Low Frequency Effects of Macroeconomic News on Government Bond Yields

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Low Frequency Effects of Macroeconomic News on Government Bond Yields Carlo Altavilla, Domenico Giannone, and Michele Modugno 2014-052 NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Low Frequency Effects of Macroeconomic News on Government Bond Yields∗ Carlo Altavillaa, Domenico Giannoneb, and Michele Modugnoc Abstract In this study, we analyze the reaction of the U.S. Treasury bond market to innovations in macroeconomic fundamentals. We identify these innovations based on macroeconomic news, which are defined as differences between the actual releases and market expectations. We find that that macroeconomic news explain about one-third of the low frequency (quarterly) fluctuations in long-term bond yields. When we focus on the high frequency (daily) movements, this decrease to onetenth. Thisisbecausemacroeconomicnewshaveapersistenteffecton bond yields, whereas non-fundamental factors have substantial effects on the day-to-day movements of bond yields, although their effects are shorter lived. Keywords: macroeconomic announcement, news, treasury bond yield JEL classification: E43; E44; E47; G14. ∗We would like to thank Marco Del Negro, Luca Guerrieri, Refet Gurkaynak, Nellie Liang, Roberto Motto, Kleopatra Nikolaou, and Steve Sharpe for helpful comments and discussions, and seminar participants at the Federal Reserve Board, European Central Bank, George Washington University, University of York, Universit´e libre de Bruxelles, the CSEF-IGIER Symposium on Economics and Institutions, the 22nd Symposium of the Society for Nonlinear Dynamics and Econometrics, Birmingham Macroeconomics and Econometrics conference, and EMMPA 2014 in Bucharest. Domenico Giannone was supported by the “Action de recherche concert´ee” contract ARC-AUWB/2010-15/ULB-11 and by IAP research network grant no. P7/06 from the Belgian government (Belgian Science Policy). The opinions in this paper are those of the authors and do not necessarily reflect the views of the European Central Bank, the Eurosystem, or the Board of Governors of the Federal Reserve System. aEuropean Central Bank, email: carlo.altavilla@ecb.int bLUISS University of Rome, ECARES, EIEF and CEPR, email: dgiannon@ulb.ac.be cFederal Reserve Board, email: michele.modugno@frb.gov

1 Introduction We analyze the reaction of the U.S. Treasury bond market to innovations in macroeconomic fundamentals. We identify innovations in macroeconomic fundamentals based on macroeconomic news, which we define as the differences between the actual macroeconomic releases and the median market predictions by participants for those releases. Our analysis is based on the regression of the daily changes in bond yields on macroeconomic news, in the same manner as event studies. The fit of the regression and the corresponding residuals are defined as the fundamental and non-fundamental components of bond yield changes, respectively. Identifyinginnovationsinmacroeconomicfundamentalsbasedonmacroeconomic news is a natural strategy. In most industrialized countries, various macroeconomic indicators are released by national statistical agencies and specialized private firms almost every calendar day. Policy makers, media commentators, and market participants monitor the real-time data flow constantly and somewhat obsessively. The market participants also make a prediction for almost every macroeconomic release and whenever they are surprised asset prices tend to move. Focusing on high frequency changes can facilitate the correct identification of the causal effects of macroeconomic news by reducing the effects of confounding factors and by limiting reverse causality issues (see Gurkaynak and Wright, 2013; Kuttner, 2001; and Cochrane and Piazzesi, 2002). In agreementwithpreviousstudies, wefindthatseveraltypesofmacroeconomic news are economically important and they have statistically significant im- 2

pacts on daily changes in bond yields. However, their explanatory power is quite limited, i.e., the R2 value of the regression is about 10% only. Thus, we develop a method to isolate the low frequency effects of macroeconomic news while preserving the information provided by the high frequency reaction of asset prices to the release of macroeconomic information in real-time. By summing the fit of the daily regressions over a month (quarter), we obtain the fit for the monthly (quarterly) changes in bond yields. The fundamental component become more important when focusing on these low-frequency changes; indeed, moving from daily to quarterly increases the R2 value to over 30%. This is because macroeconomic news has a persistent effect on bond yields, whereas the effect of non-fundamental factors is less persistent and it tends to average out when focusing on longer horizon changes. In other words, the importance of macroeconomic factors might be hidden by the high frequency noise that dominates the daily fluctuations in bond yields. Interestingly, the interaction between macroeconomic news and yields did not break apart after the zero lower bound became binding at the end of 2008. In agreement with Swansson and Williams (2014), we find that the high frequency effects remained stable. More interestingly, we show that macroeconomic news continued to exert an important influence at a low frequency on changes in bond yields. This evidence corroborates the view that the non-standard monetary policies adopted by the U.S. Federal Reserve, i.e., a combination of forward guidance and large-scale asset purchases, have been successful in keeping the bond yields anchored to macroeconomic news, thereby limiting non-fundamental fluctuations during a period of high economic uncertainty. 3

Our results reconcile some contrasting findings obtained in previous studies. As stressed above, event studies indicate that macroeconomic releases account for only a small proportion of the daily variation in bond yields (see Gurkainak, 2014). By contrast, macro-finance models estimated at monthly or quarterly frequencies can explain a significant fraction of the bond yield fluctuations with macroeconomic variables (see Ang and Piazzesi, 2003,Dieboldetal.,2006,Coroneoetal.,2013). Werationalizethisempirical evidence by identifying the role of the relative persistence of the fundamental and non-fundamental components in influencing bond yield fluctuations over different time spans. The remainder of the paper is organized as follows. Section 2 describes macroeconomic news and its effects on bond yields at different frequencies. We discuss how our findings affect excess bond returns for investors with different investment horizons. Section 3 demonstrates the effect of macroeconomicnewsonstockpricereturnsandexchangeratesatdifferentfrequencies. Section 4 analyzes the impact of macroeconomic news before and during the zero lower bound period. Section 5 concludes this study. 2 Effects of Macroeconomic News on Bond Yields at High and Low Frequencies The data used in this study came from various sources. We use the zerocoupon yields constructed by Gurkaynak, Sack, and Wright (2007) from 1- 4

