On the impact of macroeconomic news surprises on Treasury-bond returns

See, among others, Ederington and Lee (1993), Bollerslev et al. (2000), Balduzzi et al. (2001), Nikkinen and Sahlstrom (2001), Graham et al. (2005), Lobo et al. (2006), Rigobon and Sack (2006), Lee and Mykland (2008), Birz and Lott (2011), Lahaye et al. (2011), Erdemlioglu et al. (2012) and Ludwig and Zimper (2013).

See Ball and Brown (1968), Fama et al. (1969), Fama and Schwert (1977) and the literature surveys by Fama (1970, 1991).

See e.g. Schwert (1981), Mitchell and Mulherin (1994) and the literature review by Fama (1991).

See, e.g. Jain (1988), Becker et al. (1995), Funke and Matsuda (2006), and Rigobon and Sack (2006).

See Ederington and Lee (1993), Balduzzi et al. (2001), Andersen et al. (2007) and Dungey et al. (2007) among others.

Especially for important news, such as those related to consumer price index, durable goods orders, GDP, personal spending and retail sales.

See also Dungey et al. (2007) who show that U.S. Treasuries tend to co-jump across maturities, with more unique jumps at both ends of the term curve, providing some support for both liquidity and preferred habitat hypotheses.

Other interesting references are Boyd et al. (2005) and Andersen et al. (2007).

It should be noticed that the well-known issue of simultaneous announcements encountered in any event study is not a problem in the present analysis. Indeed, our data are not dummy variables, but surprises, i.e. figures. Consequently, if an abnormal variation in returns is detected, the relevant corresponding announcement can be precisely identified.

Also called “consensus” data, i.e. the median of individual forecasts of the announcement i made each month by professional forecasters working mostly in banks, brokerages or forecast agencies. It should be noticed that the horizon of the expectations does not intervene in the analysis, due to the procedure followed by Bloomberg for constructing its consensus data. The procedure can be summarized as follows (see also Sect. 3.1). Around 1 month prior to the news release, Bloomberg sends out surveys to various analysts—including academicians and practitioners—to provide their forecasts of the upcoming announcements. The horizon is thus the same for all participants and expectations. During the week prior to the scheduled announcement, Bloomberg compiles all requested up-to-date expectations and publishes the median forecasts of upcoming announcements.

Returns are calculated as the first log-difference of mid prices defined as the average of the bid and ask prices being quoted. While yield is mainly a measure of income—ignoring the possibility of changes in market value that result in capital gains and losses—return is a more inclusive measure of investment performance, which can be either positive or negative (in contrast to yield, which must be positive). In addition, considering price returns allows us to deal with the potential unit root issue by working on stationary series (results of unit root tests are available upon request to the authors).

Due to a limited access to high frequency data for the U.S., we consider a 15-min frequency for this country. Note however that this allows us to deal with microstructure noise issues (see Lee and Mykland (2008)), although the U.S. market is obviously more liquid than the others.

See the references in Sect. 1.

Results of unit root tests are available upon request to the authors.

China exhibits a different pattern because news are published mid-month.

All detailed results are available upon request to the authors.

Computed as the mean between bid and ask prices.

The values of the \(R^{2}\), as well as those of the estimated coefficients (while being significant), may be viewed as quite low at the first sight. However, they are consistent with those obtained in the related literature (see, e.g., Balduzzi et al. 2001; Vrugt 2010). These values obviously depend on the considered frequency. For instance, when dealing with high frequency data, macroeconomic announcements occur infrequently and thus the corresponding regressions have a low \(R^{2}\). When the frequency decreases, the \(R^{2}\) augments. In our case, the \(R^{2}\) values range in intervals that are consistent with previous studies dealing with intraday data.

According to Laurencelle (2009), it has indeed two advantages over the other tests (such as Spearman, Pearson, Gamma-based tests\(\ldots \)): it transcends the metric of measured variables and removes the need for a parametric model, making it a truly non-parametric index, suitable for categorical ordinal variables [see also Agresti (1976) and Khamis (2008)].

Descriptive statistics for the data used in our analysis are reported in Tables 12, 13, 14, 15 and 16 in “Appendix”.

Only significant results are reported (the Kendall \(\tau _b\) statistic is not significant for the 10-year maturity U.K. bond, nor for China).

These different results obtained for the U.K. and U.S. can be linked to the type and cyclical nature of significant announcements. Indeed, as shown by the results displayed in Table 2, the U.S. 2-year bonds are particularly sensitive to news related to pro-cyclical indicators, such as non-farm payrolls, U.S. consumer confidence index or new home sales, while the U.K. 2-year bonds are less reactive to these indicators. As recalled by Andritzky et al. (2007), positive surprises in such pro-cyclical indicators tend to have a strong—negative—impact on the bond market, consistent with our findings. However, more than the sign itself, it is the number of surprises that matters here, in the sense that the largest number of surprises reflects a stronger effect for large unexpected shocks.