ETC3450 Chap.3 White Noise, Random Walks and Differencing

White Noise, Random Walks and Differencing

Under the standard white-noise conditions, past values do not alter the conditional mean, which remains zero. The process is stationary and provides the innovation building block for richer models. Uncorrelated does not always mean independent unless stronger distributional assumptions are supplied. The next innovation cannot be forecast from its history, yet its mean and variance are specified.

Prediction therefore includes a point centre and a distribution of error rather than a claim that nothing is known. A sample ACF may place one spike outside an approximate band by chance. Declaring serial dependence from one isolated sample coefficient ignores multiple comparisons and sampling variation. Check the residual plot and ACF as a pattern.

Systematic clusters, decaying correlations or changing spread challenge white noise more strongly than a lone fluctuation. Model comparison should not reduce to selecting the largest fit statistic. Ask whether white noise leaves serial structure in the residuals, whether innovation is parsimonious enough for the sample, and whether parameters remain stable over the forecast origin.

Interpret serial signal against a substantive dynamic mechanism. If rival models perform similarly, report that uncertainty and prefer a specification whose assumptions can be checked transparently. In a random walk, today's value equals yesterday's value plus a new innovation. Iterating the recursion expresses the level as the initial value plus the sum of all past shocks.

Expected innovations cancel in the mean, but their variances accumulate with time, violating constant unconditional variance. A shock changes every later level forecast because there is no force pulling the series back to a fixed mean. This differs from a stationary autoregression, where the effect decays when the coefficient lies inside the stability boundary.

Simulated paths spread farther apart at later horizons even though their average remains near the starting point. Looking only at the ensemble mean would miss the defining growth in uncertainty. Expand the recursion for several periods and inspect the coefficient on each innovation. Then compute the variance using the independence assumptions. The time index in the result is the stationarity warning.

Econometric interpretation begins by fixing the stochastic object behind random walk: its conditioning information, time index and maintained assumptions. Derive the implication for accumulated shock before inserting estimated values, because a numerical answer can conceal a wrong model. Diagnostics then ask whether residual behaviour is compatible with the specification.

Treat growing variance as a decision boundary: if it fails, revise the process or inference rather than decorating the same equation. Express random walk for a generic time index before looking at one observed path. Subtracting the previous level from a random walk leaves the innovation. The differenced series can therefore be stationary even when the level is not.

Differencing changes the question from the level of a variable to its period-to-period change and should not be presented as a cosmetic cleaning step. A transformation that is unnecessary may introduce moving-average structure and amplify noise. Diagnosis should compare the level, difference and theoretical model rather than applying differences automatically to every trending plot.

A price index has a persistent upward level, while quarterly changes fluctuate around a stable average. Forecasting changes and reconstructing levels requires accumulating predicted differences and their uncertainty. Label the units after transformation: dollars becomes dollars per period change, while log difference approximates a growth rate under conditions.

Recheck outliers because differencing can place one unusual level change into adjacent comparisons. A time-series calculation should be accompanied by a probability statement. For first difference, separate a realised path from the process that could generate many paths. State which moments are stable and how innovation changes with the lag.

When the result depends on reconstruction, identify the parameter region and initial condition explicitly. The limiting case often reveals whether the claimed long-run interpretation is mathematically coherent. Turn first difference into a diagnostic sequence. Plot the series and transformations, inspect dependence over several lags, estimate a defensible candidate and then examine the residuals for structure left behind.

What this chapter covers

• 01

  White noise is innovation without serial signal

• 02

  A random walk accumulates every shock

• 03

  First differences expose the random-walk innovation

Using residual evidence for White noise is innovation without serial signal

Key terms

White-noise innovations

White noise is innovation without serial signal — Under the standard white-noise conditions, past values do not alter the conditional mean, which remains zero. The process is stationary and provides the innovation building block for richer models. Uncorrelated does not always mean independent unless stronger distributional assumptions are supplied. Check the residual plot and ACF as a pattern. Systematic clusters, decaying correlations or changing spread challenge white noise more strongly than a lone fluctuation.

Random-walk shock accumulation

A random walk accumulates every shock — In a random walk, today's value equals yesterday's value plus a new innovation. Iterating the recursion expresses the level as the initial value plus the sum of all past shocks. Expected innovations cancel in the mean, but their variances accumulate with time, violating constant unconditional variance. Expand the recursion for several periods and inspect the coefficient on each innovation. Then compute the variance using the independence assumptions. The time index in the result is the stationarity warning.

Random-walk first differences

First differences expose the random-walk innovation — Subtracting the previous level from a random walk leaves the innovation. The differenced series can therefore be stationary even when the level is not. Differencing changes the question from the level of a variable to its period-to-period change and should not be presented as a cosmetic cleaning step. Label the units after transformation: dollars becomes dollars per period change, while log difference approximates a growth rate under conditions. Recheck outliers because differencing can place one unusual level change into adjacent comparisons.

White Noise, Random Walks and Differencing FAQ

Exam move

Keep a derivation and diagnostic sheet for White Noise, Random Walks and Differencing. Write every process assumption before calculating a moment, response or forecast. Begin with white noise and reconstruct the reasoning without looking at the worked response. Then change one condition in the example and decide whether innovation still explains the outcome.

Use the chapter questions to compare direct observation with inference, and write the strongest rival account in full. Before closing the chapter, return to reconstruction and state the precise boundary it places on transfer. Check that every conclusion names an observable consequence and that uncertainty is attached to the step it affects.

A final retrieval pass should be fast enough to reproduce the method from headings and diagrams while leaving the detailed prose for checking nuance.

ETC2410 Introductory Econometrics · ETC2420 Statistical Thinking · ETC3460 Financial Econometrics