Decomposition and Seasonal Adjustment

Introduction

Time-series observations often combine several forces: long-run growth or decline, recurring calendar effects, business-cycle movement, and irregular shocks. A retailer’s monthly sales may rise because the population is growing, peak every December because of holidays, and fall in one month because a distribution centre closed. If these forces are not separated, analysts may mistake a seasonal peak for sustainable growth or interpret an exceptional shock as a change in underlying demand.

Decomposition is the systematic representation of an observed time series as the combination of interpretable components. Seasonal adjustment removes predictable seasonal effects so that analysts can compare adjacent periods, identify turning points, and estimate the underlying trend more clearly. These methods are descriptive rather than automatically causal: decomposition identifies patterns in the data, but it does not by itself explain why those patterns occurred.

This chapter develops classical decomposition, moving-average filters, additive and multiplicative structures, STL decomposition, and practical seasonal-adjustment decisions. It also considers limitations, revision risk, and the use of adjusted data in business forecasting.

Learning objectives

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

1. Distinguish trend-cycle, seasonal, and irregular components.
2. Select and interpret additive or multiplicative decomposition.
3. Calculate centred moving averages and seasonal indices.
4. Explain the logic and strengths of STL decomposition.
5. Produce and interpret seasonally adjusted observations.
6. Evaluate decomposition outputs for forecasting and managerial decisions.
7. Recognise limitations, revisions, calendar effects, and governance risks.

1. The structure of a time series

Let y_t denote the observed value at time t, where observations are equally spaced—for example, monthly or quarterly. A useful conceptual decomposition is:

• Trend-cycle component: smooth, medium- to long-term movement. Trend refers to persistent direction; cycle refers to slower fluctuations around that direction.
• Seasonal component: systematic movement recurring within a fixed calendar period, such as months within a year or quarters within a year.
• Irregular component: residual movement caused by unpredictable events, measurement error, strikes, extreme weather, or temporary disruptions.

For monthly data, the seasonal period is usually m = 12; for quarterly data, m = 4. The period must be known or estimated from the data-generating process. A seven-day retail series may have weekly seasonality, while an hourly series may have daily and weekly seasonalities.

A decomposition is not the same as a forecast. It is a representation of historical observations. Forecasting requires extrapolating components or modelling the original series and then evaluating predictive performance.

2. Additive and multiplicative models

Additive decomposition

The additive model is:

where T_t is the trend-cycle component, S_t is the seasonal component, and I_t is the irregular component. All terms have the same units as y_t, such as units sold or dollars.

Additivity is appropriate when seasonal fluctuations have approximately constant absolute size. For example, if demand is usually about 200 units higher in December than in November regardless of the overall level, an additive model may be suitable. In a full seasonal cycle, seasonal effects generally sum to zero:

This constraint makes the seasonal component a deviation from the average level rather than an additional permanent level.

Multiplicative decomposition

The multiplicative model is:

Here, the components are dimensionless relative factors except for the level represented by T_t, and the seasonal indices typically average to one:

A seasonal index of 1.20 means that the expected value is approximately 20% above the underlying trend-cycle level for that season; 0.85 means approximately 15% below it. Multiplicative decomposition is appropriate when seasonal variation grows as the series level grows—for instance, when December sales are consistently 40% higher than the underlying level rather than higher by a fixed number of units.

Taking logarithms converts a positive multiplicative model into an additive one:

This requires positive observations because ln(0) and the logarithm of a negative value are undefined. A log transformation can stabilise variance, but it changes the scale on which interpretation and back-transformation occur.

Choosing a structure

The choice should be based on subject knowledge, plots, residual behaviour, and out-of-sample validation. Ask whether seasonal amplitude is stable in absolute units or approximately proportional to the level. A high correlation between the seasonal amplitude and the level can support a multiplicative structure, but correlation alone is not decisive. Zero or negative observations, changing product mix, and promotions may make a simple multiplicative model inappropriate.

3. Moving averages and trend-cycle estimation

A moving average replaces each observation with the average of nearby observations. A centred moving average of odd order [[MATH=k]] is:

This formula applies when k is odd and observations are equally spaced. It produces a value at time t by giving equal weight to the k observations centred on t. The first and last (k-1)/2 observations cannot receive a centred value.

