Event Details

Economics Research Seminar

Abstract: "The paper is concerned with separating long- and short-memory components in strongly persistent, possibly nonstationary time series. The Fourier representation of the logarithm of the spectral density provides an additive decomposition of the log-spectrum for fractionally integrated processes. This result yields an analytical decomposition of the impulse response function and of the optimal linear predictor

into long- and short-memory contributions. An alternative real-time decomposition is also derived, identifying a fractional noise component and a transitory component driven by antipersistent dynamics. The framework applies to a broad class of models, including ARFIMA and fractional exponential specifications. Empirical applications to inflation and realized volatility illustrate that long-memory dynamics dominate

and generate highly irregular paths. Motivated by these findings, we introduce a novel roughness measure for discretely observed time series."