The impact of macroeconomic announcements on stock prices: in search of state dependence.

1. Introduction

Federal bureaus make regularly scheduled public announcements of macroeconomic variables such as employment, the price level, and the money stock. The impact of these announcements on financial markets has received considerable attention in both the popular press and the academic literature. For instance, as noted by Krueger (1996), virtually every time since 1983 that the Bureau of Labor Statistics has released employment data, on the following day, the New York Times has cited the release as influencing the bond or stock market. The emphasis on announcement effects in the popular press, however, relies merely on anecdotal evidence. In contrast, numerous academic studies have more formally analyzed the effects of announcements by employing econometric methods. In particular, Schwert (1981) and Urich and Wachtel (1984) analyze the effects of inflation announcements, while Pearce and Roley (1985) and Jain (1988) examine the effects of several different announcement variables. Such studies typically find statistical evidence that financial markets do respond systematically to macroeconomic announcements.

Recently, a line of research has emerged that seeks to establish a relationship between the impact of scheduled announcements on financial markets and the state of the business cycle. Specifically, Boyd, Jagannathan, and Hu (2001) conclude that an announcement of an unexpected increase in the unemployment rate tends to decrease stock prices during economic contractions but to increase stock prices during economic expansions. Similarly, Adams, McQueen, and Wood (1999) maintain that the impact of an inflation announcement on prices of large cap stocks depends on whether the economy's state of economic activity happens to be high or low. McQueen and Roley (1993) find similar state dependence in the effect of several different announcement variables on the Standard and Poor's (S&P) 500. Inevitably, the idea of state dependence of announcement effects has also found its way into the popular press (Mehring 2001).

This study employs the set of variables used by McQueen and Roley to reexamine the evidence of state dependence of announcement effects. The analysis is extended in two directions. First, the data are updated by 10 years, which nearly doubles the number of available observations. The larger sample is desirable because most of the inferences in McQueen and Roley (1993) are based on relatively modest numbers of observations. Second, to establish the robustness of results, the study employs several alternative approaches to defining the state of the economy. Defining the state of the economy proves to be a rather subjective endeavor. The study finds, however, that, regardless of definition, the estimated effect of announcements on the S&P 500 does not significantly differ across economic states. This result is reinforced by the application of a general test for coefficient stability due to Nyblom (1989) and Hansen (1992). The findings cast doubt on the robustness of the results of previous studies and imply that state dependence of announcement effects cannot be considered an empirical stylized fact.

Like previous studies, we do find that announcements have a correlation with prices in financial markets that is statistically significant. But the relationship found, while statistically nonzero, is not a close one. The announcement variables we examine explain together less than 2% of the variation of daily changes in the S&P 500, and no individual variable (other than the discount rate) explains more than about 3% of the variation occurring on its own announcement days. Notwithstanding the considerable attention given to announcements, their weak explanatory power calls into question their economic significance.

The article proceeds as follows. Section 2 discusses the data and the empirical model. Section 3 explores the issue of how to define economic states. Section 4 presents the results of stability tests that use slope-shift dummy variables to define specific breakpoints. In contrast, the tests in section 5 are based on a definition of the economic state that varies continuously. Section 6 yields additional results by applying Nyblom-Hansen stability tests, which do not require specification of the economic state. Finally, section 7 provides concluding remarks.

2. Data and Empirical Model

The existing empirical literature motivates the state-contingency hypothesis by making an argument based on a positive Phillips-curve relationship between economic activity and inflation. The argument states that, when economic activity is high, an announcement that employment and production are rising implies that the economy might be overheating. This leads the public to expect higher future inflation and interest rates, which causes stock prices to fall. On the other hand, if the economy is experiencing a downturn, an announcement of rising employment and production is good news for stocks because it implies economic recovery and higher future earnings. Hence, the impact of a macroeconomic announcement on the stock market is state contingent. More formally, Veronesi (1999) develops an intertemporal, rational expectations model in which one of two possible economic states, high or low, prevails but is unobserved. Investors form a posterior probability, [pi](t), that, at time t, the high state prevails. When investors perceive a high state, [pi](t) [congruent to] 1, the arrival of bad macroeconomic news not only decreases expected future dividends but also lowers [pi](t) closer to 0.5, which implies greater uncertainty regarding the economic state. This additional source of risk means that risk-averse investors require deeper discounts in asset prices, so that asset prices fall by more than just the present value of expected dividends. In the low state, [pi](t) [congruent to] 0, uncertainty increases ([pi](t) moves closer to 0.5) upon the arrival of good news. In this case, asset prices increase by less than the increase in the present value of expected dividends. Thus, assets prices are less elastic with respect to news in the low state than in the high state.

To test the hypothesis of state contingency, I examine the effect of several different announcement variables on the S&P 500 index. Specifically, I follow McQueen and Roley (1993) by considering the following eight announcement variables: the consumer price index, the producer price index, the unemployment rate, total nonfarm employment, the index of industrial production, the U.S. trade balance on goods and services, the MI money stock, and the Federal Reserve's discount rate. Although government bureaus announce many other variables, the aforementioned variables are among the most closely watched and analyzed, and they encompass a wide spectrum of macroeconomic and policy issues. (1)

The econometric model takes the form

(1) S[P.sub.t] = [alpha] + [[alpha].sub.L][Low.sub.t] + [[alpha].sub.H][High.sub.t] + [a.sup.a.sub.t][beta] + [a.sup.u.sub.t][gamma] + [Low.sub.t][a.sup.a.sub.t][[beta].sub.L] + [Low.sub.t][a.sup.u.sub.t][[gamma].sub.L] + [High.sub.t][a.sup.u.sub.t][[beta].sub.H] + [High.sub.t][a.sup.u.sub.t][[gamma].sub.H] + [D.sub.t][delta] + [epsilon].sub.1,t].

The dependent variable, S[P.sub.t], is defined as the percentage change in the daily closing value of the S&P 500 index. To indicate whether the state of the economy is high or low, we define dummy variables [Low.sub.t] and [High.sub.t]. When these dummy variables equal zero, the state of economic activity is defined as "medium." The row vectors [a.sup.a.sub.t] and [a.sup.u.sub.t] contain the anticipated and unanticipated components of the macroeconomic announcements for day t. In order to allow the estimated effects of announcements to vary with economic states, Equation 1 includes terms that interact the announcement variables with dummy variables [High.sub.t] and [Low.sub.t]. We denote the coefficients of the announcement variables with parameter vectors [beta] and [gamma] and denote the coefficients of the interaction variables [[beta].sub.L], [[beta].sub.H], and [[gamma].sub.L], [[gamma].sub.H]. Finally, [D.sub.t] represents a vector of dummy variables indicating the day of the week, [delta] is a vector of coefficients, and [[epsilon].sub.1,t] is a random disturbance with mean zero.