to 10-year horizons.1 This dataset also includes the parameters estimated for the model of Svensson (1994) to smooth the yield data. In principle, one can retrieve any desired maturity using these parameters. Section 2.3 reports the 3-month holding period excess returns computed using data generated with these parameters for the maturities that were not available in Gurkaynak, Sack, and Wright (2007) plus the 3-month Treasury bill.2 In order to simulate the macroeconomic information that is available in real time to market participants, we use the data contained in the Economic Calendars (ECO) provided by Bloomberg. For each macroeconomic release, this dataset contains the realized value and the predictions made by a panel of market participants for the same value. ECO survey forecasts normally start one to two weeks before each release and they are updated in real time until the macroeconomic variable is released officially. The survey value used in the empirical analysis is the median (consensus) forecast. Using both the official releases and the corresponding forecast for each macroeconomic variable allows us to reconstruct the size and direction of all news that hit the market at each point in time. The first column of Table 1 provides an overview of the macroeconomic variables used in the analysis. The sample period is January 1, 2000 to January 28, 2014. We consider all U.S. macroeconomic news available for the entire sample, with a total of 41 variables. For some of the listed variables, Bloomberg collects more than one release. This is the case for the 1This dataset is publicly available on the website of the Federal Reserve Board. The daily data can be obtained at www.federalreserve.gov/pubs/feds/2006/. 2These data are publicly available on the website of the Federal Reserve of St Louis at www.research.stlouisfed.org/fred2/. 5

GDP annualized QoQ and GDP price index, for which we have advanced (A), second (S), and third (T) releases; and for nonfarm productivity, unit labor costs, and the University of Michigan Confidence, for which we have preliminary (P) and final (F) releases. We treat these releases as separate variables. The second column of Table 1 shows the relevance index. The value of this index corresponds to the percentage of Bloomberg users who set an alert for a particular event. For example, over 98% of the users set an alert to be notified before the scheduled release of the change in the nonfarm payrolls variable. This index gives an idea of the releases that are important to market participants. Note that the number of releases observed for each variable depends on its frequency. The third column indicates the frequency of each variable, i.e., whether it is released on a weekly (W), monthly (M), or quarterly (Q) basis. The fourth column reports the publication delay, i.e., the average number of days from the end of the period considered for each variable and the day of release. For example, the change in the nonfarm payrolls data is usually released 4 days after the end of the reference month. A negative entry, such as the University of Michigan Confidence, means that the variable is released before the end of the reference period. 2.1 Empirical analysis based on the daily frequency First, we analyze the daily reaction of bond yields to macroeconomic news. Thus, we regress the daily change in a bond yield ∆yτ at maturity τ on day t t on a constant and on the news released on day t, according to Equation (1). If variable i was not released at time t, we set news = 0. i,t 6

n (cid:88) ∆yτ = c+ βτnews +ετ (1) t i i,t t i=1 Table 1 shows the regression coefficients (β) based on the regression described in Equation (1) for the bond yields with 1-, 5-, and 10-year maturities.3 We use boldface to denote coefficients that differ significantly from zero at the 5% confidence level. Three groups of variables are particularly important for explaining the daily changes in yields throughout the whole maturity spectrum: surveys (consumer confidence, ISM manufacturing and non-manufacturing, Philadelphia Fed. economic outlook, and University of Michigan Confidence preliminary), employment-related variables (change in nonfarm payrolls and initial jobless claims), and other macroeconomic variables (e.g., GDP annualized QoQ advanced and advanced retail sales). Surveys are important because of their timeliness, as they are the first types of information available regarding the economic condition in the current month. Jobless claims are released on a weekly basis. Similar to surveys, jobless claims are very timely, which makes them useful to market participants for understanding the labor market conditions. Finally, other variables such as GDP, nonfarm payrolls, and sales are important indicators of the state of the economy and they are monitored closely by the Federal Reserve in order to determine its monetary policy stance. Thus, these indicators are also relevant to market participants. The last row of Table 1 shows the R2 values for the regression described in Equation (1). Some of the regression parameters 3Note that to facilitate the comparison, macroeconomic news were standardized by dividing the difference between the actual and the predicted value of each variable by the corresponding sample standard deviation. 7

are statistically significant, but macroeconomic news explains only a small fraction of the daily variation in bond yields, i.e., less than 10%. INSERT TABLE 1 HERE 2.2 Empirical analysis at lower frequencies Let us define the daily news index ∆ (cid:91) 1yτ = nix1,τ as the fitted value from t t Equation (1). To analyze the persistence of the effects of macroeconomic newsonyieldchanges, weaggregatetheyieldsandnewsindicesoverdifferent time spans. Specifically, we aggregate the daily changes in bond yields to obtain longer horizon changes. h−1 (cid:88) yτ −yτ := ∆hyτ = ∆yτ (2) t t−h t t−j j=0 Similarly, we sum the daily news indices to obtain longer horizon news indices at daily frequencies, as follows. h−1 (cid:88) nixh,τ = nix1,τ (3) t t−j j=0 The effect of these aggregations on the yields is to “cleanse” the series of high-frequency fluctuations and to give more weight to fluctuations with frequencies lower than h days. The following analysis focuses on regression equations: ∆hyτ = γh,τnixh,τ +υh,τ, (4) t t t 8

