Macroeconomic surprises, market environment and safe-haven currencies

In this paper, we study the reaction of the CHF and JPY to macroeconomic surprises and changes in the broader market environment before and during the crisis using high-frequency data. We show that both currencies are traditionally highly sensitive to macroeconomic surprises. This link, however, was signicantly magnied during the crisis and effects persisted during times when monetary authorities implemented specic measures to limit the appreciation trend. We also nd some evidence that, during the crisis, the CHF and JPY tended to respond more strongly to surprises generating an appreciation than to surprises leading to a depreciation. Both currencies also systematically respond to changes in the general market environment. This result is robust to the use of two measures of the market environment: VIX and on a novel index based on Bloomberg wires. Finally, our results suggest that negative macroeconomic surprises and deteriorations in the market environment are two distinct channels generating appreciation pressure on these two safe-haven currencies. JEL Classication: F31, G12, G14.


Introduction
The Swiss franc and the Japanese yen are prominent safe-haven currencies. Their specic safe-haven status has been documented among others by Auer (2015), De Bock and De Carvalho Filho (2015), Botman et al. (2013) as well as Ranaldo and Söderlind (2010). In times of increased global risk aversion, these are the only currencies that appreciate against all other currencies, including the US dollar (USD).
The troubled times experienced by the world economy since the outburst of the subprime crises in August 2007 have brought about further evidence of the safe-haven status of these two currencies. With the intensication of the international crisis both the Swiss franc (CHF) and the Japanese yen (JPY) experienced waves of massive appreciation. Figure 1 displays the development of the nominal effective exchange rate for the two currencies.
The magnitude of the appreciation and its consequences in terms of de-ationary pressures and output loss in the face of the zero-lower bound constraint forced monetary authorities in the two countries to adopt unconventional policies to limit the appreciation of their currencies in order to achieve their broader policy goals 1 . As depicted in Figure 2, this is reected in the unique increase in the SNB and BoJ balance sheets. In this paper, we will summarize the adoption of these unconventional monetary policies under the term regime change. This is just for convenience. By that, we do not intend to suggest that this has corresponded to any fundamental change in the monetary policy strategy or institutional framework of either the SNB 1 Between March 2009 and May 2010, the SNB intervened several times on the markets to contain the appreciation of the CHF (SNB, 2010 and 2011a). It also repeatedly made use of verbal interventions to manifest its discontent with the appreciation trend of the CHF and to signal that it stood ready to take further steps if needed. In September 2011, confronted with renewed strong appreciation pressures, the SNB introduced a formal exchange-rate oor vis-à-vis the EUR (SNB, 2011b). After lifting the exchange-rate oor in mid-January 2015, the SNB stated that it "will continue to take account of the exchange rate situation in formulating its monetary policy" and that "if necessary, it will therefore remain active in the foreign exchange market" (SNB, 2015). With respect to the Japanese currency, the stark appreciation that followed the Lehman Brothers events prompted the G7 to issue a statement expressing specic concern about the "recent excessive volatility in the exchange rate of the yen" (G7, 2008). The Bank of Japan (BoJ) started to intervene in September 2010 as the JPY hit a fteen-year high against the USD. On March 18, 2011, the G7 countries jointly intervened to dampen the JPY strength. Various waves of BoJ interventions also occurred in the following months. However, as Figure 1 shows, the success of these measures was very limited. PM candidate Abe and the government that was appointed in December 2012 forcefully made the case for a further "unlimited" monetary policy easing to relaunch the economy and ght deation. Indeed, the BoJ's easing stance became more resolute after the appointment of the new government. Markets clearly interpreted these signals as a game changer for the JPY (WSJ, 2012). 2 or the BoJ. The safe-haven status of these currencies implies that the demand for CHF and JPY, and hence their market value, is crucially driven by shocks that impact the macroeconomic outlook and raise the general level of uncertainty on the international nancial markets. Generally speaking, any piece of information that points to stable and balanced international growth will reduce markets' risk aversion and thus the demand for safe-haven assets. Conversely, news reecting unanticipated economic weakness or instabilities of various sorts will spark risk aversion and foster the demand for safe assets. This principle, however, leaves a series of questions open. To which pieces of information do safe-haven currencies really react? How was this link inuenced by the crisis? Does a stronger monetary policy focus on the exchange rate modify or even break the link? This paper investigates precisely these questions. Based on a sample of high-frequency data covering the period between January 2000 and December 2013, we study the effect of macroeconomic surprises, the general market environment and the monetary policy regime on the CHF and JPY exchange rates.
Narrowly dened macroeconomic surprises are a rst likely source of safehaven currency uctuations. Several studies have investigated the connection between macroeconomic surprises -dened as the difference between the value released and the corresponding market expectation for a given macroeconomic indicator -and asset price adjustments. These studies have shown that some macroeconomic indicators, or more precisely their unexpected component, systematically generate signicant and rapid reactions in asset prices.
However, the circle of factors that can potentially inuence markets is obviously much wider and disparate than standard macroeconomic indicators. The general market environment, as we will call it, may be affected by factors like corporate news (such as earning announcements), economic policy news (such as budget numbers or structural reform announcements), political news and more. The trouble with these potential sources of uncertainty and market volatility is that they are less systematic and much harder to measure than standard macroeconomic surprises. That is why research traditionally takes an indirect approach: A popular proxy for general market uncertainty and risk aversion is the CBOE VIX index of implied volatility of the S&P 500 index options. In this paper we rely on the standard VIX measure as well as on a novel index based on the Bloomberg wires to capture market-environment conditions.

