Classical Time Series Forecasting Models

Introduction

Time series forecasting uses observations recorded in time order to estimate future values. Examples include daily sales, monthly electricity demand, quarterly revenue, website traffic and annual employment. Unlike ordinary cross-sectional data, time series observations are usually related to one another. Recent demand may resemble previous demand, and sales may rise or fall according to a recurring seasonal pattern.

Classical forecasting methods provide an interpretable foundation for predictive analytics. They are especially useful when the analyst needs transparent models, limited computational resources and forecasts based primarily on the historical behaviour of one variable. This chapter examines decomposition, moving averages, simple exponential smoothing, Holt’s method and Holt–Winters methods. It also explains how to select models and construct forecast intervals.

Learning Objectives

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

• identify level, trend, seasonality and irregular variation in a time series;
• decompose a series using additive and multiplicative structures;
• apply simple exponential smoothing to level-only data;
• apply Holt’s method to data with trend;
• apply Holt–Winters methods to data with trend and seasonality;
• select a forecasting method using data characteristics and forecast errors;
• interpret point forecasts and forecast intervals;
• recognise the assumptions and limitations of classical models.

1. Components of a Time Series

A time series is commonly represented as y_t, where t denotes the time period and y_t is the observed value. Classical analysis separates the observed value into systematic and irregular components.

The principal components are:

• Level: the typical magnitude of the series.
• Trend: a persistent long-term increase or decrease.
• Seasonality: a pattern that repeats at a known and fixed frequency, such as every 12 months.
• Irregular or remainder variation: unpredictable movement not explained by the model.
• Cycle: longer-term rises and falls that do not necessarily have a fixed period. Cycles are distinct from seasonality, although the terms are sometimes used loosely.

A useful conceptual model is:

This is the additive decomposition, where T_t is the trend-cycle component, S_t is the seasonal component and R_t is the remainder. It is appropriate when seasonal fluctuations have approximately constant absolute size.

A multiplicative decomposition is:

Here, seasonal variation is proportional to the level. For example, a retailer whose December sales are approximately 30% above its underlying level may require a multiplicative model. Multiplicative decomposition requires positive observations because ratios and logarithms are involved in many implementations.

2. Exploring and Decomposing Time Series

Before fitting a model, plot the observations against time. A graph can reveal missing observations, outliers, changing variability, structural breaks, trend and seasonal repetition. Aggregation must also be checked: weekly data may have a different seasonal period from monthly data.

Moving averages

A centred moving average smooths short-term variation. For an odd window of length m = 2k + 1, the moving average at time t is:

The moving average has the same units as the observations. It cannot be calculated at the first and last k periods unless a boundary method is used. Moving averages are useful for visualising the underlying trend, but they are not automatically good forecasting models: a centred moving average uses future observations and therefore cannot be used directly for a real-time forecast.

For even seasonal periods, such as 12 months, analysts often calculate a centred moving average by averaging adjacent 12-month moving averages. This aligns the estimate with an actual month.

Seasonal indices

In an additive decomposition, seasonal indices usually sum approximately to zero over one complete seasonal cycle:

In a multiplicative decomposition, seasonal indices usually average approximately one:

where m is the number of observations per seasonal cycle. A multiplicative seasonal index of 1.20 means that the period is typically 20% above the underlying level, while 0.85 means 15% below it.

Decomposition is valuable diagnostically. If the remainder still displays trend or seasonality, the decomposition has not adequately captured the structure. If the remainder has increasing variance, a transformation or multiplicative model may be more appropriate.

3. Exponential Smoothing

Exponential smoothing assigns greater weight to recent observations while retaining information from the past. Unlike a fixed moving average, its effective weights decline geometrically rather than stopping abruptly.

The smoothing parameter is commonly denoted by α, where 0 < α < 1. A high value reacts quickly to recent changes; a low value produces a smoother estimate and reacts more slowly.

Simple exponential smoothing

Simple exponential smoothing (SES) is designed for a series with a changing level but no systematic trend or seasonality. The level equation is:

The one-step-ahead forecast is:

For all future horizons, the forecast remains constant:

Here, ℓ_t is the estimated level after observing period t, y_t is the actual observation, ŷ_t+h|t is the forecast made at time t for period t+h, and h is the forecast horizon. All these quantities have the same units as y_t.

