First Differencing: Algebra and Interpretation

Why First Differencing Matters

A time series records observations in time order, so the change from one observation to the next often has a direct interpretation: a monthly increase, a quarterly change, or a year-to-year movement. First differencing formalizes this comparison. It replaces each level with its change from the preceding period, while its algebraic form makes clear exactly which observations are being compared.

This operation is useful when a series has a changing level, a linear trend, or the structure of a random walk. It does not mean that every differenced series is automatically stationary or otherwise suitable for a particular model. The interpretation depends on the original series and on the time intervals represented by its observations.

Definition: Let $X_t$ be a series observed at consecutive, equally spaced times, and let $B$ denote the backshift operator, defined by $BX_t=X_{t-1}$. The first difference of the series at time $t$ is $$ \Delta X_t=X_t-X_{t-1}=(1-B)X_t. $$ Thus, $1-B$ means “the current observation minus the previous observation.” First differences are defined only when both observations are available.

The first difference is measured in the same units as $X_t$, but it describes a change over one observation interval. For example, if $X_t$ is a monthly dollar amount, $\Delta X_t$ is the change in dollars from the previous month. If observation times are not equally spaced, a difference is still a change, but changes over intervals of unequal duration are not directly comparable as per-period rates.

Basic Algebra and the Endpoint Check

The operator $1-B$ is linear. If $a$ and $c$ are constants, then $$ \Delta(aX_t+cY_t)=a\Delta X_t+c\Delta Y_t. $$ This follows by subtracting the preceding value from the current value: $$ aX_t+cY_t-(aX_{t-1}+cY_{t-1}) =a(X_t-X_{t-1})+c(Y_t-Y_{t-1}). $$ A constant level disappears under differencing because its current and preceding values are equal: $\Delta c=c-c=0$.

A useful check is that consecutive first differences add to the net change over the full interval. For observations $X_0,\ldots,X_n$, expanding the sum gives $$ \sum_{t=1}^{n}\Delta X_t =(X_1-X_0)+(X_2-X_1)+\cdots+(X_n-X_{n-1}) =X_n-X_0. $$ Every intermediate observation appears once with a positive sign and once with a negative sign, so those terms cancel. This telescoping identity checks arithmetic; it does not say that individual changes must be alike.

Worked Example: Monthly Changes and the Endpoint Check

A hypothetical insurer tracks a monthly claims-cost index. Its values, in index points, are $X_0=1{,}000$, $X_1=1{,}025$, $X_2=1{,}037$, $X_3=1{,}057$, and $X_4=1{,}090$. Calculate the first differences, interpret them, and check their sum against the endpoint change. Treat these as fixed observations; no probability distribution is assumed.

Method. Apply $\Delta X_t=X_t-X_{t-1}$ for each month after month 0.

For month 1, $\Delta X_1=1{,}025-1{,}000=25$ index points.

For month 2, $\Delta X_2=1{,}037-1{,}025=12$ index points.

For month 3, $\Delta X_3=1{,}057-1{,}037=20$ index points.

For month 4, $\Delta X_4=1{,}090-1{,}057=33$ index points.

The successive changes are therefore $25,12,20,33$. Their sum is $25+12+20+33=90$ index points. The endpoint change is $X_4-X_0=1{,}090-1{,}000=90$ index points, so the endpoint check agrees. The average change over the four monthly intervals is $90/4=22.5$ index points per month.

Interpretation and mistake avoided. The first difference is the change during a particular month, not the index level at that month. Subtracting in the reverse order would make every change negative and would not follow the defined operator $1-B$.

Effect on a Linear Trend

A series with a deterministic linear trend has the form $X_t=a+bt$, where $a$ is its level at $t=0$ and $b$ is its change per time period. Applying the definition gives $$ \Delta X_t=(a+bt)-\{a+b(t-1)\}=b. $$ The fixed level $a$ cancels, and the difference is the slope $b$. This result is exact for every $t$ for which both observations are defined. It does not depend on estimating a model from data.

The result also identifies an important interpretive distinction. In levels, $a+bt$ changes over time when $b\ne 0$. In first differences, the same deterministic linear trend becomes a constant sequence. The difference is not zero unless the slope itself is zero.

Worked Example: Differencing a Deterministic Trend

A hypothetical portfolio index follows the exact deterministic formula $X_t=240+7t$, where $t$ counts quarters and $X_t$ is measured in index points. Calculate the levels for $t=0,1,2,3,4$, difference them, and interpret the result. No random component or probability distribution is assumed.

Step 1: Calculate the levels. Substituting each time index gives $X_0=240+7(0)=240$, $X_1=240+7(1)=247$, $X_2=240+7(2)=254$, $X_3=240+7(3)=261$, and $X_4=240+7(4)=268$.

Step 2: Calculate each first difference. The differences are $\Delta X_1=247-240=7$, $\Delta X_2=254-247=7$, $\Delta X_3=261-254=7$, and $\Delta X_4=268-261=7$ index points per quarter.

Step 3: Check the result algebraically. From the formula, $\Delta X_t=(240+7t)-[240+7(t-1)]=240+7t-240-7t+7=7$. The calculated differences agree with the algebra.

Step 4: Check the total change. The four differences sum to $7+7+7+7=28$, and the endpoint change is $X_4-X_0=268-240=28$. The checks agree. The index rises by 7 points in each quarter; differencing removes the starting level of 240 but retains the slope.

Common mistake avoided. It would be incorrect to conclude that a linear trend differences to zero merely because differencing removes its level. The constant part $a$ cancels, but the slope $b$ remains.