3
Our results allow us to make several points. First, we show that, in spite of their safe-haven status, CHF and JPY are not uniquely driven by foreign surprises. Some domestic variables matter in a very systematic and robust way. In fact, their role seems to be more stable across states of the economy than the one of foreign variables. Secondly, full sample regressions provide clear evidence that negative foreign surprises provoke an appreciation of the CHF and the JPY, while positive surprises lead to a depreciation. As such, this result simply conrms the outcome of various other investigations. We do, however, obtain relatively high  2 for this kind of regression, signaling the specic relevance of macroeconomic surprises for safe-haven currencies. In particular, we nd strong evidence that the sensitivity of the two currencies to macroeconomic surprises was magnied by the crisis. This observation is resilient to the changes in monetary policy regimes cited above: we nd no evidence that the effects of macroeconomic surprises on the CHF and JPY were muted in any signicant way. Also, we nd some evidence of asymmetries: both currencies tend to display larger reactions when a surprise provokes an appreciation than when it causes a depreciation. Our results conrm the dominance of US macroeconomic indicators as both currencies are indeed strongly driven by US macroeconomic surprises. In addition, we test the importance of several euro area variables for both the CHF and the JPY. As questions on the fate of the European Monetary Union have taken center stage in the crisis and repeatedly rocked the markets, one might expect euro area macroeconomic surprises to have a strong impact on the CHF and JPY. Contrary to this assumption, while some German variables turn out to have a systematic inuence, French, Italian and Spanish variables do not appear to be relevant, while aggregate euro area variables only have an unsystematic impact. Even when focusing specically on the crisis period, macroeconomic surprises stemming from these economies do not gain in importance. Hence, the truly market-relevant news coming from these economies must have taken another channel to impact safe-haven currencies.
This other channel is possibly reected in our market-environment variables. Indeed, according to our results, swings in the general market environment affect exchange rates on top of standard macroeconomic surprises. Market environment and macroeconomic surprises are thus complementary factors in explaining exchange rate movements.
The Bloomberg-based variable turns out to be strongly correlated with the VIX index. On the one hand, this further legitimizes the use of VIX as a proxy for market-environment conditions. On the other hand, however, our alternative market-environment measure displays various convenient features when compared to the VIX. First, it is not inuenced by the macroeconomic surprise variables that we employ. Secondly, it is not bounded at zero and 4 can thus uctuate in an untruncated manner. In doing so, it captures phases of strongly positive market sentiment better than VIX. Our results do indeed suggest that the Bloomberg-based variable is better positioned than VIX to capture the impact of extreme conditions on safe-haven currencies.
This paper is organized as follows. The next section summarizes the lessons learnt from the existing literature. Section 3 describes our dataset and denes our variables. Section 4 sketches our empirical approach and presents the results while section 5 concludes.