For an even seasonal period, such as 12 months, a 12-term moving average lies halfway between months. It must therefore be centred by averaging two consecutive 12-term moving averages:

The resulting centred moving average estimates the trend-cycle component under the assumption that seasonal effects average out over a complete year and that irregular effects are not systematically aligned.

Worked example: a three-period centred moving average

Suppose quarterly sales are 100, 112, and 120 units in three consecutive quarters. The centred three-quarter moving average for the middle quarter is:

The moving average is not a forecast of the second quarter; it is a smoothed estimate of the local level. It uses a future observation relative to the second quarter and is therefore unsuitable as a real-time forecasting rule unless a one-sided version is used.

Moving averages have several important properties. A longer window produces greater smoothing but responds more slowly to turning points. A shorter window retains more short-run movement but may leave seasonal or irregular noise. Equal weights are transparent, but they are not always statistically optimal.

4. Classical decomposition procedure

Classical decomposition is a practical sequence rather than a single estimator. For an additive model, the usual procedure is:

1. Estimate T_t using a moving average appropriate to the seasonal period.
2. Compute detrended values y_t − T_t.
3. Group detrended values by season—for example, all January values.
4. Average within each season to estimate preliminary seasonal effects.
5. Normalise the seasonal estimates so they sum to zero.
6. Calculate irregular values as I_t = y_t − T_t − S_t.

For a multiplicative model:

1. Estimate T_t.
2. Compute ratios y_t/T_t.
3. Group ratios by season and average them.
4. Normalise indices so their average is one.
5. Calculate irregular values as I_t = y_t/(T_t × S_t).

The method assumes that the seasonal pattern is reasonably stable over the sample and that the moving average removes most seasonal movement. It loses observations at the ends and can be distorted by outliers or structural breaks.

Seasonal adjustment

Once the seasonal component is estimated, the seasonally adjusted series is obtained by removing it. Under an additive model:

Under a multiplicative model:

The adjusted value retains the original measurement units. If monthly sales are 1,200 units and the estimated December seasonal index is 1.20, the adjusted value is:

The interpretation is that sales were equivalent to 1,000 units at a neutral seasonal position. This does not mean that 200 units were literally caused by seasonality; it is an estimated statistical adjustment.

5. A compact classical example

Consider an illustrative multiplicative series whose estimated trend-cycle for a month is 500 units, seasonal index is 1.30, and irregular factor is 0.96. The reconstructed observation is:

The seasonally adjusted value is:

The adjusted value differs from the trend-cycle estimate because the irregular factor is below one. The seasonal adjustment removes the recurring month effect but does not remove the unusual shock represented by the irregular factor.

For an additive example, suppose T_t = 800, S_t = 90, and I_t = −20 units. Then:

The adjusted value is 870 − 90 = 780 units. Again, adjustment removes seasonality but leaves the irregular movement.

6. STL decomposition

STL means Seasonal-Trend decomposition using Loess. Loess, or locally estimated scatterplot smoothing, fits local regressions using nearby observations and weights them according to distance. STL repeatedly estimates seasonal and trend components through a structured smoothing procedure.

A common additive STL representation is:

where R_t is the remainder. STL is particularly useful because it can:

• allow the seasonal pattern to change gradually over time;
• use robust weighting to reduce the influence of outliers;
• handle many seasonal periods and long series flexibly;
• separate smooth trend movement from evolving seasonality without requiring a fixed seasonal pattern.

STL requires the analyst to choose smoothing spans, including a seasonal window and trend window. A fixed seasonal window estimates a relatively stable seasonal pattern; a larger or effectively infinite window allows slower seasonal evolution. Robust STL downweights observations with large residuals, which can prevent a single shock from distorting the estimated seasonal pattern.

STL is not automatically superior. Its output depends on smoothing choices, can be difficult to reproduce if settings are undocumented, and does not by itself provide a probabilistic forecasting model. For forecasting, STL may be combined with a model for the seasonally adjusted remainder, such as exponential smoothing or autoregression.

7. Classical decomposition compared with STL

Neither method establishes causality. If sales rise each December, decomposition identifies recurrence but cannot distinguish holidays, promotions, weather, or reporting practices without additional information.

8. Calendar effects and practical seasonal adjustment

Seasonality is sometimes confused with calendar composition. A month may contain different numbers of weekdays, weekends, trading days, or public holidays. Easter, for example, moves between months and may create an apparent seasonal change that is not fixed to a particular month. Leap years and changing month lengths also matter.

