalizes large errors quadratically

D. It requires stationary data

Correct Answer: B. It cannot be calculated

Question 5

Which method is designed for level data without systematic trend or seasonality?

A. Simple exponential smoothing

B. Holt’s linear trend method

C. Seasonal ARIMA only

D. Multiplicative decomposition only

Correct Answer: A. Simple exponential smoothing

Question 6

In a weighted moving average, what condition preserves the scale of the forecast?

A. All weights must be negative

B. Weights must sum to zero

C. Weights must sum to one

D. Every weight must equal two

Correct Answer: C. Weights must sum to one

Question 7

In ARIMA(p,d,q), what does d represent?

A. The number of seasonal cycles

B. The order of nonseasonal differencing

C. The number of explanatory variables

D. The forecast horizon

Correct Answer: B. The order of nonseasonal differencing

Question 8

Why is random train-test splitting generally inappropriate for ordinary time-series validation?

A. It always reduces sample size to one observation

B. It can allow future information to influence training

C. It prevents calculation of MAE

D. It removes all seasonality

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

Question 9

What is the main purpose of residual diagnostics?

A. To guarantee that demand will rise

B. To identify remaining structure or systematic forecast failure

C. To replace the need for a benchmark

D. To convert units from cases to dollars

Correct Answer: B. To identify remaining structure or systematic forecast failure

Question 10

Which statement best distinguishes forecasting from causal price analysis?

A. Forecasting always estimates elasticity

B. Forecasting can describe historical continuation, whereas elasticity requires identifying price effects

C. Causal analysis never uses time-series data

D. Forecasting does not require historical observations

Correct Answer: B. Forecasting can describe historical continuation, whereas elasticity requires identifying price effects