Giacomini-White Test of Conditional Predictive Ability
Also known as: GW Test, Conditional Predictive Ability Test, Giacomini-White CPA Test, Koşullu Tahmin Yeteneği Testi
The Giacomini-White (GW) test, introduced by Raffaella Giacomini and Halbert White in 2006, evaluates whether two competing forecasting methods have equal conditional predictive ability given information available at the time of forecast. Unlike unconditional tests such as the Diebold-Mariano test, it asks whether one method systematically outperforms the other in specific economic or market conditions, making it especially useful for practitioners who need state-dependent forecast comparisons.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Apply the GW test when comparing two non-nested forecasting methods over a rolling estimation window and when the research question concerns whether one method is systematically better given observable conditioning information. It is appropriate for macroeconomic, financial, and energy forecasting contexts. Key assumptions include a fixed rolling window (not recursive with expanding sample) and correct specification of the conditioning information set. The test is not designed for nested model comparison. Alternatives include the Diebold-Mariano test for unconditional comparison and the Model Confidence Set for multi-model settings.
Strengths & limitations
- Directly tests conditional rather than unconditional predictive ability, capturing state-dependent forecast performance
- Applicable to any loss function and does not require knowledge of the data-generating process
- Accommodates non-nested models and allows for misspecified forecasting methods
- Asymptotically valid under general forms of heteroskedasticity and serial correlation via HAC estimation
- Requires a fixed rolling window, which may discard useful early observations and reduce power in small samples
- Power and size depend critically on the choice of conditioning variables in the test function vector
- Not suitable for comparing nested models without modification, limiting applicability in model selection
- Performance in very short evaluation samples is not well-characterized and may lead to size distortions
Frequently asked
How does the GW test differ from the Diebold-Mariano test?
The Diebold-Mariano test evaluates unconditional equal predictive accuracy—whether, on average across all periods, two forecasts differ. The GW test conditions on information available at forecast time, allowing it to detect whether one method is systematically better in specific economic states. Additionally, GW requires a fixed rolling estimation window to maintain the conditional framework, whereas DM does not impose this constraint.
Why must the estimation window be fixed rather than expanding?
With an expanding window, the estimator uses an increasing fraction of the sample, which changes the asymptotic distribution of the test statistic. Giacomini and White's theoretical results are derived assuming a fixed rolling window of constant length, ensuring that the forecasting procedure—not just the model—is being evaluated and that the chi-squared limiting distribution is valid.
What should I use as the test function vector?
The test function vector should contain variables in the forecaster's information set at forecast origin—for instance, lagged loss differentials, economic indicators, or regime indicators. Setting the vector to a constant scalar reduces the GW test to an unconditional version. Adding economically motivated variables increases power against specific alternatives but risks data snooping if chosen post-hoc.
Sources
- Giacomini, R., & White, H. (2006). Tests of conditional predictive ability. Econometrica, 74(6), 1545–1578. DOI: 10.1111/j.1468-0262.2006.00718.x ↗
How to cite this page
ScholarGate. (2026, June 2). Giacomini-White Test of Conditional Predictive Ability. ScholarGate. https://scholargate.app/en/econometrics/giacomini-white-test
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Diebold-Mariano TestEconometrics↔ compare
- Model Confidence SetEconometrics↔ compare
- Time-Series Cross-ValidationEconometrics↔ compare