Let [a.sub.i,t] denote day t's actual announcement of variable i. The announcement takes the form of a percentage change occurring since the variable's previous announcement date. We define the anticipated component of the percentage change, [a.sup.a.sub.i,t], to be the mathematical expectation of day t's actual announced percentage change, conditional on information available on day t - 1,

(2a) [a.sup.a.sub.i,t] = E[[a.sub.i,t] | [I.sub.t-1].

The difference between the actual and anticipated announcements defines the unanticipated announcement,

(2b) [a.sup.u.sub.i,t] = [a.sub.i,t] - [a.sup.a.sub.i,t].

The need to distinguish between anticipated and unanticipated components of announcements arises from the theory of efficient markets. In an efficient market, prices reflect the information set, [I.sub.t-1]. Accordingly, only information that is new on day t can move stock prices on day t. Thus, the anticipated component on day t can have no effect because it reflects only information available on day t-1, which is not news on day t. The definition of market efficiency implies that only the unanticipated component can affect stock prices.

An issue generally overlooked by the literature on announcements, however, is whether announcements could ever be unanticipated in an efficient market. According to so-called strong versions of market efficiency, market prices reflect all available information, public or private (see Ross 1987). No additional information is revealed by government announcements because they typically reflect only existing information gathered by government survey. For instance, inflation and employment announcements are computed from surveys conducted by the Bureau of Labor Statistics. Because the surveys merely assemble existing information, they create no new information and should not move stock prices. In other words, each announcement is fully contained in the information set [I.sub.t-1]. The anticipated component, E[[a.sub.i,t]|[I.sub.t-1]], therefore equals the actual announcement, and the unanticipated component is zero. Perhaps counterintuitively, an empirical finding that asset prices respond significantly to movements in a constructed series of "unanticipated" announcements does not yield prima facie evidence of market efficiency; rather, it contradicts market efficiency in its strong form. Such a result, however, can concur with a weaker version of market efficiency that accounts for information and trading costs. In this view, asset prices reflect information up to the point where the marginal benefit of additional information equals the marginal cost of obtaining and acting on the information (Jensen 1978). If information and trading costs prohibit the public from profitably replicating the government survey, then a meaningful unanticipated component of announcements can exist.

To proxy for market expectations, studies of macroeconomic announcements typically use survey forecasts obtained from a San Francisco-based firm, Money Market Services (MMS). MMS reports the median forecast from a survey of several dozen market analysts, with the survey generally conducted on the Friday of the week before the announcement. (2) Thus, for announcement day t, the survey occurs on business day t - n, where n normally varies from 1 to 5. Because Equation 2a assumes the anticipated announcement to include information available as late as day t - 1, we update the MMS data by obtaining so-called revised expectations. Following McQueen and Roley and several other studies, we define the revised expectation as the fitted value from a regression of the actual announcement on the MMS forecast and the change in the three-month T-bill rate occurring between days t - n and t - 1. (3) Our revised expectations fulfill the essential properties of rational expectations; the forecast errors have mean zero and do not exhibit significant serial correlation.

To estimate Equation 1, it remains only to determine the classification of economic states and hence to define the structural shift dummy variables [Low.sub.t] and [High.sub.t]. The next section addresses this issue.

3. Classification of Economic States

The existing literature on announcement effects follows two approaches to classifying the state of business conditions. Boyd, Jagannathan, and Hu (2001) classify economic states according to official NBER reference dates for business cycles. On the other hand, McQueen and Roley (1993) and Adams, McQueen, and Wood (1999) base their classification on the linearly detrended level of the index of industrial production. The two approaches differ fundamentally because the former is based on the economy's direction of change (first derivative), while the latter is based on the economy's level relative to trend. The two approaches therefore lead to rather different classifications of economic dates. For example, a recession can depress the level of economic activity below trend at a particular date, but if by that date, the recession has already bottomed out, the current direction of change can nevertheless be positive. Whether the economic state at that particular date is classified as low or high depends on the operative definition of economic state. With no obvious way to resolve this ambiguity, the choice of definition inevitably involves an element of arbitrariness. Furthermore, defining the economic state relative to the trend level of economic activity leads to a further question of how the trend shall be estimated. For instance, McQueen and Roley (1993) and Adams, McQueen, and Wood (1999) use the ordinary least squares (OLS) fit of a linear time trend. The problem with linear detrending, however, is that it is not well supported by the data.

Using data on the monthly index of industrial production for our sample period, January 1980 through December 1998, we obtain the following OLS estimates:

(3) log(INDPRO[D.sub.t]) = 4.32 + 0.00229[TIME.sub.t] (0.02) (0.00017) N = 228; [R.sup.2] = 0.939; D.W. = 0.032

Parentheses contain standard errors computed from a Newey-West (1987) covariance matrix that is robust to serial correlation. (4) The estimates imply a long-term growth rate of industrial production equal to 0.23% per month. The problem, however, is that industrial production grew considerably faster in the 1990s than in the 1980s. To illustrate, we split the sample at its midpoint (June-July 1989) and estimate separate linear trends over each subsample. For the first half of the sample, the estimated monthly trend growth rate of industrial production is 0.23%, but for the second half of the sample, it equals 0.32%. Using the Newey-West covariance matrix, we compute a Wald statistic for testing the null hypothesis that the same linear trend applies to both subsamples. The resulting chi-square statistic with 2 d.f. equals 8.28; hence, we can reject at the 2% significance level the hypothesis that the coefficients in Equation 3 are stable over the whole sample. Note also the extremely low value of the Durbin-Watson statistic, which further suggests that the model does not efficiently utilize the information in the sample.

Apparently, simple linear detrending is too rigid and restrictive to adequately model the long-term process determining industrial production, and this accounts for its lack of recent popularity among macroeconomists. Many alternatives to linear detrending exist, but a flexible and widely used technique is the Hodrick-Prescott (1997) filter. For instance, the literature on real business cycles routinely uses the Hodrick-Prescott (HP) filter to detrend macroeconomic series. The HP filter estimates the trend path of a series by fitting a smooth, nonlinear curve to the data. The procedure is more general than linear detrending; in fact, the HP procedure encompasses linear detrending as a special case. (5)

Figure 1 plots the detrended, that is, cyclical, components of the log of industrial production as estimated by HP filtering and by the linear trend in Equation 3. The figure also indicates the thresholds for classifying the state of the economy as "low," "medium," or "'high." Following McQueen and Roley, we categorize a period's state of economic activity as high if that period's cyclical component of industrial production is among the largest 25% in the sample. In our sample, this definition implies a cutoff of 0.027 for cyclical industrial production to qualify as high. Then we make the cutoff symmetric by defining the economic state as low if cyclical industrial production falls below -0.027 (although this assigns somewhat fewer than 25% of the observations to the low category). Applying the same procedure to the HP-filtered series yields a smaller cutoff of [+ or -] 0.0076 because the HP procedure's flexibility permits a closer fit to the data. Depiction of the HP filtered series in Figure 1 is scaled to make its cutoffs coincide with those of the linearly detrended series.

