Seasonal Differences in Monthly and Quarterly Data

Why Seasonal Differences Compare Like With Like

A monthly series may rise every winter and fall every summer; a quarterly series may repeatedly have a high fourth quarter. Comparing adjacent observations mixes these seasonal movements with changes in the underlying level. A seasonal difference instead compares an observation with the value from the same season in the preceding cycle: January with the previous January, or the second quarter with the previous second quarter.

The seasonal period $s$ is the number of observations in one complete cycle. For monthly data with an annual cycle, $s=12$; for quarterly data with an annual cycle, $s=4$. The seasonal difference is an absolute difference in the original measurement units. It is not automatically a percentage change or a measure of change from the immediately preceding month or quarter.

Definition: For a regularly spaced series $Y_1,\ldots,Y_n$ and seasonal period $s$, the seasonal difference at time $t$ is $$ \nabla_s Y_t=Y_t-Y_{t-s}, \qquad t=s+1,\ldots,n. $$ It is defined only when the observation $Y_{t-s}$ is available. In lag-operator notation, with $B^sY_t=Y_{t-s}$, the seasonal difference operator is $1-B^s$.

The result $\nabla_sY_t$ is naturally attached to the later time $t$, because it describes how the current observation differs from its counterpart one cycle earlier. For example, $\nabla_{12}Y_{15}$ compares month 15 with month 3, not month 14. If there are $n$ observations and no additional earlier values are supplied, the number of calculated seasonal differences is $\max(n-s,0)$. In particular, it is $n-s$ when $n\ge s$, and zero when $n<s$.

How the Seasonal Difference Removes a Stable Seasonal Effect

A short algebraic argument explains why the comparison can be useful. Suppose a series can be represented as a nonseasonal component $T_t$, a stable seasonal effect $S_j$ for season $j$, and an irregular component $\varepsilon_t$. For a cycle of length $s$, time $t$ and time $t-s$ have the same seasonal position, so both have seasonal effect $S_j$. Then

Y_t=T_t+S_j+\varepsilon_t,\qquad Y_{t-s}=T_{t-s}+S_j+\varepsilon_{t-s},

and subtraction gives

\nabla_sY_t=Y_t-Y_{t-s} =(T_t-T_{t-s})+(\varepsilon_t-\varepsilon_{t-s}).

The stable seasonal effect cancels because the two observations share it. This does not mean seasonal differencing removes every possible seasonal pattern, or that the resulting series is necessarily stationary. For example, if the seasonal effect changes substantially from cycle to cycle, the two seasonal contributions will not be equal and will not cancel exactly. The arithmetic transformation is defined regardless; its usefulness depends on the series and the purpose of the analysis.

Worked Example: Monthly Values Compared With the Same Month One Year Earlier

A hypothetical insurer tracks a monthly claim-frequency index. The first 16 values are shown below. Calculate the available annual seasonal differences for months 13 through 16, and interpret them. The index is not a count of claims or a percentage.

Month	1	2	3	4	5	6	7	8
Index	100	120	150	180	210	240	230	220
Month	9	10	11	12	13	14	15	16
Index	190	160	130	110	108	131	155	187

Method. This is monthly data with an annual cycle, so $s=12$. Apply $\nabla_{12}Y_t=Y_t-Y_{t-12}$ to each available month.

For month 13, the comparison is with month $13-12=1$: $108-100=8$.

For month 14, the comparison is with month $14-12=2$: $131-120=11$.

For month 15, the comparison is with month $15-12=3$: $155-150=5$.

For month 16, the comparison is with month $16-12=4$: $187-180=7$.

Thus the available seasonal differences, attached to months 13–16, are $8,11,5,7$. For instance, the month-14 index is 11 index points greater than the month-2 index. These are absolute differences in index points, not percentage increases. As a check, 16 observations with $s=12$ give $16-12=4$ available differences, matching the four calculations.

Common mistake avoided: Subtracting the previous month would give $131-108=23$ for month 14. That is the ordinary one-month change, not the annual seasonal difference, which is $131-120=11$.

Quarterly Seasonal Differences

The same rule applies to quarterly observations. When the seasonal cycle is one year, $s=4$, so a quarterly value is compared with the value from the same quarter in the preceding year. For example, a third-quarter observation is compared with the third quarter one year earlier, not with the second quarter immediately before it.

Worked Example: Comparing Each Quarter With Its Prior-Year Counterpart

A hypothetical pension administrator records a quarterly service-volume index. For three years, the values are: Year 1: $40,55,70,65$; Year 2: $44,58,77,69$; Year 3: $48,62,82,73$. Find every available seasonal difference and interpret the Year 3 results.

Method. There are four quarters per annual cycle, so $s=4$. Number the observations consecutively: observations 1–4 are Year 1, 5–8 are Year 2, and 9–12 are Year 3. For each $t=5,\ldots,12$, subtract observation $t-4$.

For Year 2, the four calculations are:

First quarter: $44-40=4$.
Second quarter: $58-55=3$.
Third quarter: $77-70=7$.
Fourth quarter: $69-65=4$.

For Year 3, the comparisons are with the corresponding Year 2 quarters:

First quarter: $48-44=4$.
Second quarter: $62-58=4$.
Third quarter: $82-77=5$.
Fourth quarter: $73-69=4$.

In order from Year 2 quarter 1 through Year 3 quarter 4, the seasonal differences are $4,3,7,4,4,4,5,4$. The Year 3 fourth-quarter index is 4 points above the Year 2 fourth-quarter index. There are $12-4=8$ differences, which agrees with the eight calculations. This comparison says nothing by itself about why the index changed; it describes the same-quarter change.

