Time-series econometrics

Spurious regression, co-integration & forecasting

Theory and evidence from U.S. real disposable income and consumption.

DPIC96 × PCECC961960Q1–2026Q2Five research questions

A rainy view of Big Ben in London
Opening case Hendry's rainfall example motivates the difference between a convincing regression and a credible relationship.

An unexpected relationship

The model looked strong—then Hendry revealed the variable

Historical Bank of England five-pound note Rain in London
Fig. 2 British prices and British rainfall provide the historical setting.
01

UK prices and a variable C showed a close nonlinear relationship.

02

A quadratic regression gave a good fit and significant coefficients.

03

With an AR(1) error model, the fit became “spectacular.”

04

C was cumulative rainfall in the United Kingdom.

This case warns about misleading econometric relationships. It is not automatic evidence of co-integration.

Why ordinary inference can fail

Persistence can turn accidental movement into false confidence

1 · REMEMBERS

Persistent levels

Earlier shocks remain inside later observations.

2 · ALIGNS

Apparent co-movement

Two unrelated paths can move together for many periods.

3 · INFLATES

Misleading tests

Conventional t-tests use the wrong reference behavior.

4 · MISSTATES

False economic story

A high R² becomes an explanation the data do not justify.

Fig. 3 The formula can be familiar while its time-series assumptions fail.

Controlled demonstration · synthetic data

Independent random walks still produce a significant OLS slope

Two independent synthetic random walks Scatter plot and OLS line for independent random walks
Fig. 4 Independent innovations generate β̂ = 0.455, R² = 0.372, and conventional p < 0.001; the residual is not stationary.

Why stationarity matters

Three paths can rise, but they require different treatments

I(0): stable level

Mean and variance stay broadly stable. Shocks fade.

Trend-stationary

A predictable trend changes the mean. Remove the trend.

I(1): stochastic trend

Each shock shifts the future path. Difference once.

Fig. 5 A visible upward path alone does not identify the type of trend.

Random walk → first difference

Differencing removes the accumulated path

Xt = Xt−1 + utEach new shock stays in every later level.
→
ΔXt = utThe first difference contains only the new shock.
Fig. 6 If ΔXt is I(0), then Xt is I(1).

Differencing prevents a naive level regression, but it may remove a genuine long-run relation.

From spurious regression to co-integration

Co-integration asks whether the gap stays stable

Synthetic co-integrated series and their stable spread
Fig. 7 The levels drift, but their estimated long-run gap stays bounded.
ut = Yt − α − βXt

X and Y may each be I(1). They are co-integrated when this linear combination is I(0).

Boundary: statistical stability does not prove causation.

The diagnostic difference

The residual reveals whether co-movement is stable

Persistent residual beside a mean-reverting residual
Fig. 8 A spurious residual wanders; a co-integrating residual returns toward a stable mean.

Long-run equilibrium and short-run correction

Yesterday's gap creates today's adjustment

Last period

Y is 10 units above equilibrium.

+10
×

Adjustment speed

λ = −0.30

−0.30
=

This period

Correction contribution

−3

How to test the relationship

The tests form a sequence, not a checklist

1 · CLASSIFY

ADF

Is each series I(0) or I(1)?

H₀: unit root

2 · TEST GAP

Engle–Granger

Is one estimated residual stationary?

H₀: no co-integration

3 · TEST RANK

Johansen

How many long-run relations exist?

H₀: rank ≤ r

4 · MODEL

ECM / VECM

How does the system correct and forecast?

Estimate after rank choice

How to read the sequence

Each result determines whether the next step is valid

STEP
WHAT WE TEST
WHAT SUPPORT LOOKS LIKE
WHAT IT ALLOWS NEXT
1 · ADF
Each variable

Does the level contain a unit root?

Level: do not reject

First difference: reject

2 · ENGLE–GRANGER
One residual

Does the estimated long-run gap return to a stable mean?

Reject no co-integration

The residual is stationary.

3 · JOHANSEN
The whole system

How many independent stationary combinations exist?

Choose rank r

Trace and max-eigen tests support the same main rank.

4 · VECM
Adjustment and forecasts

How do variables respond after disequilibrium?

Significant adjustment

Diagnostics and out-of-sample errors are acceptable.

Decision guide A failed prerequisite changes the model: the sequence is a chain of decisions, not four unrelated tests.

Step 1 · Reading the ADF test

ADF checks the variables before we test their relationship

Level test

H₀: the series has a unit root
  • p > .05: do not reject the unit-root null.
  • The level may be non-stationary.
  • This is not proof that a unit root exists.

First-difference test

ΔXt = Xt − Xt−1
  • p < .05: reject the unit-root null for the change.
  • Non-stationary level + stationary first difference supports I(1).
  • Co-integration testing is appropriate when the variables have compatible integration orders.

Memory aid: ADF checks each variable. It does not, by itself, show that two variables share a long-run equilibrium.

Why VECM comes last

The model is valid only after two empirical gates are passed

GATE 1 · INTEGRATION

Are the variables I(1)?

ADF evidence should support non-stationary levels and stationary first differences.

→
GATE 2 · CO-INTEGRATION

Is the long-run gap stable?

Engle–Granger tests the residual; Johansen identifies the number of stable combinations.

→
MODEL · VECM

Short run + correction

The VECM combines recent changes with the previous equilibrium error. The selected rank fixes its long-run structure.

Model gate Without I(1) variables and co-integration evidence, an error-correction term has no justified long-run interpretation.

Step 2 · Engle–Granger in practice

Co-integration requires a stationary long-run residual

1 · ESTIMATE

Long-run equation

Ct = α + βYt + ut

OLS estimates the level relationship between consumption and income.

