Time variation in the standard forward premium regression: Some new models and tests

• Published in: Journal of empirical finance Vol. 29; pp. 52 - 63
• Main Authors: Baillie, Richard T., Cho, Dooyeon
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2014
• Subjects: Capital market; Error; Financial crisis; Forward premium anomaly; Memory; Non-linear models; Regression analysis; Time-varying parameter regression; Forward premium anomaly; Time-varying parameter regression
• ISSN: 0927-5398; 1879-1727
• DOI: 10.1016/j.jempfin.2014.03.005

This paper makes two contributions to trying to understand the forward premium anomaly and the apparent breakdowns of Uncovered Interest Rate Parity (UIP). First, investigation of the time series properties of the forward premium reveals either four or five breaks in the last twenty three years and evidence of long memory within each sub period. In fact the forward premium is highly nonlinear and appears to defy classification as a process with a constant order of integration. The second aspect of the paper is concerned with the time varying nature of the estimate of the slope parameter when spot returns are regressed on the lagged forward premium. We compare rolling type regression estimates, with Bayesian estimation and also a new Time Varying Parameter (TVP) method that is motivated by the TVP autoregression of Giraitis et al. (2014). The procedure is a form of kernel weighted regression and delivers relatively tight standard errors on the parameter estimates. We find the existence of the forward premium anomaly with large negative beta coefficients in the 1980s and 1990s. For some currencies there is also evidence of large positive coefficients and a reversal of the forward premium anomaly after the financial crisis of 2008. •This paper is concerned with the forward premium anomaly•The forward premium is highly nonlinear, with breaks and long memory in sub periods.•The slope parameter, beta is estimated by several methods which allow time variation.•We introduce a new TVP kernel weighted method which has tighter confidence intervals.•There is evidence of large negative betas in the 1980s, 90s and reversal after 2008.