[FIGURE 1 OMITTED]

Figure 1 reveals that the two detrending methods yield cyclical industrial production series that are quite different. As a consequence, the methods lead to rather different classifications of economic states. Table 1 provides complete lists of the particular months that each method designates as high or low, with the other months falling implicitly in the medium category. As Table 1 and Figure 1 show, the two methods concur that the 1990 recession depresses the economy to the low state around November or December 1990. According to the HP series, the economy recovers to the medium state as early as June 1991. The linearly detrended series, however, languishes continually in the low state for almost three more years, until May 1994. Furthermore, the rapid economic expansion of the late 1990s propels the linearly detrended series into the high category for all months after January 1997. But because this rapid growth can be accommodated by the HP procedure's greater flexibility, the HP series attains the high category only for the period September 1997 to January 1998. Indeed, for the months of July and December 1998, the HP procedure designates the economic state as low, a conclusion diametrically opposed to that of linear detrending.

Table 1 also displays the official NBER dates of U.S. economic contractions. Clearly, the choice of NBER reference dates, linear filtering, or HP filtering has a considerable impact on the categorization of economic dates. The expression "state of the economy" fails to supply an objective, nonarbitrary way to classify economic dates. Hence, considerable ambiguity surrounds the definitions of dummy variables such as [High.sub.t] and [Low.sub.t] for modeling the timings of structural breakpoints in the marginal effects of announcements. To demonstrate the robustness of results and to facilitate comparison with previous studies, we obtain estimates of Equation 1 with timings of structural shifts defined using each of the three alternative methods. The next section discusses the results.

4. Tests for Structural Shifts at Specific Breakpoints

To examine the hypothesis that announcement effects are state contingent, we obtain OLS estimates of Equation 1 that allow for coefficient shifts at specific breakpoints. The coefficient shifts are modeled by interacting announcement variables with economic-state dummy variables. Table 2 presents Wald test statistics derived from four different specifications of Equation 1. Model (a) imposes constant marginal effects of announcements across economic states ([[beta].sub.L], [[beta].sub.H], and [[gamma].sub.L], [[gamma].sub.H] are set to zero), while models (b), (c), and (d) permit coefficients to shift across economic states. Model (b) employs the classification of economic states derived from linearly detrending the log of industrial production, as in McQueen and Roley (1993). Model (c) uses the classification obtained from HP filtering, as described in the previous section. Model (d) allows for coefficient shifts across NBER-defined expansions and contractions, as in Boyd, Jagannathan, and Hu (2001). Because the estimated models exhibit significant evidence of heteroscedasticity, we compute Wald test statistics from a heteroscedasticity-consistent covariance matrix recommended by MacKinonn and White (1985). (6)

According to the Wald chi-square statistics and p-values displayed in Table 2, the unanticipated announcements consistently demonstrate joint significance hut the anticipated announcements do not. For instance, in the model (a) that imposes constant marginal effects, the unanticipated announcement variables are jointly significant at the 0.01 level. The unanticipated announcements attain joint significance in the structural shift models as well, although only at the 0.10 level in models (b) and (c). (For models [b] and [c], the joint hypothesis involves a relatively large number of coefficients (24), which perhaps lowers the power of the test). In contrast, the announcements that are anticipated achieve significance at conventional levels only in model (b). Hence, we can generally reject [gamma] = [[gamma].sub.L] = [[gamma].sub.H] = 0, but not [beta] = [[beta].sub.L] = [[beta].sub.H] = 0. (7)

To evaluate the hypothesis of state contingency of announcement effects, we test the null hypothesis that unanticipated announcement variables have coefficients that are constant across economic states: [[gamma].sub.L] = [[gamma].sub.H] = 0. Table 2 presents Wald chi-square statistics and p-values. None of the reported p-values are lower than 0.31, so none of the three structural shift models can reject at conventional levels the hypothesis of constant marginal effects.

An additional test that might have greater power against the null hypothesis can be applied to models (b) and (c). In these models, even if the estimated marginal effects in the low and high states do not differ significantly from those in the medium state, they can nonetheless differ significantly from each other. Indeed, McQueen and Roley obtained their strongest statistical results by comparing coefficients in high and low states. They found in particular that for real variables--industrial production and the unemployment rate--the null hypothesis of equal coefficients in high and low states was rejected at about the 1% level. Consequently, we test the hypothesis that marginal effects are equal in high and low states, [[gamma].sub.L] - [[gamma].sub.H] = 0, and display Wald chi-square statistics and p-values in Table 2. The p-value for the joint test of all announcements is 0.38 for model (b) and 0.53 for model (c). Again, we cannot reject constant marginal effects. To focus on the McQueen-Roley result, we also present the joint test of the coefficients on the unemployment and industrial production variables. For model (b), the p-value of 0.11 is perhaps suggestive but not significant at conventional levels. Moreover, for model (c), the p-value equals 0.39, which does not approach conventional levels. Thus, we conclude that tests of joint hypotheses contradict the notion of state contingency of announcement effects, and this result is robust across model specifications.

State contingency is also rejected by significance tests of individual coefficients. Tables 3-5 display estimated coefficients for each of the announcement variables, including the economic-state interaction variables. Because theory predicts the announcements to have no effect if anticipated and because anticipated announcement variables fail to consistently demonstrate statistical significance, we focus on a more parsimonious version of Equation 1 that includes only unanticipated announcements. Consequently, the estimates in Tables 3-5 are based on models of the following form:

(4) S[P.sub.t] = [alpha] + [[alpha].sub.L][Low.sub.t] + [[alpha].sub.H][High.sub.t] + [a.sup.u.sub.t][gamma] + [Low.sub.t][a.sup.u.sub.t][[gamma].sub.L] + [[High].sub.t][a.sup.u.sub.t][[gamma].sub.H] + [D.sub.t][delta] + [[epsilon].sub.2,t].

