Granger Causality Testing – Full Guide With Python Examples

The Granger causality test is a statistical hypothesis test used to determine whether one time series can predict another. Developed by the economist Clive Granger in 1969, it has become a standard tool in econometrics, neuroscience, finance, and signal processing. Unlike ordinary correlation, Granger causality explicitly tests for directional predictive relationships by examining whether past values of one variable improve the forecast of another variable beyond what the target variable’s own past can provide.

The core logic is straightforward: if adding lagged observations of variable X significantly improves the prediction of variable Y, then X is said to Granger-cause Y. This does not imply true causation in the philosophical sense—only that X carries useful information for forecasting Y. Researchers use this test as a diagnostic check for temporal precedence and predictive usefulness.

Over the decades, the test has been extended to multivariate settings, non-linear relationships, and high-dimensional datasets. Its implementation in widely used statistical packages such as Python’s statsmodels, R’s lmtest, and MATLAB has made it accessible to analysts across many disciplines. For those new to time-series methods, a broader overview of research approaches can be found in this Qualitative vs Quantitative Research Guide.

What Is Granger Causality Test Used For?

• Granger causality does not imply true causation; it only tests predictive usefulness.
• Stationarity of time series is a critical assumption; differencing may be needed.
• Lag length selection heavily influences test results; use information criteria (AIC/BIC).
• The test can be extended to multivariate and non-linear contexts.
• Common misinterpretation: significant p-value does not mean one variable ’causes’ another in a philosophical sense.

Understanding the Concept of Granger Causality

Granger causality rests on a simple premise: if past values of X help predict Y after accounting for Y’s own history, then X Granger-causes Y. This is evaluated by comparing a restricted model (Y predicted from its own lags only) against an unrestricted model (Y predicted from its own lags plus lags of X). If the unrestricted model yields a significantly better fit, the null hypothesis is rejected.

The mathematical framework typically involves vector autoregressive (VAR) models. For two time series, the unrestricted model can be written as a bivariate VAR where each variable is regressed on its own lags and the lags of the other variable. The significance of the cross-lagged terms is then assessed using an F-test or a likelihood ratio test. A detailed formal definition is available on Wikipedia.

Key Applications of Granger Causality

Economists use Granger causality to test whether money supply predicts inflation, whether consumer sentiment forecasts spending, or whether asset prices lead economic activity. In neuroscience, the test helps map functional connectivity between brain regions by analyzing whether activity in one region predicts activity in another. Climate scientists apply it to examine relationships between ocean temperatures and atmospheric pressure patterns.

Financial analysts often employ the test to study lead-lag relationships between stock indices, exchange rates, and commodity prices. For instance, testing whether gold price movements predict currency fluctuations is a common application—a topic explored in this Gold Price AUD analysis. The test also appears in signal processing, where it helps identify causal flows in multivariate sensor data.

How to Interpret Granger Causality Test Results?

Understanding P-Values and F-Statistics

The primary output of a Granger causality test is a p-value and an F-statistic for each lag length tested. The p-value tells you whether the observed improvement in prediction is statistically significant. A p-value below the chosen threshold (commonly 0.05) leads to rejection of the null hypothesis, meaning the predictor variable provides statistically significant predictive information.

The F-statistic measures the relative improvement in fit between the restricted and unrestricted models. A larger F-statistic indicates stronger evidence against the null. Most software outputs include several test variants—SSR-based F-test, chi-square test, and likelihood ratio test—but the p-values from these variants usually lead to the same conclusion.

Common Pitfalls in Interpretation

One frequent mistake is interpreting Granger causality as evidence of a direct causal mechanism. The test cannot account for latent confounders—a third variable driving both series can produce a significant result even when no direct relationship exists. For example, rainfall might Granger-cause both umbrella sales and ice cream sales without a direct link between the two.

Another pitfall involves ignoring bidirectional causality. X may Granger-cause Y, Y may Granger-cause X, or both. Researchers should test both directions and report results transparently. Seasonality, trending behavior, and structural breaks can also distort inference if not properly addressed. A practical guide with Python examples is provided by Machine Learning Plus.

What Are the Assumptions of Granger Causality Test?

Stationarity and Lag Selection

Stationarity is a fundamental requirement. If the time series contain trends or seasonal patterns, the test may produce spurious results. Researchers typically apply differencing, detrending, or seasonal adjustment before testing. Unit root tests such as the augmented Dickey-Fuller test help determine whether transformations are needed.

Lag selection is equally critical. Choosing too few lags may omit relevant predictive information, while too many lags wastes degrees of freedom and can inflate standard errors. Information criteria such as Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), or Hannan-Quinn Criterion (HQIC) provide systematic guidance. Domain knowledge and residual diagnostics also play a role.