2 · EXTRACT

Equilibrium error

ût = Ct − α̂ − β̂Yt

The residual measures how far the variables are from the estimated long-run relation.

3 · TEST

Residual stationarity

H₀: no co-integration

Use residual-based critical values—not the ordinary ADF table.

Decision rule: a unit-root residual can drift indefinitely → no co-integration. A stationary residual is bounded and mean-reverting → evidence of co-integration.

Empirical design · FRED / U.S. BEA

Income and consumption share persistent long-run growth

Indexed U.S. real disposable income and real consumption with the forecast test start marked
Fig. 12 The two quarterly series rise together, but formal tests must decide whether their gap is stable.
266QUARTERS
1960Q1–2026Q2
258TRAINING
THROUGH 2024Q2
8TEST QUARTERS
2024Q3–2026Q2

We use natural logs. Every unit-root test, lag choice, co-integration test, and model estimate uses the training sample only.

Empirical evidence · Step 1

Both variables behave like I(1) processes

Log income · level

p = .273

ADF with trend does not reject a unit root.

H₀: UNIT ROOT

Δ log income

p < .001

The first difference rejects a unit root.

RESULT: STATIONARY CHANGE

Log consumption · level

p = .800

ADF with trend does not reject a unit root.

H₀: UNIT ROOT

Δ log consumption

p < .001

The first difference rejects a unit root.

RESULT: STATIONARY CHANGE
Fig. 13 Non-stationary levels and stationary first differences support the I(1) classification required for standard co-integration tests.

Empirical evidence · Steps 2 and 3

Two main tests identify one long-run relation

Engle–Granger · intercept

p = .00036

Reject the null of no co-integration.

H₀: NO CO-INTEGRATION

Engle–Granger · trend

p = .00473

The conclusion survives a trend specification.

H₀: NO CO-INTEGRATION

Johansen trace

rank = 1

Reject rank zero; do not reject rank at most one.

VAR LAG 4 · BIC

Maximum eigenvalue

rank = 1

The second rank test reaches the same main conclusion.

5% CRITICAL VALUES
Fig. 14 Engle–Granger and the prespecified Johansen tests agree, although Johansen rank changes under some alternative specifications.

Empirical evidence · Error correction

Income carries the significant adjustment toward equilibrium

Estimated income-consumption equilibrium error returning around zero
Fig. 15 The estimated equilibrium error is bounded around zero in the training sample.
Income adjustmentα = −0.1616
p < .001
Consumption adjustmentα = +0.0445
p = .081
Residual dynamicsWHITENESS p = .540
NO STRONG AUTOCORRELATION

Under this normalization, income responds significantly to the previous equilibrium error. Consumption adjustment is not significant at 5%.

Empirical evidence · Forecasting

Test whether equilibrium information improves forecasts

1 · ESTIMATEFit the rank-1 VECM on 258 training quarters.
2 · FORECASTUse short-run changes + the previous equilibrium error.
3 · BENCHMARKForecast the same 8 quarters with four competing models.
4 · EVALUATECompare RMSE and MAE; lower error means better accuracy.
Actual income and consumption compared with VECM, differenced VAR, and ARIMA forecasts
Fig. 16 Fixed-origin forecasts use the same 2024Q3–2026Q2 test period for every model.
Income · VECMRMSE 225.8 · MAE 196.0
41.6% lower RMSE than drift
Consumption · VECMRMSE 114.3 · MAE 96.5
6.7% lower RMSE than drift
ResultVECM HAS THE LOWEST RMSE AND MAE FOR BOTH VARIABLES
SHORT 8-QUARTER HOLDOUT LIMITS GENERALIZATION

Answers to the five research questions

A stable residual turns shared growth into useful information

Q1

Why spurious?

Persistent unrelated trends can create significant OLS results.

Q2

What is co-integration?

I(1) variables have a stationary long-run combination.

Q3

Why equilibrium?

The deviation stays bounded and enters the next period's adjustment.

Q4

How tested?

ADF classifies the series; Engle–Granger and Johansen test stability.

Q5

How forecast?

VECM combines short-run changes with the previous equilibrium error.

Conclusion The income–consumption evidence supports one long-run relation and useful VECM forecasts, but the eight-quarter test window limits generalization.

Time-series econometrics

Spurious regression
or co-integration?

Shared movement is not enough. The residual must stay stable for the long-run relationship to be credible.

DPIC96 × PCECC96FRED data10–15 minutes

Two rising series with an unstable gap Two paths rise together, but the distance between them changes over time. gap changes Both series rise but the distance may drift time →
Fig. 1 Co-movement in levels does not prove a stable equilibrium.

The research route

Five questions move from illusion to forecasting

01

Spurious regression

Why can unrelated trends look significant?

02

Co-integration

When is a level relationship meaningful?

03

Equilibrium correction

How do short-run gaps close?

04

Statistical tests

What do ADF, Engle–Granger, and Johansen test?

05

Forecasting

Does the equilibrium improve predictions?

The logic is sequential: test persistence before trusting levels, then test equilibrium before using a VECM.

THEORY → TESTS → EVIDENCE → FORECASTS

Question 1 · Spurious regression

Persistence creates false confidence

Mechanism of spurious regression Persistent shocks create accidental co-movement, which makes ordinary p-values misleading. Persistent shockstoday changes tomorrowand many later periods allows Accidental alignmentindependent paths candrift in one direction inflates t-statisticlooks largep looks small misleads False economic storythe usual reference distributionno longer applies two unrelated random walks
Fig. 2 The regression formula looks ordinary; its sampling logic is not.
Yt = α + βXt + εt
01

Observations are strongly dependent across time.

02

OLS mistakes shared drift for repeated evidence.

03

The usual t-test no longer has its usual meaning.

A small p-value can describe a statistical artifact, not an economic relationship.