where γh,τ measures the impact of the sum of the news on the change in yields over h days and it enforces the orthogonality between the fundamental and non-fundamental components at any horizon.4 The fitted value of (cid:91) Equation (4), ∆hyτ, represents the part of the h−days changes in the bond t yields attributable to macroeconomic fundamentals. On average there are (cid:92) 22 trading days per month, thus ∆22yτ and ∆22yτ approximately correspond t t to the actual and fitted monthly changes in the bond yields at maturity τ, (cid:92) respectively. By contrast, ∆66yτ and ∆66yτ refer to quarterly changes. The t t (cid:91) residual, ∆hyτ − ∆hyτ, defines the component driven by non-fundamental t t (cid:91) (cid:91) factors. In the following, we refer to ∆hyτ and ∆hyτ − ∆hyτ as the fundat t t mental and non-fundamental components of the h−days changes in the bond yields with maturity τ, respectively. For simplicity, we define the part of the bond yields that is not explained by macroeconomic news as the non-fundamental part; however, we have to consider that this part includes fundamental innovations that we cannot extract. The types of macroeconomic news considered in this study are only a subsample of the innovations in macroeconomic fundamentals that may affect U.S. Treasury yields. We do not consider surprises related to policy announcements regarding monetary and fiscal policy interventions. Moreover, we only consider U.S. variables, but international factors could also have played important roles. Most importantly, we only control for partial measures of the news because Bloomberg collects the expectations for headline information whereas macroeconomic data releases include numerous disaggregated details. Recently, Gurkaynak (2014) showed that considering 4Note that all of these results can be confirmed qualitatively if we do not include γ. 9

unmeasured news greatly increases the explanatory power of macroeconomic releases that occur at a high frequency. Figure 1 shows the actual and fitted values for the daily, monthly, and quarterly yield changes (h = 1,22,66) in government bond yields with τ = 1,5,10-year maturities. Figure 2 shows the R2 values for the regressions. INSERT FIGURES 1 HERE INSERT FIGURE 2 HERE If we consider only the actual values, i.e., the actual changes in bond yields, we can explain the effect of filtering better. Thus, the more we filter, i.e., aggregate the daily changes into monthly and quarterly changes, the more we “cleanse” our series of high frequency noisy fluctuations, thereby highlightingthelowfrequencyfluctuations. Inotherwords,filteringidentifies the long-term patterns or low-frequency fluctuations in our variables. If we consider the fitted values obtained from Equation (1), it is evident that the fit is quite poor for the daily changes. However, if we consider the monthly and quarterly changes, the fitted values can capture a larger fraction of the variation in the bond yield changes. It is useful to introduce a measure of persistence to better understand what drives the observed increase in the R2 value with the horizon of the changes. According to Cochrane (1988) and Cochrane and Sbordone (1988), the persistency of a series, such as x , can be assessed by considering 1/h t times the variance in the h-period change, i.e., 1/h var(x − x ), as a t t−h 10

function of h. If all the shocks to x tend to be incorporated immediately and t permanently, then the series comprises white noise and 1/h var(x −x ) is t t−h constant with respect to h. However, if the effect of shocks on x are partially t reversed after some time, the reversion will be reflected in the decline of 1/h var(x − x ) from a given horizon onward. By contrast, if it takes t t−h time for the shocks to be incorporated, then 1/h var(x −x ) will tend to t t−h increase. Since the R2 for different horizons can be written as (cid:16)(cid:91)(cid:17) 1/h var ∆hyτ t R2(h,τ) := , (cid:16)(cid:91)(cid:17) (cid:16) (cid:91)(cid:17) 1/h var ∆hyτ +1/h var ∆hyτ −∆hyτ t t t it follows that the increased importance of macroeconomic news for changes in government bond yields over longer horizons can be explained by the relative persistence of the fundamental and non-fundamental components. Figure 3 reports 1/h times the variance in the bond yields, and their fundamental and non-fundamental components at different maturities, for daily (h = 1), monthly (h = 22), and quarterly (h = 66) changes. It is evident that the change in 1/h times the variance in the h-period decreases for the non-fundamental part when moving from daily to monthly and from monthly to quarterly horizons. This decrease is particularly evident with medium and long maturities. By contrast, the change in 1/h times the variance in the h-period does not decline for the fundamental components. Therefore, wecanconcludethattheincreaseintheR2 valueisattributableto the fact that shocks to the fundamental components tend to be incorporated immediately with long-lasting effects, but less time tends to be needed for 11

shocks to the non-fundamental components to be reverted. INSERT FIGURE 3 HERE In summary, these results indicate that after the high frequency fluctuations in yields are filtered out via aggregation, the macroeconomic news component has a strong explanatory power, i.e., up to 25% for the monthly aggregation and 35% for the quarterly aggregation. This is because although theeffectsofmacroeconomicnewsonyieldsarepersistent,thehigh-frequency fluctuations due to non-fundamental factors tend to be short-lived, thus they are aggregated and less evident within the course of one month (or quarter). The impact of macroeconomic news tends to be long-lasting, thus these type of news are more suitable for explaining low frequency fluctuations in government bond yields. These results reconcile the findings of the high frequency event-study literature with the macro-finance literature. Ang and Piazzesi (2003), Diebold et al. (2006), and Coroneo et al., (2013) show that when estimating models at monthly or quarterly frequencies, significant proportions of the bond yield fluctuations are driven by macroeconomic variables that measure real activities and prices. Our findings explain why this correlation exists at low frequencies, i.e., macroeconomic news persistently affects the portfolio strategies of fixed-income market participants. 12

2.3 Implications for Excess Returns The low frequency fluctuations in yields are closely related to bond returns for investors with holding periods longer than one day. To observe this relationship, we define rxτ,k as the k-day holding period t excess bond return. We have: rxk,τ = −(τ −k)yτ−k +τyτ −yk, (5) t+k t+k t t where −(τ −k)yτ−k is the (log) price at which the bond is sold at time t+k t+k for selling a bond with maturity τ −k, −τyτ is the (log) price paid at time t t when the bond reaches maturity τ, and yk is the interest paid for borrowing t money for the period k. Thus, Equation (5) can be rewritten as: k (cid:88) rxk,τ = −(τ −k)yτ−k −(τ −k) ∆yτ−k +τyτ −yk. (6) t+k t t+i t t i=1 Fork = 66,whichisequivalenttoonequarter,bysubstituting (cid:80)66 ∆y i=1 t+i with the fit obtained from Equation (4), we obtain the fitted k-days holding period excess bond return: rˆx66,τ = −(τ −66)yτ−66 −(τ −66)γq,τ−66nixq,τ−66 +τyτ −y66. (7) t+66 t t+66 t t In order to compute the k-days holding period excess bond returns with maturities of τ= 12, 24, 36, 48, 60, 72, 84, 96, 108, and 120 months, we need to generate the yields with maturity τ − k. These yields can be generated using the parameters of the model proposed by Svensson (1994), which are 13

