Why Differencing a Random Walk Matters
A random walk can move substantially over time, even when each new movement is generated by a simple, stable process. The key is to separate the level of the series from its period-to-period change. First differencing does exactly that: it subtracts the previous level from the current one. For a random walk, this subtraction recovers the increment that was added at that time.
The identity itself is algebraic; it does not require the increments to be independent or normally distributed. The claim that the differences are white noise is different. That claim depends on the increment process having the appropriate mean, variance, and covariance properties. Keeping these two points separate is essential when interpreting a differenced series.
The Differencing Identity
Substitute the random-walk equation into the definition of the first difference:
Thus, at every time \(t\), the first difference equals the increment added at that time. This conclusion holds for any sequence of increments, including increments that are dependent, have a nonzero mean, or have changing variance. Those properties affect what kind of process the differences form, but not the identity.
The reverse calculation also follows directly. If the initial level \(X_0\) and the differences are known, then each later level is obtained by adding the differences cumulatively:
This is why differencing removes the accumulated level while retaining the one-step movements. It does not discard the increments; it expresses them directly.
Worked Example: Recovering Increments From Observed Levels
A hypothetical index is observed at five consecutive times, including time 0, with levels \(X_0=18\), \(X_1=21\), \(X_2=19\), \(X_3=24\), and \(X_4=23\). Calculate each first difference, identify the corresponding increments, and check by rebuilding the levels.
Method. For each time after time 0, subtract the preceding level: \(\Delta X_t=X_t-X_{t-1}\).
At time 1, \(\Delta X_1=21-18=3\). At time 2, \(\Delta X_2=19-21=-2\). At time 3, \(\Delta X_3=24-19=5\). At time 4, \(\Delta X_4=23-24=-1\). The differences, and therefore the increments in a random-walk representation, are \(3,-2,5,-1\).
Check by reconstruction. Starting at \(X_0=18\), add the first difference: \(18+3=21\). Add the second: \(21+(-2)=19\). Then \(19+5=24\), and finally \(24+(-1)=23\). Each reconstructed level matches the stated level.
This calculation establishes the differencing identity for the observed path. It does not establish that the increments are white noise; a finite list of observed values alone does not verify the process assumptions.
When Are the Differences White Noise?
For this tutorial, weak white noise means a sequence with mean zero, finite and constant variance, and zero covariance between values at distinct times. These conditions concern the probability model for the sequence, not merely the values in one observed sample. Independence is sufficient for zero covariance when second moments exist, but independence is not required for weak white noise.
The justification follows from the identity \(\Delta X_t=\varepsilon_t\). In particular, \(E[\Delta X_t]=E[\varepsilon_t]=0\) and \(\operatorname{Var}(\Delta X_t)=\operatorname{Var}(\varepsilon_t)=\sigma^2\). For distinct times \(t\) and \(s\), \(\operatorname{Cov}(\Delta X_t,\Delta X_s)=\operatorname{Cov}(\varepsilon_t,\varepsilon_s)=0\). Each required white-noise property is therefore inherited directly from the increments.
Conversely, if the differences are weak white noise, then the increments are weak white noise because they are the same sequence. The random-walk equation alone does not guarantee that result. It only says that each level is the previous level plus an increment.
The increments need not be normally distributed for this weak-white-noise conclusion. Nor does zero covariance imply independence in general. If a model specifically assumes independent, identically distributed increments, that assumption gives more than weak white noise; it should not be silently inferred from the weaker mean, variance, and covariance conditions.
Worked Example: A Random Walk With Drift
A hypothetical monthly index follows \(X_t=X_{t-1}+2+\varepsilon_t\), with \(X_0=10\). Suppose the random terms \(\varepsilon_t\) are independent and identically distributed as \(N(0,4)\), where the second parameter is the variance. For one realized path, the four random terms are \(-1,0,3,-2\). Calculate the levels and differences, then decide whether the differences themselves are zero-mean white noise.
Method. The change is \(X_t-X_{t-1}=2+\varepsilon_t\), so the drift of 2 must be included along with each random term.
The four differences are \(2+(-1)=1\), \(2+0=2\), \(2+3=5\), and \(2+(-2)=0\). Starting from 10, the successive levels are \(X_1=10+1=11\), \(X_2=11+2=13\), \(X_3=13+5=18\), and \(X_4=18+0=18\).