The initial level ℓ₀ must be specified. A common choice is the first observation, although the average of several early observations or an estimated value may be preferable. The effect of initialization declines over time when 0 < α < 1.

Worked example: simple exponential smoothing

Suppose the initial level is 100, α = 0.30, and the first three observations are 110, 105 and 120.

After period 1:

After period 2:

After period 3:

The forecast for periods 4, 5 and 6 is therefore 108.52 units. The method does not extrapolate a trend; it estimates the current level.

4. Holt’s Method for Trend

Simple exponential smoothing is unsuitable when a trend is persistent because its forecasts are flat. Holt’s linear trend method estimates both level and trend.

The level equation is:

The trend equation is:

The forecast is:

where b_t is the estimated change per time period and 0 < β < 1 is the trend-smoothing parameter. The level and trend must use compatible time units. For monthly data, b_t is measured in units per month and h is measured in months.

Holt’s method may produce unrealistic forecasts when a trend continues indefinitely. A damped-trend version limits long-range growth by multiplying successive trend contributions by a damping parameter 0 < φ < 1:

The damped method is useful when recent trend is real but indefinite linear extrapolation is not credible.

5. Holt–Winters Methods for Trend and Seasonality

Holt–Winters methods extend Holt’s method by estimating a seasonal component. They are appropriate when a series contains level, trend and repeating seasonality. Let m be the number of observations per season—for example, m = 12 for monthly data with annual seasonality.

Additive Holt–Winters method

The additive equations are:

For the final equation, the seasonal term is selected from the appropriate position in the repeating seasonal cycle; operational implementations commonly express this using a seasonal index such as s_{t−m+1+((h−1) mod m)}. The parameters satisfy 0 < α, β, γ < 1.

Additive seasonality is suitable when the seasonal effect is approximately the same number of units regardless of the level.

Multiplicative Holt–Winters method

The multiplicative equations are:

Multiplicative Holt–Winters requires positive observations and positive seasonal factors. It is appropriate when seasonal variation is proportional to the level. A forecast of 1,000 units multiplied by a seasonal factor of 1.25 becomes 1,250 units.

Parameter estimation is usually performed by minimising an objective function such as the sum of squared one-step-ahead errors. The parameters should be estimated using training data only; otherwise, information from the evaluation period can leak into the model.

6. Choosing an Appropriate Model

Model choice should begin with the data-generating pattern rather than with a preference for a particular algorithm.

A time series should be divided chronologically into training and test sets. Random splitting is generally inappropriate because it allows future information to enter the training data. Rolling-origin evaluation is stronger: fit the model using an initial history, forecast the next period or several periods, expand the history and repeat.

Common error measures include:

Here, n is the number of forecast errors, y_i is an actual value and ŷ_i is its forecast. MAE is in the original units. RMSE penalises large errors more heavily. MAPE is undefined when an actual value is zero and can be misleading when actual values are close to zero. For intermittent or zero-valued demand, MAE, RMSE or scaled measures are usually safer.

Accuracy alone is not sufficient. A slightly less accurate model may be preferred if it is easier to explain, less sensitive to unusual observations, cheaper to maintain or more consistent with operational decisions. Residuals should have approximately zero mean, no remaining trend or seasonality, and limited autocorrelation. A residual plot and an autocorrelation diagnostic are therefore important.

7. Forecast Intervals

A point forecast is a single best estimate. A forecast interval communicates uncertainty. For a forecast ŷ_t+h|t, an approximate two-sided 95% interval is often written:

where SE_h is the standard error of the forecast at horizon h. The factor 1.96 comes from the standard normal approximation and is appropriate only when the approximation is defensible. The interval is not a statement that there is a 95% probability that a fixed future observation has already been randomly located inside the interval; rather, under the model and repeated-sampling interpretation, approximately 95% of similarly constructed intervals are expected to contain the future value.

For simple exponential smoothing, a rough method may use the residual standard deviation, but forecast uncertainty generally increases with horizon because more future disturbances can accumulate. For Holt and Holt–Winters models, analytical standard errors depend on the model structure and parameter estimates. Software implementations may use analytical formulas, simulation or bootstrap methods.

Prediction intervals should be distinguished from confidence intervals. A confidence interval concerns uncertainty about an estimated parameter or mean. A forecast interval concerns uncertainty about a future observation and is normally wider because it includes both estimation uncertainty and future random variation.