Common mistake avoided: The Year 3 second-quarter seasonal difference is $62-58=4$, not $62-48=14$. The value 48 is the previous year's first quarter, so it is not the same-season comparison.

Difference Counts, Data Boundaries, and Interpretation

Seasonal differencing requires a matched earlier observation. Without one, a difference cannot be calculated. This is a data-boundary issue, not a zero-valued difference: the first $s$ observations have no seasonal difference within the observed series. If the series is shorter than $s$, none of its observations has a prior same-season value in that series.

This counting rule can be written in one expression:

\text{number of available seasonal differences} =\max(n-s,0).

For example, a series of 7 monthly observations with $s=12$ has $\max(7-12,0)=0$ available annual seasonal differences. It is incorrect to report $-5$ differences: a count cannot be negative. If earlier observations outside the displayed data are explicitly supplied, they may provide the missing comparisons; absent such values, they must not be assumed.

A seasonal difference is measured in the same units as the original series. If $Y_t$ is measured in dollars, $\nabla_sY_t$ is in dollars; if it is an index, the difference is in index points. A positive value means the current observation exceeds its same-season counterpart, a negative value means it is lower, and zero means the two values are equal. The sign alone does not indicate a percentage change or establish a persistent trend.

The ordinary first difference and seasonal difference answer different questions:

Transformation	Formula	Comparison
Ordinary first difference	$Y_t-Y_{t-1}$	Current observation versus the immediately preceding observation
Seasonal difference	$Y_t-Y_{t-s}$	Current observation versus the same season in the preceding cycle

The two differences need not have the same sign or magnitude. A series can decline from one month to the next while remaining above the value for that month a year earlier. Always identify the lag in the question before subtracting.

Worked Example: A Short Quarterly Series and Two Different Comparisons

A hypothetical quarterly operating index has seven consecutive values, beginning in the first quarter of Year 1: $72,84,96,90,78,89,103$. Calculate all available annual seasonal differences, compare the last one with its ordinary first difference, and state how many seasonal differences are unavailable.

Method. Quarterly annual seasonality has $s=4$. The series has $n=7$, so the number of seasonal differences is $\max(7-4,0)=3$, attached to observations 5, 6, and 7.

At observation 5, $78-72=6$. At observation 6, $89-84=5$. At observation 7, $103-96=7$. The three seasonal differences are therefore $6,5,7$.

The ordinary first difference at observation 7 is $103-89=14$, because observation 6 is the immediately preceding quarter. Thus the last seasonal difference is 7, while the adjacent-quarter difference is 14. The first four observations have no seasonal differences computed from the supplied series. The count check gives $\max(7-4,0)=3$, as calculated.

Common mistake avoided: Using $103-89$ as the seasonal difference confuses the one-quarter lag with the four-quarter lag. The correct same-quarter comparison is $103-96$.

Common Exam Traps

Using the wrong lag. Annual seasonality means 12 for monthly observations and 4 for quarterly observations; it does not mean subtracting the previous observation.
Pairing different seasons. Check that $t$ and $t-s$ represent the same month or quarter before calculating.
Reversing the subtraction. The convention is current minus prior same-season value, $Y_t-Y_{t-s}$. Reversing it changes the sign and the interpretation.
Calling a difference a percentage. A value of 8 is an increase of 8 original units. A percentage change would require a separate calculation and a stated denominator.
Reporting a negative number of differences. Use $n-s$ only when $n\ge s$; in general use $\max(n-s,0)$.
Assuming the transformation proves a model property. Seasonal differencing cancels a stable seasonal component in the additive setup shown above, but the calculation alone does not prove that the transformed series is stationary.

Key takeaway: Choose the seasonal period $s$, calculate $Y_t-Y_{t-s}$, and interpret the result as the current observation's absolute change from the same season in the preceding cycle. Verify that the earlier value exists and that the number of results is $\max(n-s,0)$ when no outside observations are available.

Check Your Understanding

For each problem, use the stated seasonal period and show that the comparison is with the correct earlier observation.

A monthly index has $s=12$, with $Y_2=46$ and $Y_{14}=51$. Find the seasonal difference at month 14 and interpret its units as index points. (Answer: $\nabla_{12}Y_{14}=51-46=5$ index points.)
A quarterly series has consecutive observations $30,34,41,38,33,39$, beginning in quarter 1 of Year 1. Find every annual seasonal difference and state the number of calculated differences. (Answer: $33-30=3$ and $39-34=5$; there are $6-4=2$ differences.)
A monthly series contains 9 observations and no earlier values. Its seasonal period is 12. How many seasonal differences can be calculated, and why? (Answer: $\max(9-12,0)=0$; no same-month observation from the previous year is supplied.)
Quarterly values for observations 1 through 6 are $50,63,72,68,54,66$, with $s=4$. Calculate the seasonal difference and the ordinary first difference at observation 6. (Answer: Seasonal $=66-63=3$; ordinary $=66-54=12$.)
A monthly series has 15 observations, with no outside data, and $s=12$. State the number of seasonal differences and identify the observations at which they are calculated. (Answer: $15-12=3$ differences, at observations 13, 14, and 15.)

Transformation	Formula	Comparison
Ordinary first difference	\(Y_t-Y_{t-1}\)	Current observation versus the immediately preceding observation
Seasonal difference	\(Y_t-Y_{t-s}\)	Current observation versus the same season in the preceding cycle

What You'll Learn