Relevant literature
A large body of literature has used high-frequency data to study the impact of macroeconomic surprises on asset prices: equities, exchange rates, treasury bills, bonds and forward rates. Generally speaking, there is broad-based evidence that macroeconomic surprises can produce signicant changes in asset prices 2 . These changes tend to occur very rapidly after the surprise is observed (Pearce and Solakoglu (2007); Andersen et al. (2003); Dominguez (1999)). Not all assets are impacted equally. Goldberg and Leonard (2003) show that the effect of macroeconomic surprises are systematically stronger on the short end of the yield curve than on the long end. Bartolini, Goldberg and Sacany (2008) and Andersen et al. (2007) nd that macroeconomic surprises have their strongest impact on interest rates while the impact is less pronounced for exchange rates and equity prices. The effects on interest rates and exchange rates appear to be longer-lasting than the effects on equity prices. Swanson and Williams (2013a) nd that interest-rate reactions to macroeconomic surprises are less pronounced when the economy operates at the zero lower bound. Similarly, Goldberg and Grisse (2013) argue that the responsiveness of asset prices to macroeconomic surprises can vary over time with various states of the economy. A prominent feature of different studies is the dominance of US macroeconomic variables. Goldberg and Leonard (2003) show that while US macroeconomic surprises have a distinct impact on European yields, the evidence for a reverse inuence is very limited. German yields appear to be more responsive to US surprises than to German or euro area surprises. Andersen et al. (2003) show that the impact of US surprises on the USD/DM exchange rate is much more substantial than the impact of German macroeconomic surprises.
The safe-haven status of the CHF and JPY has been documented by Ranaldo and Söderlind (2010), based on an analysis of all major currencies' reactions to various nancial market shocks over the period 1993-2008. CHF and JPY appreciate in a systematic way when risk peaks. In the most extreme instances, the appreciation of these two currencies is non-linear in the increase in risk. The measure of risk that appears to matter most is FXmarket volatility, i.e. FX-specic risk rather than general measures of risk such as the VIX. In their effort to identify what characterizes a safe-haven currency, Habib and Stracca (2011) also fail to nd a stable link between the VIX and the currencies of advanced economies. Instead they nd that an increase in the VIX is systematically associated with depreciations of emerging market currencies. De Bock and de Carvalho Filho (2013) use spikes in the VIX to dene risk-off episodes and show that in such periods returns on the CHF and JPY outperform those of all other currencies. Yeşin (2016) nds that the VIX displays stronger co-movements with the CHF real effective exchange rate than do capital ows to and from Switzerland. Thus the evidence on the usefulness of the VIX as a measure of global risk aversion to explain the behavior of safe-haven currencies is mixed.

Exchange rates
We work with a dataset covering the period from January 1, 2000 up to the end of 2013, at a ve-minute frequency. All in all, the full sample includes 1,472,832 observations for each exchange rate cross. Nominal exchange rate data are taken from the Swiss National Bank database and originate from the EBS trading platform. We focus the on following exchange rate crosses: EURCHF, USDCHF, EURJPY, USDJPY. The notation of the currencies is such that a rising (falling) value in the quotes represents a depreciation (appreciation) of the safe-haven currency. For each point in time we dispose of bid and ask quotes as posted on EBS, and compute transaction prices as the average between the two. In the absence of formal quotes, the price at time t is computed on the basis of the latest quotes observed. We remove weekends from our data sample and focus on trades occurring between Sunday 11:00 pm and Friday 9:00 pm British Standard Time (BST). For simplicity's sake, all variables are matched to a continuous time vector (without summer time). Returns are computed as logarithmic differences: where the interval between  and  + 1 in our baseline specication lasts ve minutes.

Macroeconomic surprises
The database on macroeconomic news and market expectations is provided by Haver Analytics, which has taken over surveys that were originally conducted by Money Market Services. The database includes the precise time of the official data release, the number released and the corresponding market expectation. Expectations are dened as the median forecast over all market participants included in the survey. These data sources are the ref- Our initial dataset groups together information for ten economies: Switzerland, US, euro area, Germany, France, Italy, Spain, Belgium, the UK, and Japan.
Following a well-established approach, we compute macroeconomic surprises at time t for any given macroeconomic indicator k as the difference between the released number (R  ) and the value expected by market participants, E  . Also, in order to facilitate a comparison of the relevance of various surprises across macroeconomic indicators, surprise variables are standardized as in Balduzzi et al. (2001) by dividing the surprise by its sample standard deviation.
Hence, our surprise variables are dened as follows: (1) If the data released correspond precisely to the expected value, S  will be equal to zero. Note that, by construction, each time series S  is predominantly composed of missing values, as announcements for any variable k only occur once a month or once a quarter. Regressions below will focus only on exchange rate changes when surprises are available.
S  is a true surprise only under the assumption that E  actually incorporates all information available up the announcement time. In other words, if the amount of time elapsing between the market participants' survey and the actual announcement is large enough, surveys might provide an inappropriate proxy of rational expectations at time t as they do not incorporate information made available in between. This question has been investigated by Andersen et al. (2003) and Swanson and Williams (2013b), for example. Formal tests do not nd any evidence that the reliability of surveys as rational expectation proxies should be questioned.