Tables 3-5 display estimated models that use the classifications of economic states obtained from linear filtering, HP filtering, and NBER reference dates, respectively. Several announcements demonstrate statistical significance, including nonfarm payrolls, the consumers' price index (CPI), the producer price index (PPI), and the discount rate. But the coefficient estimates for the interaction terms reveal that no announcement exhibits a structural shift that attains significance at the 0.05 level across models. Parameter constancy can be rejected at the 0.05 level for the trade balance, but only for the breakpoints determined by HP filtering. When the breakpoints are determined by the methods used in previous studies--linear filtering or NBER reference dates--parameter constancy cannot be rejected at the 0.05 level for any of the individual announcements. In sum, the estimates yield little evidence of structural shifts, and this result is robust across three different methods of determining specific breakpoints. In the next section, we present additional evidence by considering general tests of coefficient stability that do not require specification of breakpoints.

5. A Continuous Economic State Defined by Contemporaneous Market Prices (8)

Our three aforementioned methods of classifying states all use ex post data, that is, data not available to agents at the time they form their expectations. If the ex post model of the economic state (e.g., Equation 3), is estimated imprecisely or if the model is misspecified, the resulting inferences can be misleading. An alternative approach uses ex ante data available to agents as they make decisions in real time. In this regard, a convenient ex ante measure of the state of business conditions that is available on a daily basis is the default premium on bonds. Indeed, the default premium is closely cyclical (Chen, Roll, and Ross 1986), and it has predictive power for returns on stocks, not just bonds (Fama and French 1989).

Another advantage of the default premium is that it correlates with the conditional variance of stock returns. The volatility of returns creates a state of uncertainty. In theory, this state of uncertainty might provide the rationale for announcement effects to be state dependent. The announcement effect can depend on the volatility of returns if, for instance, investors are guided by a Sharpe ratio. Evidence indicates the conditional variance of stock returns and the default premium are both countercyclical (Harrison and Zhang 1999). Furthermore, the two are causally related because increased volatility during downturns causes a flight to relatively safe assets, increasing the default premium.

We therefore perform additional tests of the state dependence of announcement effects by defining the relevant state as a continuous function of the contemporaneous default risk premium. Specifically, for day t. let the variable DE[F.sub.t] represent the difference between the yield on Baa and Aaa corporate bonds. The econometric model takes the form

(5) S[P.sub.t] = [[alpha].sub.1] + [[alpha].sub.1]DE[F.sub.t] + [[alpha].sup.u.sub.t] + DE[F.sub.t][a.sup.u.sub.t][[gamma].sub.2] + [D.sub.t][delta] + [[member of].sub.e,t].

The interaction terms allow announcement effects to vary continuously with DE[F.sub.t]. In case DE[F.sub.t] is unduly noisy or is itself influenced by the contemporaneous announcement, we also estimate a version of Equation 5 with a smoothened DE[F.sub.t] series (consisting of a two-sided, five-period moving average). The resulting estimates of Equation 5 support the previous inferences. In particular, a chi-square test of the null of no state dependence yields a p-value of 0.632 with DE[F.sub.t] smoothened and 0.627 with DE[F.sub.t] not smoothened. Hence, we cannot reject the null at conventional levels.

6. Tests for Structural Shifts with Breakpoints Unspecified 

Testing for structural shifts at specific breakpoints can be problematic if the timing of the breakpoints is not known a priori. As pointed out in section 3, considerable ambiguity surrounds the issue of how to estimate the timing of changes in the state of the economy. Even if the appropriate estimating technique were known, the timings of shifts could only be estimated, not known with certainty. If the specified timings of structural change differ from the true timings, the stability tests can lack power against the null hypothesis of parameter constancy. On the other hand, attempts to find the true timings through various forms of pretest analysis such as visual inspection, estimation over various subsamples, or searches over a variety of structural breakpoints, are tantamount to data mining and can lead to overrejection of the null hypothesis.

When the timing of potential structural shifts is uncertain, a useful approach is to apply a general test of coefficient stability. Recently, Nyblom (1989) and Hansen (1992) have developed stability tests that do not require specification of the timings of shifts. The Nyblom-Hansen tests evaluate the null hypothesis of coefficient stability against the alternative hypothesis that at least one coefficient does shift at some unspecified breakpoints. The tests are not designed for determining the timings of structural breaks; they merely test the null of parameter constancy. The Nyblom-Hansen tests are based on cumulative sums (CUSUM) of residuals, similar to CUSUM and CUSUM of squares tests, but the Nyblom-Hansen tests have asymptotic local power against a wider and more useful range of alternatives. In fact, the tests have locally optimal power, and they are asymptotically robust to heteroscedasticity.

To apply the tests, we impose the null hypothesis of constant coefficients on Equation 4. Setting the structural shift coefficients ([[alpha].sub.L], [[alpha].sub.H], and [[gamma].sub.L], [[gamma].sub.H]) equal to zero yields the following estimating equation:

(6) S[P.sub.t] = [alpha] + [a.sup.sub.t] [gamma] + [D.sub.t][delta] + [[member of].sub.4,t].

The resulting Nyblom-Hansen statistic for the joint stability of the coefficients [gamma] of the unanticipated announcements equals 1.52, which is not large enough to reject the null of coefficient stability at conventional levels. When we add to Equation 6 the anticipated announcements, the test statistic for the joint stability of all 15 announcement variables, anticipated and unanticipated, equals 3.29. Again, this falls short of the 0.05 critical value of 3.54. (9)

We can also use the Nyblom-Hansen procedure to test the stability of coefficients individually. The individual coefficient tests evaluate the null hypothesis of stability of a single coefficient under the assumption that all other coefficients in the model are stable. Table 6 displays the individual test statistics for each of the unanticipated announcement coefficients as estimated from Equation 6. All eight test statistics fall short of the 0.05 critical value of 0.470. Hence, the Nyblom-Hansen results suggest that no structural breakpoints exist in the sample.

To formulate a stability test with possibly greater power than the Nyblom-Hansen test would require considerable information regarding the timings of potential structural shifts. But as shown in section 3, the timings of economic states involve considerable uncertainty. Among stability tests that do not require specification of breakpoints, the Nyblom-Hansen tests have asymptotically optimal power. The Nyblom-Hansen tests therefore provide strong evidence that our data contain no significant evidence of coefficient instability. To quote Hansen (1992, p. 532):

   If the test statistics are insignificant, then the investigator
   can be reasonably confident that either the model has been constant
   over that sample or the data is [sic] not sufficiently informative
   to reject this hypothesis.

The Nyblom-Hansen tests thus reinforce the results of the Wald stability tests of section 4. Apparently, the data in question yield no significant evidence of state dependence of announcement effects.