Limitations and When Not to Use Granger Causality

Granger causality assumes a linear model structure. If the true relationship between variables is non-linear, the standard test may fail to detect predictive links. Non-linear extensions—such as kernel-based Granger causality or transfer entropy—are available but require additional methodological care.

The test also assumes no severe multicollinearity among lagged predictors and requires a sufficient sample size. Small datasets produce unstable estimates. When omitted variables are likely, or when temporal ordering is ambiguous, alternative causal discovery methods such as directed acyclic graphs (DAGs) or the PC algorithm may be more appropriate.

How to Perform Granger Causality Test in Python (with Example)?

Step-by-Step Implementation with statsmodels

The most common Python library for Granger causality testing is statsmodels, which provides the grangercausalitytests function. The function takes a DataFrame where the first column is the target variable and the second column is the predictor variable, along with a maximum lag length.

Below is a basic example that simulates two related time series and tests whether x Granger-causes y. The data is structured so that y depends on lagged values of x plus noise.

from statsmodels.tsa.stattools import grangercausalitytests
import pandas as pd
import numpy as np

np.random.seed(0)
n = 200
x = np.random.randn(n).cumsum()
y = np.roll(x, 1) + np.random.randn(n) * 0.5

df = pd.DataFrame({'y': y, 'x': x}).dropna()
results = grangercausalitytests(df[['y', 'x']], maxlag=4, verbose=True)

The function prints test statistics and p-values for each lag. For a more flexible multivariate approach, the VAR class from statsmodels allows testing multiple predictors simultaneously and supports lag selection via information criteria. A detailed tutorial with expanded code is available on Aptech.

Interpreting the Python Output

For each lag, the output includes several test statistics. The most commonly reported are the F-test p-value and the chi-square test p-value. If the p-value is below 0.05, the null hypothesis (no Granger causality) is rejected. The test statistic itself indicates the magnitude of the improvement in model fit.

In the example above, because y was constructed to depend on lagged x, the p-values should be small for lags where the dependency exists. A Medium article with a worked example demonstrates the full workflow from data preparation to interpretation (link).

How Has the Granger Causality Test Evolved Over Time?

1. 1969 — Clive Granger introduces the concept in the paper “Investigating Causal Relations by Econometric Models and Cross-Spectral Methods” (JSTOR).
2. 2003 — Granger is awarded the Nobel Memorial Prize in Economic Sciences (shared with Robert Engle).
3. 2000s — Extensions to non-linear Granger causality emerge, including kernel-based and information-theoretic approaches.
4. 2010s — Widespread adoption in neuroscience for functional connectivity analysis, in climate science for studying Earth system interactions, and in finance for lead-lag detection.
5. 2021 — Shojaie & Fox publish a comprehensive review (PMC: 10571505) covering modern advancements, including high-dimensional and non-parametric methods.

What Is Known and What Remains Uncertain About Granger Causality?

What Is the Role of Granger Causality in Modern Data Analysis?

The Granger causality test remains a foundational tool in econometrics and time-series analysis. Its simplicity and interpretability make it widely used, but practitioners must be aware of its limitations. The test’s reliance on linear autoregressive models means it may miss non-linear dependencies unless extended.

In the age of big data, more advanced causal inference methods—such as directed acyclic graphs, do-calculus, and structural causal models—sometimes overshadow Granger causality. However, the test remains a quick and effective diagnostic check for temporal precedence. Its low computational cost and straightforward implementation make it an attractive first step in exploratory analysis.

Modern implementations in Python and R have lowered the barrier to entry, increasing the need for clear guidance on interpretation and assumptions. As datasets grow larger and more complex, extensions to high-dimensional and non-parametric settings continue to expand the test’s relevance.

What Are the Key Sources on Granger Causality?

The foundational source is Clive Granger’s 1969 paper, which first articulated the concept in the context of econometric modeling. The Nobel Prize committee’s recognition in 2003 cemented the test’s importance in economic analysis. A comprehensive review by Shojaie & Fox (2021), published in an NIH-indexed journal, covers modern developments and applications across multiple disciplines (NIH/PMC).

“The definition of causality is based entirely on the predictability of some series.” — Clive Granger (1969)

“Granger causality has become a popular tool for analyzing time series data in many application domains, from economics to neuroscience.” — Shojaie & Fox (2021)

“The Granger causality test is a statistical hypothesis test for determining whether one time series is useful in forecasting another.” — Wikipedia

What Should You Remember About Granger Causality Testing?

Granger causality is a forecasting-based statistical test that checks whether lagged values of one time series improve prediction of another. It is widely used in econometrics, neuroscience, finance, and signal processing, but it requires careful attention to stationarity, lag selection, and interpretation. The test measures predictive utility, not philosophical causation, and should be applied with a clear understanding of its assumptions and limitations.

Frequently Asked Questions