Check the model-based conclusion. At any time, \(\Delta X_t=2+\varepsilon_t\), so \(E[\Delta X_t]=2+0=2\) and \(\operatorname{Var}(\Delta X_t)=4\). The differences are independent across times under the stated model, but they do not have mean zero. Therefore, they are not zero-mean white noise under the definition above. After subtracting the drift, \(\Delta X_t-2=\varepsilon_t\), the resulting sequence is independent and identically distributed \(N(0,4)\).
The realized differences \(1,2,5,0\) need not average exactly to 2. The mean and variance calculations describe the assumed distribution across repeated realizations, not a requirement that a short observed sequence have its theoretical mean.
What Can Prevent the Differences From Being White Noise?
The direct equality \(\Delta X_t=\varepsilon_t\) makes the assumption check straightforward: inspect the properties required of the increments. A nonzero mean means the differences are not zero-mean white noise, even if their variance is constant and they are uncorrelated over time. Dependence between increments can also rule out white noise if it creates nonzero covariance at distinct times. Changing increment variance violates the constant-variance condition.
A random walk with drift is often written \(X_t=X_{t-1}+c+\varepsilon_t\), where \(c\) is a constant and \(\varepsilon_t\) has mean zero. Its differences are \(c+\varepsilon_t\). They have constant mean \(c\); subtracting \(c\) gives the zero-mean innovations \(\varepsilon_t\). It is important to say whether a claim concerns the raw differences or the differences after centering.
Worked Example: Dependent Increments Are Not White Noise
Consider a zero-drift random walk with increments \(\varepsilon_t=Z+\eta_t\), where \(Z\) has distribution \(N(0,1)\), the \(\eta_t\) are independent and identically distributed \(N(0,1)\), and \(Z\) is independent of every \(\eta_t\). Find the mean and variance of a difference and the covariance between two differences at distinct times. Decide whether the differences are weak white noise.
Method. The differencing identity gives \(\Delta X_t=\varepsilon_t=Z+\eta_t\). Since both terms have mean zero, \(E[\Delta X_t]=0+0=0\). Independence of \(Z\) and \(\eta_t\) gives \(\operatorname{Var}(\Delta X_t)=1+1=2\).
For \(t\ne s\), expand the covariance:
The other covariance terms are zero: each \(\eta\) is independent of \(Z\), and \(\eta_t\) and \(\eta_s\) are independent when \(t\ne s\). The covariance between distinct differences is therefore 1, not 0. Despite having mean zero and constant variance 2, the differences are not weak white noise.
For a path check, if \(Z=1\) and four realized values of \(\eta_t\) are \(-1,0,2,-2\), the increments and differences are \(0,1,3,-1\). The theoretical covariance calculation, not this particular four-value sequence, is what establishes the dependence in the model.
Common Mistake: Treating the Level as the Increment
A frequent calculation error is to call \(X_t\) itself the increment at time \(t\). For example, in the first worked example, this error would identify the increment at time 1 as \(X_1=21\). That is the level, not the change. The correct increment is \(X_1-X_0=21-18=3\). At time 2, the correct difference is \(19-21=-2\), not the level \(X_2=19\).
A related mistake is to conclude that differences are white noise merely because they equal increments. Equality identifies which sequence to assess; it does not establish the sequence's mean, variance, or dependence properties. Check those conditions separately. These distinctions are aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.
Check Your Understanding
For each question, distinguish the arithmetic of differencing from the assumptions about the increment process.
- A series has levels \(X_0=7\), \(X_1=10\), \(X_2=8\), \(X_3=8\), and \(X_4=13\). Find the four first differences. (Answer: \(3,-2,0,5\).)
- A random walk starts at \(X_0=4\) and has increments \(1,-3,2,0\). Reconstruct \(X_1\) through \(X_4\). (Answer: \(X_1=5,\ X_2=2,\ X_3=4,\ X_4=4\).)
- A zero-drift random walk has independent increments, each distributed as \(N(0,9)\). State the distribution of a difference and whether the differences are weak white noise. (Answer: Each difference is \(N(0,9)\); yes, they are independent, mean-zero white noise with variance 9.)
- A random walk has increments with mean 1, variance 4, and zero covariance between distinct times. Are its raw differences zero-mean white noise? What if 1 is subtracted from every difference? (Answer: No; the raw differences have mean 1. After subtracting 1, the centered differences have mean 0, variance 4, and zero covariance across times, so they are weak white noise.)
- A zero-drift random walk has mean-zero increments with variance 2, but the covariance between increments at any two distinct times is 0.5. Are its first differences weak white noise? (Answer: No; the covariance between distinct differences is also 0.5, not zero.)