7. Concluding Remarks

This study's findings differ from those of previous studies, including Boyd, Jagannathan, and Hu (2001) and McQueen and Roley (1993). The latter study considers the same set of announcements as does the present study and likewise proxies for market expectations by using Money Market Services (MMS) surveys. The study by McQueen and Roley, however, covers only about 10 2/3 years, whereas the present study covers 19 years. The difference is important, because the inferences for most announcement variables are based on relatively modest numbers of observations. For instance, McQueen and Roley's estimating equation uses 932 announcement day observations, but each of those days typically records announcements for only one or two of the announcement variables. The variables not announced on a particular announcement day take the value zero. The money stock announcement is nonzero for most of the 932 days because the announcement was made weekly. But the remaining variables were announced no more than once per month. The monthly announcement variables such as unemployment, industrial production, and the CPI, each possess only about 127 nonzero observations. Moreover, by definition, the slope-shift dummy variables [Low.sub.t] and [High.sub.t] take nonzero values over only about 25% of the sample. This implies that inferences regarding the state dependence of the monthly announcements are based on only about 30 observations per announcement. By comparison, the inferences from the present study are typically based on nearly twice as many observations because the overall number of available announcement days (1739) is larger by about 87%.

Unlike most studies of announcement effects, the study by Boyd, Jagannathan, and Hu (200l) does not use MMS surveys; instead, the authors proxy market expectations by estimating a forecasting model. They analyze a relatively long time series, reaching back to 1948, but they consider announcements of only a single variable, the unemployment rate. A cause for concern is that relevant information about labor markets can also reside in the level of total nonfarm employment, which is announced simultaneously with the unemployment rate. Ignoring the announcement of nonfarm employment creates the potential for omitted variable bias. In fact, we find that nonfarm employment is statistically significant in most models, while the unemployment rate is never significant. This finding concurs with those of Krueger (1996) and Balduzzi, Elton, and Green (2001).

As far as the state contingency of employment news is concerned, the present study finds the evidence to be weak and not robust. Our estimated marginal effect of nonfarm employment is significantly more positive during recessions, but only at the 0.10 level, not at the 0.05 level. Even this weak result no longer holds if the model defines the economic state relative to the level of economic activity. Turning to the unemployment rate, a stability test for this variable yields a Nyblom-Hansen statistic of 0.376, which rejects stability at the 0.10 level, but not at the 0.05 level. But if the impact of the unemployment announcement does experience structural shifts, when these shifts occur is far from clear because none of the four economic-state models reveal any significant evidence of structural shifts. Furthermore, if the impact of employment news were state contingent, this property should be reflected not just by the unemployment rate but also by total employment. Yet the Nyblom-Hansen statistic for nonfarm employment does not reject stability at conventional levels.

As noted above, announcements that are anticipated generally do not have significant explanatory power for changes in the S&P 500. This finding concurs with theories of market efficiency. The results, however, indicate that announcements do have statistically significant explanatory power if unanticipated, which contradicts strong (full informational) efficiency. But the evidence might nonetheless be consistent with semistrong efficiency, which requires merely the absence of unexploited profit opportunities. Correlation of announcements with stock prices can concur with semistrong efficiency so long as agents cannot profitably replicate the government survey in advance of the announcement. The profitability of replication depends on the cost of conducting the survey, the survey's predictive power, the cost of acting on the information, and so on. Unfortunately, the issue of profitability cannot be addressed by the literature on announcement effects. Studies of announcements and security prices can establish only statistical correlation, with no obvious way to specify a loss function for determining how much correlation might be too much to be consistent with semistrong efficiency. The dilemma provides a classic example of McCloskey's (1985) distinction between statistical significance and economic significance.

Unanticipated announcements explain a fraction of daily variation in stock prices that is statistically nonzero yet rather small. For instance, as indicated by the multiple [R.sup.2] reported in Table 6, the eight unanticipated announcement variables together explain only about 1.6% of the variation in daily closing values of the S&P 500. Table 6 also reports partial [R.sup.2]s for each of the individual announcement variables. We compute each variable's partial [R.sup.2] using only the observations for days on which the particular variable was announced (or the following market days if the variable was announced after hours). (10) With the exception of the discount rate, our computations indicate that no individual variable explains as much as 3% of the variation occurring on its own announcement days. In light of this result, the considerable press coverage devoted to government announcements appears misplaced. (11)

Exceptionally, the change in the discount rate explains more than 9% of the S&P variation occurring on its announcement days. (12) In this respect, our results resemble those of Pearce and Roley (1985), who consider a similar set of announcement variables and find that only the discount rate consistently demonstrates significance. The result is intriguing because the discount rate also happens to be the only variable in the set that is not produced by government survey. Instead of reflecting only preexisting information, the change in the discount rate presumably reflects the judgment and policy discretion of the authorities at the Federal Reserve. Hence, unlike other announcements, changes in the discount rate can provide an exclusive source of news regarding discretionary shifts in public policy. This distinction might account for the discount rate's relatively greater explanatory power.

Table 1. Classification of States of the Economy

                                      Economic State

Methodology                 Low                         High

Linear         1982: May-December       1980: January-May,
  detrending                                  September December
  ([lambda] =  1983: January-July       1981: January-October
  [infinity]   1990: November-December  1984: May-August
               1991: January-December   1987: December
               1992: January-December   1988: April, July-August,
                                              November-December
               1993: January-December   1989: January-April
               1994: January-April      1997: February-December
                                        1998: January-December
Hodrick-       1980: May-September      1980: January-March,
  Prescott                                    November-December
  filter       1982: January,           1981: January-November
  ([lambda]          May-December
  = 4800)      1983: January-June       1982: February
               1986: March-November     1983: September December
               1987: January            1984: January-November
               1990: December           1985: February-March
               1991: January-May,       1987: December
                     December
               1992: January            1988: August, November-December
               1993: August             1989: January, March-May
               1996: January, March     1990: March, May-October
               1998: July, December     1994: December
                                        1995: January-February
                                        1997: September-December
                                        1998: January
                 Economic Contraction

NBER turning   1980: February-July
  points       1981: August-December
               1982: January-November
               1990: August-December
               1991: January-March

The linear detrending method classifies the state of the economy
as high (low) if the log of industrial production exceeds (falls
short of) its trend by at least 0.027. The trend is defined as the
OLS fit of a linear time trend to the log of the index of industrial
production from January 1980 to December 1998. Alternatively, the
trend component is estimated by applying a Hodrick-Prescott filter
with smoothing parameter [lambda] = 4800. Using the Hodrick-Prescott
filter, the state of the economy is high (low) if the log of industrial
production exceeds (falls short of) its trend by at least 0.0076. The
cutoff values of 0.(1076 and 0.027 are chosen so that the high state
occurs for 25% of the observations. Economic contractions are
identified by referring to official NBER turning points for U.S.
business cycles.