Analysts should distinguish:

• Seasonal effects: recurring effects tied to the position within a regular cycle.
• Calendar effects: variation caused by the number or type of working days, moving holidays, or trading days.
• Event effects: one-off interventions such as a product launch or strike.

A robust adjustment process may regress the series on calendar and intervention variables before or alongside decomposition. For official statistics, seasonal-adjustment systems often use model-based filters, diagnostics, outlier detection, and revision policies rather than a simple moving average alone.

Year-over-year comparisons can reduce seasonal influence, but they do not remove all calendar or irregular effects. For a monthly series, a year-over-year growth rate is often expressed as:

This formula is meaningful when the denominator is non-zero and the measurement scale supports percentage interpretation. It compares the same month across years but may hide a recent turning point and can be distorted by an unusual value in the comparison month.

9. Applied case: regional grocery demand

Background

The following is a fictional illustrative case. A regional grocery chain has three years of monthly sales data. Management wants to determine whether sales growth is genuine or simply the result of increasingly strong holiday demand. The data show a rising level, large December peaks, and several unusual months during a supply disruption.

Problem

The finance team uses raw month-to-month changes to set inventory targets. This produces repeated over-ordering before December and under-ordering after the holiday period. Executives also disagree about whether the supply disruption permanently reduced demand.

Analysis

The analyst plots the series and observes that seasonal amplitude increases with sales level, supporting a multiplicative starting model. A 12-month centred moving average estimates the trend-cycle. Ratios of actual sales to the moving average are grouped by month to estimate seasonal indices. December has an index of 1.28, while February has an index of 0.82. The indices are normalised to average one.

Because the disruption creates unusually low observations, the analyst also performs STL with robust weighting. The STL seasonal component changes modestly over time, while the trend falls sharply during the disruption and recovers afterward. This suggests that the event is not merely seasonality. It is an irregular or intervention effect that should be modelled separately when forecasting.

Approach and solution

Management receives three series: raw sales, seasonally adjusted sales, and a forecast based on adjusted demand plus separately modelled holiday and intervention effects. Inventory decisions use the raw series for expected physical demand in each month, but adjusted sales for assessing underlying momentum. The analyst documents the seasonal period, STL settings, outlier treatment, data revisions, and the date on which each estimate was available.

Outcome

The chain finds that underlying demand grew moderately before the disruption and recovered gradually afterward. The apparent dramatic December growth was mostly seasonal. Forecast accuracy improves because the model no longer treats every December peak as permanent trend growth. Inventory policy is also improved by combining seasonal indices with lead-time and service-level analysis.

Learning points

1. A seasonal adjustment is a decision aid, not a replacement for operational knowledge.
2. A fixed seasonal profile may be inadequate when customer behaviour changes.
3. Exceptional events should not be absorbed into seasonal indices without examination.
4. Raw and adjusted data serve different purposes: planning physical demand often needs seasonality, while monitoring underlying momentum often benefits from adjustment.

10. Diagnostics and quality control

A decomposition should be audited rather than accepted automatically. Useful checks include:

• Plot the original, trend, seasonal, and remainder series.
• Confirm that additive seasonal effects sum approximately to zero or multiplicative indices average approximately one.
• Inspect the remainder for residual seasonality, changing variance, autocorrelation, and outliers.
• Check whether the estimated seasonal pattern is plausible to subject-matter experts.
• Compare alternative additive, multiplicative, and STL specifications.
• Use rolling-origin or time-based validation when decomposition supports forecasting.
• Examine endpoint estimates, which are often revised when new observations arrive.
• Record transformations, missing-value treatment, smoothing windows, and software versions.

A useful residual measure for an additive decomposition is the remainder R_t = y_t − T_t − S_t. If systematic structure remains in R_t, the decomposition has not captured all relevant dynamics. For a multiplicative decomposition, the corresponding remainder is R_t = y_t/(T_t × S_t). A remainder near one indicates close reconstruction; a remainder near zero in an additive model indicates little unexplained movement.

11. Limitations and professional judgement

End-point problem

Centred filters require observations on both sides of a date. At the start and end of a sample, one-sided filters or extrapolation are needed. Consequently, recent seasonally adjusted values may change substantially when new data arrive. Users should not interpret the latest estimate as final.

