Newey-West t-statistic
Education only · our voice · free public data
Definition
A t-statistic computed with a HAC (heteroskedasticity- and autocorrelation-consistent) standard error, so significance tests on mean returns or regression coefficients are not fooled by serial correlation or changing variance in the residuals.
How to read it
Ordinary t-stats assume returns are independent with constant variance. Trading returns rarely are: trends, overlapping holding periods, and illiquid marks all induce autocorrelation that makes the naive standard error too small and the t-stat too big. Newey-West widens the standard error to account for correlation out to a chosen lag L, giving an honest t-stat. If a signal's significance evaporates when you switch from OLS to Newey-West, the apparent edge was mostly autocorrelation.
How practitioners use it
Used as context among multiple indicators — never as a standalone signal to act.
Less common professional uses
Overlapping-return regressions (e.g. predicting 20-day forward returns sampled daily) have mechanically induced MA(19) autocorrelation; the lag must be at least the overlap or the t-stat stays inflated. Hansen-Hodrick is the classic alternative for this exact case. The Bartlett kernel weights used by Newey-West guarantee a positive-semidefinite covariance estimate but can be conservative; the choice of kernel and bandwidth is a bias-variance tradeoff, and too-large L injects noise into the standard error itself. Because so many predictors have been mined from the same return series, Harvey-Liu-Zhu argue the effective significance hurdle for a NEW factor is a t-stat near 3.0, not 1.96 - the cross-sectional analogue of Deflated Sharpe.
Sources & provenance
Newey & West (1987), Econometrica; Harvey, Liu & Zhu (2016), '...and the Cross-Section of Expected Returns', Review of Financial Studies
This page is educational content published by Pachira Aquatica Global LLC. It is not investment advice and not a recommendation.