Table 2. Wald Test Statistics

                                    Chi-Square Test Statistic
                                  [Degrees of Freedom] (p-Value)

Null Hypothesis            Model (a)  Model (b)  Model (c)  Model (d)

Coefficients equal zero,   11.4 [7]   35.7 [21]  23.3 [21]  19.0 [14]
  anticipated               (0.122)    (0.024)    (0.329)    (0.166)
  announcement variables
Coefficients equal zero,   21.0 [8]   33.5 [24]  33.4 [24]  31.2 [16]
  unanticipated             (0.007)    (0.094)    (0.095)    (0.013)
  announcement variables
Coefficients equal zero,   33.9 [15]  76.8 [45]  57.3 [45)  52.8 [30]
  all announcement          (0.004)    (0.002)    (0.104)    (0.006)
  variables
Coefficients constant         --      14.1 [16]  12.9 [16]  9.34 [8]
  across economic states,              (0.591)    (0.680)    (0.314)
  unanticipated
  announcement variables
Coefficients equal in         --      8.53 [8]   7.09 [8]      --
  high and low states,                 (0.383)    (0.527)
  unanticipated
  announcement variables
Coefficients equal in         --      4.35 [2]   1.87 [2]
  high and low states,                 (0.114)    (0.393)
  real unanticipated
  announcement variables

Model (a) restricts coefficients to be constant across economic states;
models (b), (c), and (d) include dummy variables and interaction terms
to allow for coefficient shifts across economic states. Model (b)
classifies economic states relative to the least squares fit of a
linear trend of the monthly log of industrial production. Model (c)
classifies economic states relative to the trend component of
industrial production as estimated by a Hodrick-Preston filter. Model
(d) classifies economic states according to official NBER reference
dates for U.S. business cycles. Parentheses contain p-values; brackets
contain degrees of freedom (d.f.). Wald statistics are computed from
the heteroscedasticity-consistent covariance matrix recommended by
MacKinnon and White (1985).

Table 3. Estimated Effects of Government Announcement Surprises
1739 Available Announcement Day Observations, January 1980
through December 1998. Economic States Classified as High or
Low Relative to the Least Squares Fit of a Linear Trend to
the Monthly Log of Industrial Production

                                           Interaction Terms

                                      Announcement   Announcement
Announcement                              High           Low

                            (a)           (b)            (c)

Industrial production    -0.162         0.104          0.674
                         (0.220)       (0.484)        (0.494)
Unemployment rate         0.0016        0.0049        -0.0119 *
                         (0.00467)     (0.0120)       (0.0070)
Money stock (M1)          0.004         0.041         -0.157
                         (0.103)       (0.162)        (0.175)
Nonfarm payrolls         -2.26 *       -0.44           2.42
                         (1.28)        (2.97)         (1.78)
CPI                      -0.812 *      -0.054         -0.417
                         (0.469)       (0.850)        (0.963)
PPI                      -0.666 *       0.442          0.324
                         (0.368)       (0.698)        (0.499)
Trade balance             2.95         -2.96          -3.99
                         (1.86)        (4.32)         (3.65)
Discount rate            -0.00602 **   -0.00325        0.00647
                         (0.00258)     (0.00682)      (0.00742)

Summary statistics        [R.sup.2] = 0.024      DW = 1.94

Dependent variable: Daily percentage change in the closing value
of the S&P 500 index. Each explanatory variable is defined as the
part of the percentage change in the announcement that is
unanticipated, except the discount rate variable, which represents
the absolute change in the rate.

Announcement * High and Announcement * Low stand for explanatory
variables that interact government announcement surprises with
dummy variables High and Low. The dummy variable High takes the
value one for months when the log of industrial production exceeds
its trend by at least 0.027; the dummy variable Low takes the value
one if the log of industrial production falls short of its trend by
at least 0.027. The cutoff value of 0.027 was chosen so that High
equals one for 25% of the observations. The trend is defined as the
OLS fit of a linear time trend to the log of the index of industrial
production from January 1980 to December 1998.

Parentheses contain asymptotic standard errors. The standard
errors are computed from the heteroscedasticity-consistent
covariance matrix recommended by MacKinnon and White (1985).

* and ** Indicate statistical significance
at the 0.10 and 0.05 levels, respectively.

Table 4. Estimated Effects of Government Announcement Surprises on
Stock Prices: 1739 Available Announcement Day Observations, January
1980 through December 1998. Economic States Classified as High or
Low Relative to the Trend Component of Industrial Production as
Estimated by a Hodrick-Prescott filter

                                                 Interaction Terms

                                           Announcement    Announcement
Announcement                                  x High          x Low

                             (a)               (b)             (c)

Industrial production       0.005             0.101          -0.184
                           (0.232)           (0.426)         (0.517)
Unemployment rate          -0.0012            0.0080         -0.0051
                           (0.0063)          (0.0107)        (0.0086)
Money stock (MI)           -0.008            -0.023          -0.015
                           (0.011)           (0.163)         (0.202)
Nonfarm payrolls           -1.48             -0.37           -1.06
                           (1.08)            (2.49)          (4.23)
CPI                        -0.386            -0.996          -0.830
                           (0.565)           (0.797)         (0.969)
PPI                        -0.389            -0.142          -0.154
                           (0.350)           (0.706)         (0.561)
Trade balance               4.75 *           -8.70 **        -0.32
                           (2.67)            (3.54)          (4.34)
Discount rate              -0.00762          -0.00156         0.00571
                           (0.00517)         (0.00677)       (0.00710)

Summary statistics            [R.sup.2] = 0.024      DW = 1.95

Dependent variable: Daily percentage change in the closing value of
the S&P 500 index. Each explanatory variable is defined as the part
of the percentage change in the announcement that is unanticipated,
except the discount rate variable, which represents the absolute
change in the rate.

Announcement High and Announcement Low stand for explanatory variables
that interact government announcement surprises with dummy variables
High and Low. The dummy variable High takes the value one for months
when the log of industrial production exceeds its trend by al least
0.0076; the dummy variable Loa, takes the value one if the log of
industrial production falls short of its trend by at least 0.0076.
The cutoff value of 0.0076 was chosen so that High equals one for 25%
of the observations. The trend component is defined by a
Hodrick-Preston filter with smoothing parameter [lambda] = 4,800.

Parentheses contain asymptotic standard errors. The standard errors
are computed from the heteroscedasticity-consistent covariance matrix
recommended by MacKinnon and White (1985).

