New methods of estimation and inference for regressions that use overlapping observations

I use Monte Carlo methods to investigate the empirical performance of a new econometric approach for handling time-series regressions that use overlapping observations. The approach resolves size distortions that arise from using heteroskedasticity-and-autocorrelation-consistent (HAC) standard errors to correct for mechanically-induced serial correlation in the regression residuals. But size distortions that arise from small-sample effects, such as small-sample bias, are substantial in many cases, especially when the contemporaneous innovations to the dependent variable and regressors are highly correlated. To illustrate the approach, I regress U.S. stock market returns on lagged measures of the market dividend yield using data for 20-year and 80-year sample periods. I find that testing the null hypothesis of no predictability produces the lowest p-values for regressions whose features tend to generate oversized tests according to Monte Carlo experiments. I also find that quadrupling the number of observations used to conduct the tests does little to increase the values of the chi-squared test statistics and actually produces lower values of these statistics in some cases.