included in the dataset of Gu¨rkaynak, Sack, and Wright (2007). Figure 4 shows the excess returns based on a 3-month holding period for an equally weighted portfolio of bonds with maturities ranging from 1 to 10 years and the implied fitted value obtained from the regression described in Equation (7).5 We study the external validity of the model based on a pseudo-out-ofsample exercise. We compute the fitted return using the parameters estimated from the data up to December 15, 2008. The sample split is selected to coincide with the date when the monetary policy reached the Zero Lower Bound.6 Thefittedreturnsfortheremainingpartofthesamplearecomputed using these parameters. The out-of-sample fit is shown in Figure 4, where the shaded area highlight the period used for the out-of-sample validation. Figure4showsclearlythatthemacroeconomicfundamentalsperformwell in tracking the 3-month holding period excess bond return and they explain 35% of its fluctuations. The in-sample and out-of-sample fits are remarkably similar and almost undistinguishable, thereby indicating that the importance of macroeconomic news in driving bond returns is a robust result and not an artifact due to overfitting or to other spurious effects. INSERT FIGURES 4 HERE Empirical research into financial economics has shown that a significant fractionofthevariationinexcessbondreturnsispredictable. FamaandBliss 5The return of this portfolio and the relative fit, respectively, are defined as: r¯x66 = t 1 (cid:80) rxτ,66 and r¯ˆx66 = 1 (cid:80) rˆxτ,66. 10 τ=[12,24,...,120] t t 10 τ=[12,24,...,120] t 6We return to this issue in Section 4. 14

(1987) and Campbell and Shiller (1991) showed that excess returns can be predicted based on the forward rate spreads and yield spreads. Cochrane and Piazzesi (2005) found that about one-third of the variation in excess bond returns can be predicted using a linear combination of forward rates. Understanding the reasons and the sources of such predictability are important questions in economics and finance. The decomposition of the bond returns derived above may help to shed some new light on this old debate. In the following, we show that the predictability of returns is due to nonfundamental fluctuations because the component of bonds returns driven by macroeconomic news is unpredictable. We construct a factor similar to that used by Cochrane and Piazzesi (2005), except it is for a 3-month holding period (we refer to this as the CP factor), and we find that it can only predict the non-fundamental part of the excess bond returns. We construct the CP factor from the available yields with maturities of 12 to 120 months and from the generated yields with maturities of 9, 21, 33, 45, 57, 69, 81, 93, 105, and 117 months. First, we compute the bond log prices: pτ ≡ −τyτ t t and then the log forward rate between time t+τ −66 and t+τ as: fwτ ≡ pτ−66 −pτ. t t t We collect the intercept, the 3-month Treasury bill, and the forwards in the vector g = [1,y3,fw12,fw24,...,fw120](cid:48), and estimate the parameters of t t t t t the following equation: 15

r¯x66 = ρg +(cid:15)¯, t t−66 t where we define CP = ρˆg . To understand how much of the excess bond t t returns can be predicted by the yield curve itself, we perform the following predictive regression: x = c+β CP +w , (8) t 2 t−66 t where x is in turn r¯x66, the observed 66-day holding period excess bond t t return; f66 ≡ − 1 (cid:80) (τ −66)γq,τ−66nixq,τ−66 is the fundamental t 10 τ=[12,24,...,120] t+66 part of the 66-day holding period excess bond return; and nf66 = r¯x66−f66 t t t is the non-fundamental part. Table 2 shows the coefficients and the relative R2 values of these regressions for the three dependent variables. The results show that the CP factor predicts a large proportion (20%) of the excess bond returns. Unsurprisingly, this proportion is related mainly to the non-macroeconomic news part, nf (18%), whereas the CP factor t explains almost nothing about the news-related part. This result is not surprising if we consider the nature of the elements we are analyzing, where the forward prices are determined by market participants at time t − 66 given the information available at that time. By definition, macroeconomic news comprises surprises for market participants (i.e., innovations to their information set) that occur between time t−65 and t, thus they cannot be predicted by forward rates that are based on the information available to market participants at time t−66. Insummary,thepredictablecomponentofbondreturnsisnon-fundamental 16

because shocks to this component generate predictable dynamics, as they tend to be reverted. INSERT TABLE 2 HERE 3 Macroeconomic News, Stock Prices, and Exchange Rates The impact of macroeconomic news has been previously studied for other assets/markets. More precisely, several event studies have analyzed the daily and intra-daily fluctuations in stock prices and exchange rates (e.g., see Andersen et al. 2003b and 2007, and Faust et al. 2007). The general finding of these studies indicates that also these assets are sensitive to macroeconomic news. We analyze the impact of macroeconomic news on longer horizon changes in the trade-weighted U.S. dollar index (major currencies) and the S&P 500 stock price index to assess whether these assets have the same low frequency sensitivity to macroeconomic news as bond yields. We are aware that foreign macroeconomic news can have an important impact, especially on the exchange rate, thus our analysis is incomplete. Nevertheless, U.S. economic fundamentals should play a predominant role in determining these asset prices. Table 3 shows the coefficients obtained from the regression of the daily returns of these assets on the macroeconomic news, which is equivalent to 17