Market environment
Surprises stemming from standard macroeconomic indicators are not the only factors that move asset prices. Qualitative information often plays a crucial 7 role. Any unexpected piece of news that might impact the (future) performance of the economy can affect nancial markets' mood -risk aversion and market expectations -and thus move asset prices. The spectrum of potentially relevant information is broad. It includes further macroeconomic news (such as growth projections, budget or public debt numbers), macroeconomic policy announcements (such as structural reforms), individual companies' information, and national or international political events. We summarize these elements under the term market environment.
A major problem naturally exists in the concrete measurement of the factors affecting the market environment. In order to incorporate this crucial aspect in our analysis we take two different routes. The rst is an indirect route. We identify the market environment measure with the inverse of the CBOE VIX index of implied volatility of the S&P 500 index options, a proxy for general nancial market uncertainty and risk aversion. This same approach is taken in an abundant number of contributions in the literature. We name this rst market-environment variable, VIX  .
The second and more novel approach is based on Bloomberg wires. A Bloomberg algorithm allows us to count stories as reported in the wires on a daily basis according to desired ltering criteria. Users have various options. They can create their own story counts by using key words or by selecting and combining specic topics. Alternatively, they can rely on pre-dened story counts as provided by Bloomberg itself. The algorithm distinguishes between positive and negative news according to their historical relevance for markets. Our second market-environment variable, BB  , is derived as the simple difference between two very broad-based pre-dened topics, 'POSITIVE' and 'NEGATIVE', as computed by Bloomberg itself. Under the rst topic the algorithm counts all 'potentially positive news'. Under the 'NEGATIVE' topic the algorithm records all stories which appear to be 'potentially negative news'. The denition of 'potentially positive' and 'potentially negative' is entirely empirical and rests with the Bloomberg algorithm. To provide the reader with a sense of which pieces of news are incorporated in this variable, Table A1 in the appendix lists the 38 top-ranked 'NEGATIVE' pieces of news taken into account in the computation of the story count on the day that BB  reached its lowest absolute value in our sample (March 13, 2011). It is important to stress that BB  does not contain macroeconomic data releases and thus does not overlap with our standard macroeconomic surprises.
BB  is so dened that negative values correspond to days when negative news dominates. Figure 3 compares normalized versions of BB  and the inverse of VIX (corresponding to our VIX  ), so that negative values correspond in both cases to a risk-averse environment. The two variables are highly correlated but not identical. At a daily frequency, BB  is much more volatile.

8
In the 2003 to 2007 period -generally speaking a time of expansion and stability of the world economy -BB  still displays repeated negative spikes while VIX  does not. From 2011 onwards, the two variables diverge. While VIX  goes back to the positive territory, in our logic pointing to a reduced demand for safe havens, BB  remains mostly negative. Figure 4 shows the daily changes in the two variables. Clearly, the second part of the sample displays a systemically higher volatility in marketenvironment conditions than the pre-crisis period. In this period it is not uncommon to observe uctuations in BB  in the order of magnitude of four standard deviations from one day to the next.

