Using the Haar wavelet transform in the semiparametric specification of time series

Using theoretical arguments for nonparametric wavelet estimation, we devise regression-based semiparametric wavelet estimators to dissect linear from nonlinear effects in a time series. The wavelet estimators localize in both time and frequency so that distortion due to outliers is lessened. Our regression-based approach also lends itself to ease of replication, clarity, flexibility, timeliness and statistical validity. We demonstrate the efficacy of the approach via rolling regressions on time series of quarterly U.S. GDP growth rates, monthly Hong Kong/ U.S. exchange rates, weekly 1-month commercial interest rates and daily returns on the S&P 500.