Equation (1). The fits of the returns due to macroeconomic news over different horizons are shown in Figure (5) INSERT TABLE 3 HERE INSERT FIGURES 5 HERE For the trade-weighted U.S. dollar index, six types of news have statistically significant impacts: changes in nonfarm payrolls, ISM manufacturing, producer price index (excluding energy and food), unemployment rate, advancedGDP,andthefinalreleaseofnonfarmproductivity. However,macroeconomic news do not have a persistent effect on the exchange rate. As shown in Table 3, the R2 value for the daily changes, Equation (1), is equal to 2%, and the R2 values became lower as we filter out more dependent variables, i.e., the monthly and quarterly R2 values from Equations (4) are equal to zero. In our analysis of S&P 500 returns, we find that only four types of macroeconomic news have coefficients that differ significantly from zero: capacity utilization, ISM manufacturing and non-manufacturing, and retail sales. However, in contrast to the exchange rate results, the effect of U.S. macroeconomic news on the S&P 500 stock price index, similar to bond yields, tends to increase with the horizon, where R2 value is 2% for daily changes, 5% for monthly changes, and 15% for quarterly changes. Although the increase in the explanatory power of macroeconomic news with a longer 18

horizon is qualitatively similar to that observed for bond yields, the effect of macroeconomic news on S&P 500 returns is quantitatively much smaller. It is likely that international macroeconomic news is more important for stock returns than for bond yields, but this is a topic for future research. 4 Government Bonds and Macroeconomic News at the Time of the Zero Lower Bound The normal implementation of monetary policy provides a link between macroeconomic news and Treasury bond yields. In normal times, the central bank reacts to macroeconomic news by changing the short-term policy rate, thereby influencing a broad spectrum of fixed income asset classes. However, this mechanism cannot work at the zero lower bound because the interaction between macroeconomic news and yields may break apart, thus the low frequency effect of macroeconomic news might disappear. Therefore, we analyze whether our results change since the zero lower bound became binding. First, we estimate the regression model described in Equation (1) augmented with a zero lower bound dummies interacting with each type of news: n n (cid:88) (cid:88) ∆yτ = c+ βτnews + δτ(zlb ×news )+ετ, (9) t i i,t i t i,t t i=1 i=1 where zlb is an indicator variable that takes a value of 1 when the zero lower t bound was binding, i.e., from December 16, 2008 to January 28, 2014, and 0 before, i.e., from January 1, 2000 to December 15, 2008. The coefficient δτ measures whether the impact of each type of news on the change in bond i 19

yields varied after the policy rate reached the zero lower bound. The estimation results are shown in Table 4 for the maturities at τ = 1,5,10 years. The results suggest that although some of these coefficients change quantitatively in the two subsamples, their differences are rarely statistically significant, especially for long maturities. The unchanged responsiveness of bond yields with long maturities at high frequency was recently interpreted by Swanson and Williams (2013) as evidence that unconventional policy actions appear to have helped offset the effects of the zero bound on medium- and longer-term rates. The fact that these effects have remained persistent and sizeable over longer horizons lends additional support to this view. INSERT TABLE 4 HERE INSERT FIGURES 6 HERE Figure 6 shows the R2 values computed during the pre-zero lower bound periodandduringthezerolowerboundperiod. Thethreepanelsinthefigure show the R2 values for the daily changes, as in Equation (1), and monthly changes and quarterly changes, as in Equation (4). Interestingly, the interaction between macroeconomic news and yields did not break apart. These results provide evidence that the measures adopted by the Federal Reserve at the zero lower bound, i.e., forward guidance and large-scale asset purchases, did not weaken the relationship between macroeconomic news and 20

bond yields at low frequencies. These results also suggest that the introduction of explicit macroeconomic targets in the central bank communication did not influence the sensitivity of sovereign bonds to macroeconomic news.7 As a consequence, market participants continued to pay attention to macroeconomic news to understand the state of the economy and to anticipate the decisionsoftheFederalReserveregardingthefuturemonetarypolicystance.8 5 Conclusions Using high frequency data, we found that macroeconomic fundamentals have sizeable low frequency effects on sovereign bond yields. This feature cannot be detected by looking at high frequency fluctuations since the effect of macroeconomic fundamentals is persistent but low in terms of impact whereas the effect of non-fundamental factors is shorter lived but large in terms of impact. Animportantimplicationofourresultsisthatmacroeconomicnewshasa considerable effect on the dynamics of excess bond returns when the holding period extends beyond a single day. Interestingly, this is a specific feature of bond yield returns. The explanatory power of macroeconomic factors for stock and exchange rate returns also remains low at low frequencies. 7Forexample, theFOMCStatementofAugust2011statedthat“Committeecurrently anticipates that economic conditions [....] are likely to warrant exceptionally low levels for the federal funds rate at least through mid-2013.” On December 12, 2012 the FOMC indicated that a federal funds rate close to zero would remain appropriate at least as long as the unemployment rate remained above 6.5 % and inflation expectations continued to be well anchored. 8These results are robust and they are not due to overfitting. Indeed, we obtained the same fit using the parameters estimated in the pre-zero lower bound period for the zero lower bound period. 21

Although we considered a large dataset of news, we probably underestimated the importance of macroeconomic fundamentals for bond yields for several reasons. First, fundamental events may have an immediate effect on bond yields, but they cannot be captured immediately by macroeconomic data. Second, Bloomberg does not collect market expectations for all of the released variables. Third, we only considered U.S. macroeconomic news, but innovations in the fundamentals of other countries could also be important for U.S. bond yields. Overall, these considerations suggest that we underestimated the effect of macroeconomic fundamentals on bond yields. Thus, it is highly likely that fundamentals explain more than one-third of the lowfrequency fluctuations in bond yields. 22