Effect on a Random Walk

A random walk without drift is defined by $X_t=X_{t-1}+\varepsilon_t$, with a specified starting value $X_0$. Here the innovations $\varepsilon_t$ are independent and identically distributed random variables with mean 0 and finite variance $\sigma^2$. If they are normally distributed, the additional assumption is $\varepsilon_t\sim N(0,\sigma^2)$. Subtracting $X_{t-1}$ from both sides gives $$ \Delta X_t=X_t-X_{t-1}=\varepsilon_t. $$ Thus the first differences are exactly the innovations in the random-walk equation. Under the stated assumptions, the differences are independent, have mean 0 and variance $\sigma^2$, and have the same distribution across time. In the normal case, $\Delta X_t\sim N(0,\sigma^2)$.

This identity explains the role of differencing for a random walk: it changes a series described recursively through its previous level into the sequence of changes that drive that level. The difference operation does not make the realized innovations equal; it reveals them one period at a time. A random walk with a fixed drift $c$, defined instead by $X_t=X_{t-1}+c+\varepsilon_t$, has $\Delta X_t=c+\varepsilon_t$. The no-drift and drift cases should not be confused.

Worked Example: Recovering Random-Walk Innovations

A hypothetical index follows the no-drift random walk $X_t=X_{t-1}+\varepsilon_t$, with fixed starting value $X_0=50$. Assume the innovations are independent with $\varepsilon_t\sim N(0,9)$, so their mean is 0 and variance is 9. The realized innovations for four periods are $3,-2,4,1$. Calculate the resulting levels and first differences, and verify their interpretation.

Step 1: Build the levels recursively. Starting from $X_0=50$, $X_1=50+3=53$. Next, $X_2=53+(-2)=51$. Then $X_3=51+4=55$, and $X_4=55+1=56$.

Step 2: Difference the levels. We obtain $\Delta X_1=53-50=3$, $\Delta X_2=51-53=-2$, $\Delta X_3=55-51=4$, and $\Delta X_4=56-55=1$. These match the four innovations, as the model equation requires.

Step 3: Check the total change. The differences sum to $3+(-2)+4+1=6$. The endpoint change is $X_4-X_0=56-50=6$, so the check agrees. The observed average change over these four periods is $6/4=1.5$, which is a finite-sample average and does not contradict the assumed innovation mean of 0.

Interpretation and mistake avoided. The theoretical distribution of each first difference is $N(0,9)$, not $N(0,3)$, because the second parameter in $N(\mu,\sigma^2)$ is the variance. The realized differences need not have an average of exactly zero in a finite sequence.

How to Apply the Operator Reliably

For a finite time series, write the indexing explicitly before differencing. If the available observations are $X_0,\ldots,X_n$, the first difference at $t$ uses $X_t$ and $X_{t-1}$, so there are $n$ first differences. The starting observation $X_0$ has no first difference within this data range unless an earlier value is supplied.

State the operator.
Use $\Delta X_t=X_t-X_{t-1}=(1-B)X_t$. This fixes the subtraction order.

Match observations by time.
For each $t$, subtract the value at the immediately preceding observation time, not the starting value of the entire series.

Interpret the result in context.
Name what changed, the direction of change, and the interval over which the change was measured.

Check the total.
Sum the differences and compare with $X_n-X_0$. A mismatch signals an arithmetic or indexing error.

A frequent error is to report $X_t$ itself as the first difference, or to calculate $X_{t-1}-X_t$. For example, if $X_{t-1}=80$ and $X_t=86$, the first difference is $86-80=6$, not 86 and not $-6$. Another error is to interpret a difference as a percentage change. The defined quantity $X_t-X_{t-1}$ is an absolute change; a percentage change would require a separate calculation, such as $(X_t-X_{t-1})/X_{t-1}$, and is not the first-difference operator.

The main patterns can be summarized as follows. A constant series differences to zero; a deterministic linear trend differences to its constant slope; and a random walk without drift differences to its innovation. These statements follow from the defining equations, but they apply to those structures specifically. They do not imply that differencing any observed series will produce independent values or remove every form of time variation.

Key takeaway: First differencing is the operator $1-B$, so $\Delta X_t=X_t-X_{t-1}$. Its algebra exposes the period-to-period change: a fixed level cancels, a linear slope remains as a constant difference, and in a random walk the difference is the innovation.

Check Your Understanding

For each problem, apply the stated definition and check that your interpretation matches the indexing and model assumptions.

A hypothetical reserve index has values $X_0=620$, $X_1=635$, and $X_2=629$. Find the two first differences and verify their sum using the endpoints. (Answer: $15$ and $-6$; their sum and the endpoint change are both $9$.)
A series is $X_t=90-4t$, with $t$ measured in years. Find $\Delta X_t$, state its units if $X_t$ is measured in thousands of dollars, and explain the sign. (Answer: $-4$ thousand dollars per year; the level decreases by 4 thousand dollars each year.)
For $t=0,1,2,3$, a sequence has levels $12,16,15,21$. Calculate its first differences and the average change over the three intervals. (Answer: $4,-1,6$; average change $3$ units per interval.)
A random walk has $X_0=100$ and realized innovations $-3,5,2$. Write the levels through $X_3$, then give the differences and their sum. Assume a no-drift model; no distributional assumption is needed for this realized path. (Answer: $X_1=97$, $X_2=102$, $X_3=104$; differences $-3,5,2$, sum $4=X_3-X_0$.)
A series follows $X_t=X_{t-1}+2+\varepsilon_t$, where the $\varepsilon_t$ are independent with mean 0 and variance 16. Identify the first-difference mean and variance, and state its distribution if $\varepsilon_t\sim N(0,16)$. (Answer: $\Delta X_t=2+\varepsilon_t$, with mean $2$ and variance $16$; under the normal assumption, $\Delta X_t\sim N(2,16)$.)

What You'll Learn