Structural change

A pandemic, regulatory change, product redesign, or new competitor can alter both trend and seasonal behaviour. Historical averages may then be poor estimates of current seasonality.

Outliers and interventions

An extreme observation can inflate a seasonal index if it occurs in the same calendar period as other observations. Robust methods help, but analysts should investigate the event and consider intervention variables rather than silently deleting data.

Aggregation and disaggregation

Seasonal adjustment performed on an aggregate series may not equal the sum of adjustments performed separately on component series. Organisations should define whether consistency across products, regions, and totals is required and apply an appropriate reconciliation policy.

Identification and causality

A repeated pattern is not necessarily a seasonal cause. Marketing campaigns, pricing, capacity constraints, and reporting deadlines can create recurring-looking patterns. Decomposition should be combined with business knowledge and, where relevant, causal designs.

12. Technology, AI, and governance

Modern forecasting software can automate decomposition, select smoothing parameters, detect outliers, and generate adjusted series. Machine-learning systems may include seasonal features, Fourier terms, calendar variables, or decomposition layers. These tools are useful, but automation does not remove the need for validation.

Governance requires a clear data lineage: the raw series, adjustment method, parameter settings, release date, and revision history should be retained. Analysts should distinguish an estimate available in real time from a retrospectively revised estimate. This distinction is essential when evaluating forecast accuracy or making performance claims.

AI-generated explanations should be checked against the numerical output. A model may label a change as seasonal even when it is caused by a policy intervention or data-quality problem. Sensitive commercial series also require access controls, privacy safeguards, and careful consideration of whether external services may retain uploaded data.

Best-practice workflow

1. Define the business decision and time frequency.
2. Inspect plots, units, missing values, and calendar structure.
3. Identify the seasonal period using domain knowledge and diagnostics.
4. Compare additive and multiplicative representations.
5. Estimate components using classical decomposition and, where appropriate, STL.
6. Investigate outliers, interventions, holidays, and structural breaks.
7. Validate the remainder and compare forecast performance out of sample.
8. Communicate raw, adjusted, and forecast values separately.
9. Document settings and expected revisions.
10. Reassess the decomposition when products, policies, or customer behaviour change.

Chapter Summary

Decomposition separates an observed time series into interpretable trend-cycle, seasonal, and irregular components. Additive models express these components as sums and are suitable when seasonal variation is stable in absolute units. Multiplicative models express them as products and are suitable when seasonal variation is proportional to the level. Moving averages provide transparent estimates of the trend-cycle, while classical decomposition estimates seasonal effects by averaging detrended values. STL uses Loess smoothing to permit evolving seasonality and robust treatment of outliers.

Seasonal adjustment removes estimated seasonal effects: subtraction for additive models and division for multiplicative models. Adjustment supports comparisons of underlying momentum, but it does not remove irregular shocks or establish causality. Calendar effects, moving holidays, endpoint uncertainty, structural breaks, revisions, and outliers require professional judgement. In practice, analysts should compare methods, inspect remainders, validate forecasting performance, document settings, and match raw or adjusted data to the decision being made.

Key Terms

Decomposition: Representation of a time series through distinct underlying components.

Trend-cycle: Smooth long- and medium-term movement in a series.

Seasonal component: Recurring movement associated with a fixed position in a calendar cycle.

Irregular component: Unsystematic residual movement after other components are estimated.

Additive model: Decomposition in which components are summed.

Multiplicative model: Decomposition in which components are multiplied.

Seasonal index: Estimated seasonal factor or effect for a particular period.

Seasonal adjustment: Removal of estimated seasonal effects from observations.

Moving average: Average of observations within a specified moving window.

Centred moving average: Moving average aligned with the middle observation or period.

Trend-cycle estimate: Smoothed estimate intended to suppress seasonal and irregular variation.

STL: Seasonal-Trend decomposition using Loess.

Loess: Local regression smoothing based on nearby, distance-weighted observations.

Remainder: Residual component left after trend and seasonality are removed.

Calendar effect: Variation caused by working days, holidays, month length, or related calendar composition.

Intervention: Explicit representation of an unusual event or structural change.

Endpoint problem: Uncertainty caused by limited observations at the beginning or end of a series.

Robust decomposition: Decomposition designed to reduce the influence of outliers.