6 References Andersen, T. G., T. Bollerslev, F. X. Diebold, and C. Vega (2003). “Micro Effects of Macro Announcements: real time Price Discovery in Foreign Exchange Markets,” American Economic Review Volume 93, Issue 1, pages 38-62. Andersen, T. G., T. Bollerslev, F. X. Diebold, and C. Vega (2007). “Real time Price Discovery in Global Stock, Bond and Foreign Exchange Markets,” Journal of International Economics Volume 73, pages 251-277. Ang, A. and M. Piazzesi (2003). “A no-arbitrage vector autoregression of term structure dynamics with macroeconomic and latent variables,” Journal of Monetary Economics, Elsevier, Volume 50, Issue 4, pages 745-787. Cochrane, J. (1988). How Big Is the Random Walk in GNP?, Journal of Political Economy, Volume 96, Issue 5, pp. 893-920. Cochrane J, and Piazzesi M (2002) The Fed and interest rates: a highfrequency identification, American Economic Review, 92:90-101. Cochrane, J., and M. Piazzesi (2005).“Bond risk premia,” American Economic Review, Volume 94, Issue 1, pp. 138-160. Cochrane, J. and A. Sbordone (1988).“Multivariate Estimates of the Permanent Components of GNP And Stock Prices,” Journal of Economic Dynamics and Control, Volume 12, pp. 255-296. Coroneo,L.,D.Giannone,andM.Modugno(2013). “UnspannedMacroeconomic Factors in the Yields Curve,” ECARES Working Papers 2013-07, Universite libre de Bruxelles. Diebold, F. X., G. D. Rudebusch and S. B. Aruoba (2006). “The Macroe- 23

conomy and the Yield Curve: A Dynamic Latent Factor Approach,” Journal of Econometrics, Volume 131, pages 309-338. Faust, J., J. H. Rogers, S. B. Wang, and J. H. Wright (2007). “The High- FrequencyResponseofExchangeRatesandInterestRatestoMacroeconomic Announcements,” JournalofMonetaryEconomics, Volume54, Issue4, pages 1051-1068. Gu¨rkaynak, R.S.(2014). “IdentifyingEffectsofPartially-MeasuredNews Surprises,” Unpublished Manuscript, Bilkent University Gu¨rkaynak, R. S., B. Sack, and J. H. Wright (2007). “The US Treasury yield curve: 1961 to the present,” Journal of Monetary Economics,Volume 54, Issue 8, pages 2291-2304 Gu¨rkaynak, R. S. and J. H. Wright (2013). “Identification and Inference Using Event Studies,” CEPR Discussion Papers 9388. Kuttner K. (2001) Monetary policy surprises and interest rates: evidence from the fed funds futures market, Journal of Monetary Economics, 47:523- 44. Ludvigson, S.C., and S. Ng (2009). “Macro factors in bond risk premia,” Review of Financial Studies, Volume 22, Issue 12, 5027-5067. Swanson, E. and J. Williams (2014). “Measuring the Effect of the Zero Lower Bound on Medium- and Longer-Term Interest Rates,” The American Economic Review (forthcoming). 24

Figures 25

0.5 0 −0.5 02 04 06 08 10 12 yliad 1−year 5−year 10−year 0.5 0.5 0 0 −0.5 −0.5 02 04 06 08 10 12 02 04 06 08 10 12 1 0.5 0 −0.5 −1 −1.5 02 04 06 08 10 12 ylhtnom 1 1 0.5 0.5 0 0 −0.5 −0.5 −1 −1 −1.5 −1.5 02 04 06 08 10 12 02 04 06 08 10 12 1 0 −1 −2 02 04 06 08 10 12 ylretrauq 1 1 0 0 −1 −1 −2 −2 02 04 06 08 10 12 02 04 06 08 10 12 actual fit Figure 1: Daily, monthly, and quarterly bond yield changes Notes: The figure shows the daily, monthly, and quarterly yield changes for 1-, 5- and 10-year maturities, where their fits were obtained using Equations (1) and (4) and estimated for the entire sample from January 1, 2000 to January 28, 2014. 26

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 1 2 3 4 5 6 7 8 9 10 maturities (in years) 2 R Daily (h=1) Monthly (h=22) Quarterly (h=66) Figure 2: R2 for the daily, monthly, and quarterly bond yield changes Notes: The figure shows the R2 values from the regressions of the daily, monthly, and quarterly changes in yield at different maturities based on the daily, monthly, and quarterly news indexes, as in Equations (1) and (4). 27

x 10−3 yields x 10−3 fundamental x 10−3 non fundamental 4 4 4 3.5 3.5 3.5 3 3 3 2.5 2.5 2.5 2 2 2 1.5 1.5 1.5 1 1 1 0.5 0.5 0.5 0 0 0 2 4 6 8 10 2 4 6 8 10 2 4 6 8 10 maturities (in years) Daily (h=1) Monthly (h=22) Quarterly (h=66) )x(rav h/1 t Figure 3: 1/h times variance of the h difference in bond yield, the fit, and the residuals. Notes: The figure shows 1/h var (cid:0) ∆hyτ(cid:1) (left panel), 1/h var (cid:16) ∆ (cid:92) hyτ (cid:17) (middle panel) and t t (cid:16) (cid:92)(cid:17) 1/h var ∆hyτ −∆hyτ (right panel), multiplied by 100, for different maturities (τ) and different t t horizons: h=1 (daily), h=22 (monthly), and h=66 (quarterly). 28

10 8 6 4 2 0 −2 −4 −6 −8 −10 Jul02 Jan05 Jul07 Jan10 Jul12 actual fit out−of−sample Figure 4: Three-month holding period excess bond returns Notes: The figure shows the 3-month holding period excess bond returns average across maturities (blue line), the fit obtained with the macroeconomic news (red line) using Equation (7), and the out-of-sample (green line). The shaded area indicates the out-of-sample period. 29

2 0 −2 −4 00 02 05 07 10 12 yliad TWEX S&P500 10 5 0 −5 00 02 05 07 10 12 10 5 0 −5 00 02 05 07 10 12 actual fit ylhtnom 20 0 −20 00 02 05 07 10 12 10 0 −10 00 02 05 07 10 12 ylretrauq 20 0 −20 −40 00 02 05 07 10 12 Figure 5: Other assets. Notes: The figure shows the daily, monthly, and quarterly asset returns for the trade-weighted U.S. dollarindex(TWEX)andtheS&P 500, wheretheirfitswereobtainedusingEquations(1)and(4), and estimated based on the entire sample from January 1, 2000 to January 28, 2014. 30

