Reading the current lesson
Learn
Interactive·

Time-Series Regression

If advertising and revenue rise together, is that really an advertising effect?

Monthly revenue depends not only on current advertising but also on previous revenue, long-run trend, and seasonality. Ignoring time order can leave strong autocorrelation in the residuals.

Time-series regression includes past information and time structure so that a shared trend is not automatically mistaken for a substantive relationship.

Key question

How should past values, trend, and seasonality enter a regression that explains the current outcome?

1

Concept

What makes time-series regression different?

Time-series regression connects the current outcome to present conditions and past values

Dynamic regression example

yₜ = β₀ + β₁xₜ + φyₜ₋₁ + εₜ

TermRevenue example
xₜcurrent advertising spending
yₜ₋₁last month’s revenue
tlong-run time trend
seasonal indicatorsrecurring monthly or quarterly differences

Trend and season

yₜ

A series usually carries a slow drift and a repeating shape at the same time.

Residuals that remember

corr(eₜ, eₜ₋₁)

Errors of the same sign arrive in runs, so ordinary standard errors mislead.

The lagged term

φ yₜ₋₁
yₜ₋₁yₜφxₜ

Last period's outcome enters as a predictor and carries the effect forward.

2

Why It Matters

Why are trends and spurious regression dangerous?

Two unrelated series can produce a high R² simply because both trend over time

This is spurious regression. A high R² and small p-value are not convincing when the series are nonstationary and the residuals retain trend or autocorrelation.

Stationarity

A stationary series has a reasonably stable mean, variance, and autocorrelation structure over time. Plot the series and examine unit-root or stationarity diagnostics before deciding how to model it.

3

How It Works

What do lagged variables represent?

Lags represent delayed effects and persistence

TermInterpretation
yₜ₋₁persistence from the previous outcome
xₜ₋₁a delayed predictor effect
ta deterministic linear trend
month indicatorsseasonality

Persistence

If φ=0.70, a higher value last month tends to carry substantially into the current month, provided the dynamic model is stable.

Too many lags reduce the usable sample and can create multicollinearity, so use information criteria, domain knowledge, and residual diagnostics together.

4

Example

When should we use a trend or a difference?

A trend term and differencing solve different kinds of trend

ProblemTypical response
Deterministic linear trendinclude a time trend
Stochastic trend / unit rootdifference the series
Seasonalityseasonal indicators or seasonal differencing
Residual autocorrelationlag terms or ARIMA errors
Stable long-run relationship among nonstationary seriescointegration and an error-correction model

Interpretation after differencing

A coefficient in a differenced model explains relationships between period-to-period changes, not differences in the original levels.

5

Interactive

Change trend, lag, and seasonality

Change trend, lag, and seasonality

Adjust the time structure and compare the level series, first differences, and lag-1 autocorrelation.

Series form

Level

Current-series ACF(1)

0.83

Observations

36

Time

ACF(1) here is calculated from the displayed series, not from fitted-model residuals.

What to watch

Time-series regression must explain not only relationships among variables but also the remaining structure across time.

Key takeaways

  • Time-series regression respects time order and incorporates lagged information.
  • Shared trends can create spurious regression even with high R² and small p-values.
  • Lagged terms represent persistence and delayed effects.
  • Use trend terms, differencing, seasonality, and residual diagnostics according to the source of time structure.
Regression curriculum complete