Content area
Full Text
(ProQuest: ... denotes formulae omitted.)
Are bond returns predictable? The answer depends on whom you ask. The response from some researchers might sound like this, for example: "Recent empirical research has uncovered significant forecastable variation in the excess returns of U.S. government bonds." (Ludvigson and Ng [2009]) Others might be less enthusiastic: "It is difficult to forecast bond market fluctuations, although some research shows that these fluctuations are not fully unpredictable." (Ilmanen [1997]) The former's optimism is based on the statistical evidence that bond returns can be forecast using the information contained in present yield spreads and forward rates and on more recent statistical evidence that some macroeconomic factors also have predictive power. The latter's skepticism stems from two concerns: first, few studies have linked the statistical evidence to economic value, and second, the statistical evidence itself has been called into question (Thornton and Valente [2012]).
This article presents historically profitable trading strategies, generated by models that are statistically sound. Put differently, we present new evidence of bond return predictability that is both statistically and economically significant. We start by replicating the Cochrane and Piazzesi [2005] regressions for annual returns of U.S. Treasury notes and bond futures contracts, using five forward rates. Like Cochrane and Piazzesi, we find high R2 values. Unlike them, we find that the estimated betas are erratic and do not exhibit a tent-shape pattern. We apply several well-known but rarely used regression diagnostic methods to measure the severity of these regressions' potential econometric weaknesses. Namely, we use a series of tests by Belsley [1982] and Belsley et al. [1980] to assess the effect of collinearity on the regression coefficients, and the Britten-Jones et al. [2011] transformation to assess the degree of autocorrelation caused by the use of the long-horizon, overlapping regressands. The tests show that both collinearity and autocorrelation appear to be problematic, the latter being the more serious of the two. Furthermore, an implementation of a trading strategy based on the futures returns forecasts using robust portfolio optimization results in no added economic value, relative to a passive benchmark.
We study two alternative specifications of the model. The first falls out naturally from the diagnostics of the annual returns regressions. To remove the overlap in the regressands, we use monthly...