* and ** Indicate statistical significance at the 0.10 and 0.05
levels, respectively.

Table 5. Estimated Effects of Government Announcement Surprises on
Stock Prices: 1739 Available Announcement Day Observations, January
1980 to December 1998. Economic Recessions Classified According to
Official NBER Reference Dates

                                                 Interaction Term

                           Announcement      Announcement x Recession

                                (a)                    (b)

Industrial production         -0.113                  0.359
                              (0.230)                (0.497)
Unemployment rate              0.00086               -0.00348
                              (0.00444)              (0.00946)
Money stock (MI)               0.028                 -0.279
                              (0.077)                (0.199)
Nonfarm payrolls              -2.05 **                6.46 *
                              (0.96)                 (3.32)
CPI                           -0.937 **              -0.001
                              (0.384)                (0.926)
PPI                           -0.604 **               0.982
                              (0.267)                (0.791)
Trade balance                  2.48                  -4.91
                              (1.61)                 (4.86)
Discount rate                 -0.00788 **             0.00360
                              (0.00333)              (0.00591)

Summary statistics       [R.sup.2] = 0.023           DW = 1.95

Dependent variable: Daily percentage change in the closing value of
the S&P 500 index. Each explanatory variable is defined as the part
of the percentage change in the announcement that is unanticipated,
except the discount rate variable, which represents the absolute
change in the rate.

Announcement x Recession stands for explanatory variables that
interact government announcement surprises with the dummy variable
Recession. The dummy variable Ret ession takes the value one on NBER
reference dates for U.S. business cycle contractions.

Parentheses contain asymptotic standard errors. The standard errors
are computed from the heteroscedasticity-consistent covariance matrix
recommended by MacKinnon and White (1985).

* and ** Indicate statistical significance at the 0.10 and 0.05 levels,
respectively.

Table 6. Estimated Effects of Government Announcement Surprises on
Stock Prices: 1739 Available Announcement Day Observations, January
1980 through December 1998. Coefficients held constant across
economic states

                                                Nyblom-
                                                Hansen      Partial
Announcement                   Coefficient     Statistic   [R.sup.2]

Industrial production             -0.028         0.267    <[10.sup.-3]
                                  (0.176)
Unemployment rate                  0.0001        0.376    <[10.sup.-3]
                                  (0.004)
Money stock (M 1)                 -0.0115        0.087    <[10.sup.-3]
                                  (0.0715)
Nonfarm payrolls                  -1.72 *        0.065    0.027
                                  (0.940)
CPI                               -0.960 ***     0.068    0.027
                                  (0.353)
PPI                               -0.472 *       0.190    0.015
                                  (0.249)
Trade balance                      1.42          0.001    0.004
                                  (1.59)
Discount rate                     -0.00546 **    0.049    0.092
                                  (0.00260)
Summary statistics     Multiple                DW = 1.95
                       [R.sup.2] = 0.016
Nyblom-Hansen test
 of overall model
                         Unanticipated announcements      1.52
                             All announcements            3.29

Dependent variable: Daily percentage change in the S&P 500 index.

Parentheses contain asymptotic standard errors. The standard errors
are computed from the heteroscedasticity-consistent covariance matrix
recommended by MacKinnon and White (1985).

The partial [R.sup.2] represents the fraction of variation in the
dependent variable that is explained by each announcement variable,
after accounting for the influence of the other announcement
variables. For each announcement variable, the computation of the
partial [R.sup.2] uses only the observations for the days on which
the variable was announced.

The 0.05 critical value for the Nybloin-Hansen test of the stability
of a single coefficient is 0.470. For the joint tests of stability,
the 0.05 critical values are 3.54 for the full set of announcements,
anticipated and unanticipated, and 2.11 for the model that includes
only unanticipated announcements.

*, **, and *** Indicate statistical significance at the 0.10, 0.05,
and 0.01 levels, respectively.

The author thanks Tyler Cowen, Kevin Grier, Grant McQueen, and two anonymous referees for helpful comments. John Ruggiero deserves special thanks for invaluable assistance in compiling the data.

(1) Considering a broad set of announcements can also help to avert a problem of omitted variables. During a given time period, more than one macroeconomic variable can be announced. If one or more of the announced variables is excluded from the model and there is correlation between excluded and included variables, this can lead to biased estimation. For instance, as a rule. the unemployment rate is reported at the same time as total nonfarm employment. Employment and unemployment are correlated, but if the size of the labor force changes, the two variables can nonetheless reflect different labor market effects.

(2) Note that, for my sample, no MMS survey is available for the discount rate. McQueen and Roley (1993) address this issue by simply letting the discount rate's actual announced change proxy for the unanticipated change. This article follows the same practice; that is, for the discount rate, we set [a.sup.u.sub.t] = [a.sub.t]. I thank Carl Chen for providing a time series of discount rate announcements.

(3) The regression model takes the form [a.sub.i,t] = [b.sub.0] + [b.sub.1]MM[S.sub.i,t-n] + [b.sub.2]([r.sub.t-1] - [r.sub.t-n]) + [e.sub.t], where [a.sub.i,t], is the actual announcement of variable i on day t, MM[S.sub.i,t-n] is the survey forecast, ([r.sub.t-1] - [r.sub.t-n]) is the change in the 90-day T-bill rate between the survey date t - n and day t-1, and [e.sub.t] is a random disturbance. These regression estimates are available on request. In nearly all regressions, the estimate of the coefficient be does not significantly differ from zero. This result concurs with those of Balduzzi, Elton, and Green (2001), and implies that the T-bill rate does not contain significant information for revising MMS forecasts, notwithstanding the established practice in the literature. Also, for some announcements, the estimate of [b.sub.1] significantly exceeds unity. This result implies that. while the forecasts may be unbiased in mean, the changes in the forecasts tend to understate the changes in announcements. Krueger (1996) reaches the same conclusion regarding the MMS forecasts of employment variables. Computing the revised expectations, however, corrects this systematic bias.

(4) My version of the Newey-West matrix uses Bartlett weights with lags up to 24 months.

(5) The HP filter estimates the trend path by minimizing the sum of squared errors subject to the constraint that the bend in the curve, measured by the sum of squared second differences along the curve, not be too large. The optimization problem is given by

[MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

where [[mu].sub.t] denotes the fitted trend and [lambda] represents the smoothness parameter that penalizes curvature. Hodrick and Prescott (1997) recommend setting [lambda] = 1600 for quarterly data, which implies [lambda] = 4800 for monthly data. Our procedure therefore uses [lambda] = 4800.