Intervals may be asymmetric for positive, multiplicative or transformed data. If a logarithmic model is used, intervals should be transformed back carefully; simply exponentiating the point forecast can introduce retransformation bias.

8. Applied Case: Monthly Café Demand

This is an illustrative case. A café records monthly demand for prepared meals over four years. Management needs forecasts for staffing and ingredient purchasing for the next six months.

Background and problem

A plot shows higher demand each summer, lower demand in winter and an upward long-term movement. Seasonal peaks become larger as overall demand increases. This suggests trend plus multiplicative seasonality. The business also reports occasional promotions, which create unusual observations.

Analysis and approach

The analyst first checks that all months are present and records promotional events. The final six months are reserved as a test set. A seasonal-naïve benchmark is created: each month’s forecast equals the observation from the same month in the previous year. The analyst then fits additive and multiplicative Holt–Winters models using the training period.

The multiplicative model is selected if it produces lower rolling-origin MAE and residuals without obvious remaining seasonality. However, promotional months are not automatically deleted. They are investigated because removing legitimate business events may make the forecast less realistic. If promotions will recur, their effects may need to be modelled separately or adjusted using known future information.

Outcome and interpretation

Suppose the selected model forecasts 2,400 meals for July with a 95% interval of 2,150 to 2,690. The point forecast supports a central staffing plan, while the upper interval informs contingency staffing and ingredient capacity. The interval does not guarantee that demand will stay within the range. A supply disruption, severe weather or a new competitor could produce an outcome outside it.

Learning points

The case illustrates five professional practices: inspect the data before modelling; compare against a simple benchmark; match additive or multiplicative seasonality to the observed scale; evaluate chronologically; and use intervals for operational risk planning rather than relying only on a point estimate.

9. Limitations, Governance and Professional Practice

Classical models are transparent but not universally adequate. They may perform poorly when relationships depend on external drivers, when there are abrupt structural breaks, when seasonal frequency changes, or when demand is intermittent. They also extrapolate historical patterns and cannot automatically understand business strategy, competitor entry or policy changes.

Forecasting systems should document data sources, time zones, missing-value treatment, outlier decisions, model parameters, training periods and evaluation procedures. Analysts should monitor forecast errors after deployment and establish rules for retraining. A forecast that affects staffing, credit, health or access to services should be reviewed for potential harm and should not be treated as objective merely because it is generated mathematically.

Modern forecasting platforms may automate model selection, rolling validation and interval estimation. Automation improves consistency but does not remove the need for judgment. A model must still be tested for data leakage, inappropriate seasonal periods, unstable intervals and operational plausibility. Machine-learning models may improve accuracy when many explanatory variables are available, but classical models remain important benchmarks because they are interpretable and often competitive for stable univariate series.

Best-Practice Checklist

1. Plot the series and verify its time frequency.
2. Check missing values, outliers, unusual events and structural changes.
3. Identify whether trend and seasonal variation are additive or proportional.
4. Reserve a chronological test period.
5. Compare against naïve or seasonal-naïve benchmarks.
6. Estimate parameters using training data only.
7. Examine residuals for remaining structure.
8. Report both point forecasts and forecast intervals.
9. Match the forecast horizon to the business decision.
10. Monitor performance and revise the model when the process changes.

Chapter Summary

Classical time series forecasting separates systematic behaviour from irregular variation. Additive decomposition is suitable when seasonal effects have roughly constant size, while multiplicative decomposition is suitable when seasonal effects vary with the level. Moving averages assist with smoothing and diagnosis but centred averages are not directly available for real-time forecasting.

Simple exponential smoothing estimates a changing level and produces flat future forecasts. Holt’s method adds a trend estimate, while damped trend limits excessive long-range extrapolation. Holt–Winters methods estimate level, trend and seasonality using either additive or multiplicative seasonal structures. Model selection should consider the pattern in the data, chronological validation, forecast accuracy, residual diagnostics, interpretability and operational consequences.

Forecast intervals are essential because future observations are uncertain. Their width generally increases with the forecast horizon, and their interpretation depends on the adequacy of the model and its assumptions. Responsible forecasting also requires documentation, monitoring, protection against data leakage and careful treatment of unusual events.