Revision: Change to a previously published estimate when new information becomes available.

Review Questions

5 Short Answer Questions

1. What are the three principal components in a classical time-series decomposition?
2. When is an additive model preferable to a multiplicative model?
3. Why is a 12-month moving average centred for monthly data?
4. How is a multiplicative seasonal adjustment calculated?
5. What is the endpoint problem?

5 Long Answer Questions

1. Explain the classical decomposition procedure for an additive time series, including its assumptions and limitations.
2. Compare additive and multiplicative decomposition using equations, interpretation, and an example of an appropriate business context for each.
3. Explain how STL differs from classical decomposition and evaluate the situations in which STL is likely to be useful.
4. Discuss why calendar effects, interventions, and outliers must be distinguished from ordinary seasonality.
5. Design a quality-control process for a seasonally adjusted business series used in executive reporting.

3 Analytical / Application Questions

1. A retailer’s monthly sales rise from 1,000 units in March to 1,400 units in December. The estimated December seasonal index is 1.25. Calculate the December seasonally adjusted sales and explain the result.
2. A decomposition produces a remainder that still has strong peaks every December. What does this indicate, and what steps should the analyst take?
3. A company’s seasonal amplitude grows as its overall sales level grows, but a major product launch changes the pattern halfway through the sample. Recommend a decomposition and validation strategy, giving reasons.

Multiple Choice Questions with Answers

Question 1

Which equation represents an additive decomposition?

A. y_t = T_t + S_t + I_t

B. y_t = T_t × S_t × I_t

C. y_t = T_t/S_t + I_t

D. y_t = T_t − S_t × I_t

Correct Answer: A. y_t = T_t + S_t + I_t

Question 2

In a multiplicative decomposition, what does a seasonal index of 1.20 usually indicate?

A. The observation is 1.20 units above trend

B. The observation is approximately 20% above the trend-cycle level, before irregular effects

C. The observation is approximately 20 units below trend

D. The seasonal effect is zero

Correct Answer: B. The observation is approximately 20% above the trend-cycle level, before irregular effects

Question 3

For monthly data, why is a 12-term moving average commonly centred?

A. Because 12 is an odd number

B. Because a 12-term average falls halfway between calendar months

C. To eliminate all irregular variation

D. To create a forecast for the next 12 months

Correct Answer: B. Because a 12-term average falls halfway between calendar months

Question 4

If an additive seasonal effect is 80 units and the observed value is 620 units, what is the seasonally adjusted value?

A. 540 units

B. 620 units

C. 700 units

D. 7.75 units

Correct Answer: A. 540 units

Question 5

If an observed value is 900 units and its multiplicative seasonal index is 1.20, what is the seasonally adjusted value?

A. 750 units

B. 901.20 units

C. 1,080 units

D. 899 units

Correct Answer: A. 750 units

Question 6

Which feature is particularly associated with STL decomposition?

A. It requires seasonal effects to remain exactly constant

B. It can allow seasonal patterns to evolve gradually over time

C. It removes the need for forecasting validation

D. It can only be applied to annual data

Correct Answer: B. It can allow seasonal patterns to evolve gradually over time

Question 7

Which statement about seasonal adjustment is correct?

A. It removes every unusual shock from a series

B. It proves the causal reason for a recurring pattern

C. It removes an estimated recurring seasonal effect while generally leaving irregular effects

D. It always increases forecast accuracy

Correct Answer: C. It removes an estimated recurring seasonal effect while generally leaving irregular effects

Question 8

A multiplicative decomposition has trend-cycle 500, seasonal index 1.30, and irregular factor 0.96. What observation is reconstructed?

A. 480 units

B. 624 units

C. 650 units

D. 1,300 units

Correct Answer: B. 624 units

Question 9

Which issue is most associated with the beginning and end of a centred moving-average decomposition?

A. The endpoint problem

B. Perfect seasonal identification

C. Automatic causal inference

D. Constant forecast variance

Correct Answer: A. The endpoint problem

Question 10

Why should calendar effects be distinguished from ordinary seasonality?

A. Calendar effects never recur

B. Working-day counts and moving holidays can vary across nominally comparable periods

C. Seasonality only occurs in annual data

D. Calendar effects cannot influence business data

Correct Answer: B. Working-day counts and moving holidays can vary across nominally comparable periods