(6) Applying a Breusch-Pagan heteroscedasticity test to model (b), for instance, yields a chi-square statistic of 242 with 51 d.f. Because the 0.01 critical value is only 77.4, the test easily rejects the null hypothesis of homoscedasticity. The heteroscedasticity-consistent covariance matrix takes the form [(X'X).sup.-1] X'[OMEGA]X[(X'X).sup.-1]. The MacKinnon-White estimator defines [OMEGA] to be an n X n diagonal matrix with ith element equal to [u.sup.2.sub.i]/[(1-[h.sub.i]).sup.2], where [u.sub.i] is the OLS residual and [h.sub.i] is the ith diagonal element of the least squares projection matrix, X[(X'X).sup.-1]X'. MacKinnon and White (1985) conducted a Monte Carlo analysis of the finite sample properties of this estimator and several alternative heteroscedasticity-consistent covariance estimators and found that this estimator performed best.

(7) For model (d), these hypotheses involve only two thirds as many coefficients because model (d) defines only two possible states: NBER-defined recession or expansion. If we consider recession to be the low state and expansion to be the control state, the hypotheses become [gamma] = [[gamma].sub.L] = 0 and [beta] = [[beta].sub.L] = 0. Similarly, the hypothesis of constant marginal effects can be written [[gamma].sub.L] = 0. This explains why Table 2 reports chi-square tests for model (d) that have only two thirds as many degrees of freedom.

(8) This section of the article relies heavily on the comments of two anonymous referees.

(9) The distributions of the test statistics are nonstandard. Hansen (1992) presents asymptotic critical values.

(10) This definition of the partial [R.sup.2] departs from the conventional definition, which involves all observations used to estimate the regression model. Our definition focuses on the individual variable's announcement days because each announcement variable obviously has no explanatory power for days on which it was not announced. For each announcement variable, all of its variation occurs on its announcement days because it is constant (zero) on nonannouncement days and has the same mean (zero) on both announcement and nonannouncement days. The different definition explains why our computed partial [R.sup.2]s in some cases exceed the model's multiple [R.sup.2], in contradiction of the usual inequality.

(11) I also explore the possibility that announcements exhibit greater explanatory power at a lower frequency. To this end, we estimate a version of Equation 6 based on weekly data. The resulting [R.sup.2] was 0.045 compared with the 0.016 obtained at a daily frequency. Although higher, the [R.sup.2] still reflects relatively little explanatory power. Moreover, the inference regarding the state dependence of announcement effects also remains unaltered.

(12) As noted by an anonymous referee, the Fed in 1994 adopted the practice of announcing discount rate changes at the same time as changes in fed funds rate targets. Hence. the significant estimated response from the discount rate might actually be driven by the closely correlated effect of the fed funds rate. To explore this, we reestimate Equation 6 using only the pre-1994 data, which contain 33 of the sample's 40 discount rate changes. Thus, the Fed's change in practice applies only to the last 7 discount rate changes. Using the 33 earlier discount rate changes yields a coefficient estimate for the discount rate variable that is very close to that obtained from the full model: -0.00556 compared with -0.00546. Similarly, the asymptotic t-stats are quite similar: -2.04 versus -2.10. Hence, the inference regarding the discount rate does not appear to be driven by the Fed's 1994 change of practice.

References

Adams, Greg, Grant McQueen, and Robert Wood. 1999. The effects of inflation news on high frequency stock returns. Unpublished paper, Brigham Young University.

Balduzzi, Pierluigi, Edwin J. Elton, and T. Clifton Green. 2001. Economic news and bond prices: Evidence from the U.S. Treasury market. Journal of Financial and Quantitative Analysis 36:523-43.

Boyd, John H., Ravi Jagannathan, and Jian Hu. 2001. The stock market's reaction to unemployment news: Why bad news is usually good for stocks. NBER Working Paper No. 8092.

Chen, Nai-fu, Richard Roll, and Stephan A. Ross. 1986. Economic forces and the stock market. Journal of Business 59:383-403.

Fama, Eugene F., and Kenneth R. French. 1989. Business conditions and expected returns on stocks and bonds. Journal of Financial Economies 25:2349.

Hansen, Bruce E. 1992. Testing for parameter instability in linear models. Journal of Policy Modeling 14:517-33.

Harrison, Paul, and Harold H. Zhang. 1999. An investigation of the risk and return relation at long horizons. The Review of Economics and Statistics 81:399-408.

Hodrick, Robert J., and Edward C. Prescott. 1997. Postwar U.S. business cycles: An empirical investigation. Journal of Money, Credit, and Banking 29:1-16.

Jain, Prem C. 1988. Response of hourly stock prices and trading volume to economic news. Journal of Business 61:219-31.

Jensen, Michael C. 1978. Some anomalous evidence regarding market efficiency. Journal of Financial Economics 6:95-101.

Krueger, Alan B. 1996. Do markets respond more to more reliable labor market data? A test of market rationality. Unpublished paper, Princeton University.

MacKinnon, James G., and Halbert White. 1985. Some heteroskedasticity-consistent covariance matrix estimators with improved finite sample properties. Journal of Econometrics 29:305-25.

McCloskey, Donald N. 1985. The loss function has been mislaid: The rhetoric of significance tests. American Economic Review Papers and Proceedings 75:201-5.

McQueen, Grant, and V. Vance Roley. 1993. Stock prices, news, and business conditions. The Review of Financial Studies 6:683-07.

Mehring, James M. 2001. Stocks may take another hit. Business Week, 12 February, p. 21.

Newey, Whitney K., and Kenneth D. West. 1987. A simple, positive semi-definite, heteroskedasticity and autocorrelation consistent covariance matrix. Econometrica 55:703-8.

Nyblom, Jukka. 1989. Testing for the constancy of parameters over time. Journal of the American Statistical Association 84:223-30.

Pearce, Douglas K., and V. Vance Roley. 1985. Stock prices and economic news. Journal of Business 58:49-67.

Ross, Stephen A. 1987. Finance. In The new palgrave: Finance, edited by John Eatwell, Murray Milgate, and Peter Newman. New York: W. W. Norton and Co., pp. 1-34.

Schwert, G. William. 1981. The adjustment of stock prices to information about inflation. Journal of Finance 36:15-29.

Urich, Thomas J., and Paul Wachtel. 1984. The effects of inflation and money supply announcements on interest rates. Journal of Finance 39:1177-88.

Veronesi, Pietro. 1999. Stock market overreaction to bad news in good times: A rational expectations equilibrium model. The Review of Financial Studies 12:975-1007.

Received July 2002; accepted May 2003.

Marc Poitras, Department of Economics, University of Dayton, Dayton, OH 45469-2251. USA; E-mail [e-mail].