Key Terms

Time series: Observations recorded sequentially over time.

Level: The typical magnitude of a series.

Trend: Persistent long-term movement.

Seasonality: A pattern that repeats at a fixed frequency.

Cycle: Longer-term fluctuation without a necessarily fixed period.

Remainder: Variation left after systematic components are removed.

Additive decomposition: A structure in which components are summed.

Multiplicative decomposition: A structure in which components are multiplied.

Moving average: An average over neighbouring observations used for smoothing.

Exponential smoothing: A method giving greater weight to recent observations.

Simple exponential smoothing: A level-only smoothing method.

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

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

Damped trend: A trend whose future contribution gradually decreases.

Smoothing parameter: A value controlling the weight assigned to recent information.

Forecast horizon: The number of periods into the future being forecast.

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

MAE: Mean absolute error, measured in the original units.

RMSE: Root mean squared error, which gives greater weight to large errors.

Forecast interval: A range representing uncertainty about a future observation.

Review Questions

5 Short Answer Questions

1. What is the difference between trend and seasonality?
2. When is an additive decomposition more appropriate than a multiplicative decomposition?
3. What does the parameter α control in simple exponential smoothing?
4. Why is random train–test splitting usually inappropriate for time series?
5. What is the difference between a point forecast and a forecast interval?

5 Long Answer Questions

1. Explain the components of a classical time series and describe how decomposition can support forecasting.
2. Derive and interpret the level and forecast equations for simple exponential smoothing.
3. Compare simple exponential smoothing, Holt’s method and Holt–Winters methods.
4. Explain how additive and multiplicative seasonality affect model choice.
5. Discuss the assumptions, uses and limitations of forecast intervals.

3 Analytical / Application Questions

1. A monthly series has a stable upward trend, but its seasonal peaks grow as the level increases. Recommend a forecasting method and justify your choice.
2. A business obtains lower RMSE from a complex model but the model has strongly autocorrelated residuals. Explain what this indicates and propose next steps.
3. A forecast for next quarter is 5,000 units with a 95% interval of 4,400–5,700. Explain how a manager should use both the point forecast and interval when planning capacity.

Multiple Choice Questions with Answers

Question 1

Which model is designed for a series with level but no systematic trend or seasonality?

A. Simple exponential smoothing

B. Holt’s method

C. Multiplicative Holt–Winters

D. Seasonal decomposition only

Correct Answer: A. Simple exponential smoothing

Question 2

In an additive decomposition, seasonal effects are assumed to be approximately:

A. Proportional to the level

B. Constant in absolute size

C. Always positive

D. Equal to one

Correct Answer: B. Constant in absolute size

Question 3

If α = 0.30, what weight does simple exponential smoothing assign to the previous level?

A. 0.30

B. 0.70

C. 1.30

D. 3.00

Correct Answer: B. 0.70

Question 4

Which method explicitly estimates both level and trend?

A. Simple moving average

B. Simple exponential smoothing

C. Holt’s method

D. Seasonal index calculation

Correct Answer: C. Holt’s method

Question 5

A multiplicative Holt–Winters model generally requires observations to be:

A. All negative

B. Positive

C. Normally distributed

D. Measured annually

Correct Answer: B. Positive

Question 6

For monthly data with annual seasonality, the seasonal period m is usually:

A. 4

B. 7

C. 12

D. 52

Correct Answer: C. 12

Question 7

Which error measure is expressed in the original units of the target variable?

A. MAE

B. MAPE

C. Correlation

D. Seasonal index

Correct Answer: A. MAE

Question 8

Why should time series data normally be split chronologically for evaluation?

A. To increase the sample size

B. To prevent future information entering model training

C. To guarantee normal residuals

D. To remove seasonality

Correct Answer: B. To prevent future information entering model training

Question 9

A 95% approximate normal forecast interval is commonly formed as:

A. Forecast ± 0.05 × standard error

B. Forecast ± 1 × standard error

C. Forecast ± 1.96 × standard error

D. Forecast ± 95 × standard error

Correct Answer: C. Forecast ± 1.96 × standard error

Question 10

A seasonal-naïve forecast for a monthly observation usually uses:

A. The previous hour’s value

B. The average of all observations

C. The observation from the same month in the previous seasonal cycle

D. The maximum historical value

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