0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 2 4 6 8 10 2R Daily Monthly Quarterly 0.4 0.4 0.35 0.35 0.3 0.3 0.25 0.25 0.2 0.2 0.15 0.15 0.1 0.1 0.05 0.05 0 0 2 4 6 8 10 2 4 6 8 10 maturities (in years) pre−ZLB ZLB Figure 6: R2 values for the pre-zero lower bound and zero lower bound. Notes: The figure shows the R2 values from the regressions of the daily (left-hand side panel), monthly (center panel), and quarterly (right-hand side panel) changes in yields based on the daily, monthly, and quarterly news indexes from the pre-zero lower bound and zero lower bound subsamples. 31

Tables 32

Table 1: Macroeconomic News and their Effects on Bond Yields Releases Relevance Freq Pub. Delay 1-year 5-year 10-year Advance Retail Sales 89 M 15 1.20 1.82 1.45 Business Inventories 34 M 45 -0.18 -0.18 -0.03 Capacity Utilization 61 M 16 1.20 1.52 1.28 Change in Nonfarm Payrolls 98 M 4 3.44 4.43 3.59 Consumer Confidence 95 M 2 0.87 0.96 0.94 Consumer Credit 36 M 38 -0.16 -0.34 -0.38 Consumer Price Index (MoM) 93 M 18 0.36 0.58 0.21 CPI Ex Food & Energy (MoM) 75 M 18 0.48 0.38 0.31 Domestic Vehicle Sales 30 M 3 0.90 0.30 -0.06 Durable Goods Orders 91 M 21 0.49 0.78 0.71 Employment Cost Index 71 M 31 0.18 0.36 0.29 Factory Orders 82 M 34 0.17 0.23 0.27 Housing Starts 88 M 19 0.23 0.32 0.03 Import Price Index (MoM) 78 M 11 0.05 0.00 -0.15 Industrial Production 87 M 16 -0.02 -0.26 -0.80 Initial Jobless Claims 99 W 5 -1.12 -1.57 -1.42 ISM Manufacturing 94 M 2 1.73 2.78 2.66 ISM Non-Manf. Composite 70 M 2 1.67 2.19 2.01 Leading Indicators 84 M 24 0.25 0.72 0.91 New Home Sales 90 M 25 0.40 0.59 0.74 Personal Income 83 M 21 -0.37 -0.39 -0.31 Personal Spending 83 M 21 0.31 0.16 0.13 Philadelphia Fed. 75 M -14 1.11 1.96 1.73 PPI Ex Food & Energy (MoM) 68 M 14 0.27 1.06 1.39 Producer Price Index (MoM) 85 M 14 0.04 -0.26 -0.11 Retail Sales Less Autos 62 M 15 0.71 0.96 1.25 Trade Balance 81 M 41 0.21 0.87 1.19 Unemployment Rate 88 M 4 -0.92 -0.66 -0.42 Wholesale Inventories 79 M 40 0.10 0.10 0.07 GDP Annualized QoQ A 96 Q 26 2.46 2.68 2.28 GDP Annualized QoQ S 96 Q 59 -0.44 0.05 0.08 GDP Annualized QoQ T 96 Q 80 0.03 -0.91 -1.09 GDP Price Index A 77 Q 26 0.38 0.43 0.21 GDP Price Index S 77 Q 59 0.81 1.87 1.78 GDP Price Index T 77 Q 80 0.43 -1.09 -0.82 Nonfarm Productivity P 35 Q 31 -1.43 -1.95 -1.78 Nonfarm Productivity F 35 Q 65 -1.00 -0.95 -0.70 Unit Labor Costs P 27 Q 31 0.13 0.48 0.51 Unit Labor Costs F 27 Q 65 -0.10 0.22 0.34 U. of Michigan Confidence P 93 M -23 1.00 1.65 1.42 U. of Michigan Confidence F 93 M -9 0.02 -0.13 0.09 R2 daily 0.08 0.08 0.07 monthly 0.15 0.23 0.18 quarterly 0.14 0.35 0.32 Notes: The table shows the macroeconomic releases used to compute the news indices. In each case, we show the relevance index, i.e., the percentage of users who set an alert for a particular event, the frequency, the average publication delay expressed in d3a3ys, and the values of the coefficients estimated from Equation (1)fortheyieldsofbondswithmaturitiesat1,5,and10years. Thevaluesinboldaredifferentsignificantly from zero at the 5% confidence level (t-stat based on HAC standard errors). The final three rows show the R2 values obtained from: Equations (1), daily; and Equation (4), monthly (h=22) and quarterly (h=66).

Table 2: Predictive regressions r¯x66 f66 nf66 t t t const 0 0.41 -0.35 CP 1 0.27 0.74 R2 20 4 18 Notes: ThetableshowsthecoefficientsandtheR2 valuesforequation(8),wherex isinturndefinedasr¯x66, t t the 66-day holding period excess bond returns average through different maturities; f66 is its fundamental t obtained from the macroeconomic news; and nf66 is the residual part. t 34

Table 3: Effects of Macroeconomic News on Stock Prices and the Exchange Rate Releases TWEX S&P 500 Advance Retail Sales 1.6 0.1 Business Inventories -2.5 0.7 Capacity Utilization 0.0 24.5 Change in Nonfarm Payrolls 13.9 7.2 Consumer Confidence 3.7 -6.0 Consumer Credit -1.1 5.3 Consumer Price Index (MoM) -2.1 1.2 CPI Ex Food & Energy (MoM) 1.9 -14.8 Domestic Vehicle Sales 1.7 8.5 Durable Goods Orders 1.1 9.2 Employment Cost Index -5.2 -2.8 Factory Orders 4.7 -18.0 Housing Starts 0.5 7.6 Import Price Index (MoM) 3.9 -15.4 Industrial Production 0.7 -27.4 Initial Jobless Claims 1.1 -8.0 ISM Manufacturing 10.0 18.7 ISM Non-Manf. Composite -1.5 21.6 Leading Indicators 3.2 1.8 New Home Sales -2.6 -6.2 Personal Income -2.9 -4.8 Personal Spending 1.8 15.9 Philadelphia Fed. -1.6 19.8 PPI Ex Food & Energy (MoM) -9.5 -0.9 Producer Price Index (MoM) 3.5 -4.1 Retail Sales Less Autos 5.7 37.4 Trade Balance 6.5 18.1 Unemployment Rate -8.4 -0.3 Wholesale Inventories 0.8 -5.3 GDP Annualized QoQ A 18.9 -19.1 GDP Annualized QoQ S 10.9 -12.7 GDP Annualized QoQ T -11.7 -3.5 GDP Price Index A 5.5 7.7 GDP Price Index S -0.7 6.5 GDP Price Index T -4.7 -24.3 Nonfarm Productivity P -11.4 -5.7 Nonfarm Productivity F 15.0 -18.8 Unit Labor Costs P -10.2 -15.5 Unit Labor Costs F 6.7 -8.7 U. of Michigan Confidence P 3.0 -0.7 U. of Michigan Confidence F -1.4 -9.8 R2 daily 0.02 0.02 monthly 0.00 0.05 quarterly 0.00 0.15 Notes: The table shows the macroeconomic releases used to compute the news indices and the coefficients estimatedfromEquation(1)forthetrade-weightedU.S.dollarindex(TWEX)andtheSP500logdifferences. 35 The values in bold are significantly different from zero at the 5% confidence level (t-stat based on HAC standarderrors). ThefinalthreerowsshowtheR2 valuesobtainedfrom: Equations(1),daily;andEquation (4) , monthly (h=22) and quarterly (h=66)

Table 4: Effects of Macroeconomic News on Bond Yields at the Zero Lower Bound 1-year 5-year 10-year Releases β δ β δ β δ Advance Retail Sales 1.35 -0.65 1.73 -0.08 1.11 0.83 Business Inventories -0.39 0.27 -0.35 0.23 -0.24 0.33 Capacity Utilization 1.76 -1.10 2.13 -1.15 2.02 -1.37 Change in Nonfarm Payrolls 3.63 -0.45 3.87 1.20 2.84 1.60 Consumer Confidence 1.34 -0.78 1.09 -0.28 0.71 0.28 Consumer Credit -0.02 -0.09 -0.12 -0.17 -0.43 0.20 Consumer Price Index (MoM) 0.32 -0.09 0.12 0.24 -0.26 0.25 CPI Ex Food & Energy (MoM) 0.74 -0.57 1.32 -1.60 1.35 -1.74 Domestic Vehicle Sales 1.18 -0.66 0.45 -0.27 -0.14 0.33 Durable Goods Orders 0.66 -0.35 0.78 -0.06 0.67 0.01 Employment Cost Index 0.18 0.06 0.87 -1.05 0.96 -1.32 Factory Orders 0.49 -0.61 0.79 -0.95 0.98 -1.17 Housing Starts 0.28 0.13 0.06 0.96 -0.27 1.04 Import Price Index (MoM) -0.08 0.14 -0.04 0.01 -0.04 -0.25 Industrial Production -0.19 0.18 -0.11 -0.53 -0.76 -0.30 Initial Jobless Claims -1.60 0.75 -1.83 0.40 -1.50 0.10 ISM Manufacturing 2.58 -1.43 3.60 -1.34 3.07 -0.63 ISM Non-Manf. Composite 1.86 -0.66 2.18 -0.44 1.94 -0.33 Leading Indicators 0.22 -0.03 -0.13 1.14 0.24 0.93 New Home Sales 0.45 -0.16 0.62 -0.07 0.79 -0.09 Personal Income -0.77 0.45 -1.04 0.84 -1.07 1.04 Personal Spending 0.42 -0.31 0.23 -0.17 0.16 0.04 Philadelphia Fed. 1.91 -1.22 2.79 -1.28 2.08 -0.55 PPI Ex Food & Energy (MoM) 0.23 0.14 0.89 0.52 1.10 0.80 Producer Price Index (MoM) -0.06 0.23 -0.50 0.58 -0.47 0.76 Retail Sales Less Autos 1.27 -0.42 1.19 -0.24 1.27 -0.48 Trade Balance 0.18 0.05 0.85 0.07 1.11 0.13 Unemployment Rate -1.81 1.38 -1.37 1.04 -0.78 0.55 Wholesale Inventories 0.24 -0.21 -0.09 0.28 -0.23 0.46 GDP Annualized QoQ A 2.95 -1.10 3.64 -2.16 3.36 -2.33 GDP Annualized QoQ S -0.76 0.48 -1.07 1.55 -1.15 1.66 GDP Annualized QoQ T 0.40 -0.49 -0.92 0.03 -0.68 -0.55 GDP Price Index A 0.11 0.14 0.41 -0.74 0.07 -0.61 GDP Price Index S 0.50 0.35 1.17 0.69 1.01 0.76 GDP Price Index T 0.49 -0.19 -0.49 -1.23 -0.09 -1.43 Nonfarm Productivity P -2.37 1.14 -3.05 1.07 -2.95 1.06 Nonfarm Productivity F -1.72 1.54 -2.68 2.93 -2.16 2.46 Unit Labor Costs P -0.55 0.80 -1.45 2.48 -1.84 3.03 Unit Labor Costs F 0.06 -0.18 -0.69 0.99 -0.81 1.25 U. of Michigan Confidence P 1.32 -0.66 1.47 0.03 1.12 0.21 U. of Michigan Confidence F 0.14 -0.12 0.33 -0.73 0.75 -1.11 R2 daily 0.10 0.10 0.09 monthly 0.19 0.23 0.20 quarterly 0.19 0.34 0.32 Notes: The table shows the coefficients estimated from Equation (1) for the yields of bonds with maturities 36 of 1, 5, and 10 years based on the pre-zero lower bound and zero lower bound subsamples, as well as their differences. The values in bold are significantly different from zero at the 5% confidence level (t-stat based on HAC standard errors). The final three rows show the R2 values computed from the entire sample, which were obtained from: Equations (1), daily; and Equation (4), monthly (h=22) and quarterly (h=66).