The role of macroeconomic surprises
A recurrent result in the existing literature is that US macroeconomic surprises often have a stronger impact on any country's asset prices than national surprises. Given their safe-haven function, one can expect the CHF and JPY to respond to a broader range of international macroeconomic surprises than other asset prices. Accordingly, beyond the domestic indicators, we included news originating from several other countries in our universe of potential explanatory variables: the US, the UK, Germany, France, Italy, Spain, Belgium, and the euro area as a whole. There are two reasons for including individual euro-area countries. First, some individual country data are released faster than the aggregated euro area data, and hence provide early signals for the area as a whole. Secondly, we want to be able to isolate the impact of macro-news stemming from specic countries that attracted signicant market attention during the European debt crises.
In a rst step, we estimate the following model: where ∆fx  +1 are the ve-minute returns of the exchange rate cross c.   captures for each macroeconomic variable k the magnitude of the exchange rate reaction to any surprise S that occurs at time t.   coefficients can be interpreted as the percentage point changes in the exchange rate generated by a one-standard-deviation surprise. Coefficients estimates stem from OLS 9 regressions. t-statistics are computed using heteroskedasticity and serial correlation consistent (HAC) standard errors.
The number of explanatory variables initially considered is very large. We operate a rst broad screening by regressing our four exchange rate crosses on the complete list of non-domestic S  variables. We refer to the 10% significance level as the selection threshold. Regression results are not reported in this paper. They reveal a rst clear verdict: macroeconomic news stemming from France, Italy, Spain, Belgium and the UK have no material relevance for the CHF and JPY. We thus remove them all from the set of explanatory variables. We will return later to the meaning of this outcome.
In a second step, EURCHF and USDCHF returns are regressed on all US, German and euro area macroeconomic surprises in addition to Swiss macroeconomic surprises. The same exercise is made with EURJPY and USDJPY, where Japanese surprises are used instead of the Swiss variables. Only a small number of variables appears to be consistently non-signicant across all four regressions. Thus if we took signicance in at least one regression as the selection criterion this would still leave us with a very large number of explanatory variables. Consequently, for each of the two currencies we selected only those variables that appear to be signicant both in the EUR and USD regression. Applying this criterion leads to a choice of 22 variables for the CHF and 25 variables for the JPY. Table A2 in the appendix provides descriptive statistics for all these variables.
Tables 1 (for the CHF) and 2 (for the JPY) report detailed regression results using the explanatory variables we just selected. As the p-values show all coefficients appear to be highly signicant. Generally speaking, as one might have expected, returns in both currencies are strongly affected by US macroeconomic surprises. However, the exposure to US macroeconomic surprises is more pronounced for the JPY than for the CHF: 19 US variables impact the JPY crosses, against 14 variables for the CHF. European news is more relevant for the CHF: four variables enter the CHF regressions, while only two are relevant for the JPY crosses. In the case of the CHF, only German variables have been selected, while in the case of the JPY, one euro-area-wide variable emerges. All in all, there thus seems to be surprisingly little systematic impact of non-German European variables on the two currencies.
A further interesting result concerns domestic indicators. Both currencies are signicantly impacted by surprises in their domestic macroeconomic variables. Hence, movements in the CHF and the JPY are not driven only by international macroeconomic surprises. Unexpected domestic developments do matter.
All in all, the  2 metric tells us that in this baseline specication our 10 regressions explain between 15% and 29% in the variation of 5 minutes returns for the CHF and between 11% and 35% for the JPY returns. As a comparison, the highest  2 obtained by Ranaldo and Söderlind (2010) is 8% (for USDCHF). This suggests that macroeconomic surprises are important independent drivers of safe-haven currencies. Table 1 here Table 2 here Looking at the EURCHF cross, all foreign variables display a positive sign. This corresponds to the expected result. It implies that positive surprises (S   0) favor a depreciation of the CHF vs the EUR 3 . Conversely, negative surprises (S   0) lead to an appreciation of the CHF. Domestic variables bear a negative sign. This is again the result we expected. Positive domestic surprises naturally tend to generate an appreciation pressure as they ceteris paribus imply a risk of a tighter monetary policy than previously projected. As far as the USDCHF regression is concerned, results are similar, with one prominent exception: all German variables are signicant but bear the wrong sign. The explanation is to be found via the EURUSD cross. Positive German surprises cause an appreciation of the EUR against the USD and the CHF. The former being more pronounced than the latter, the Swiss currency appreciates against the USD. Results for the JPY are qualitatively very similar. Most   coefficients are highly signicant and only one bears the wrong sign.

Impact of the crisis
Next, we want to check whether the crisis has impacted the link between macroeconomic surprises and CHF or JPY variations. The hypothesis is that, in periods of heightened uncertainty, safe-haven currencies react more to macroeconomic surprises. As a divide we take the outbreak of the subprime crisis in August 2007. Figure 1 shows that both currencies initiated an appreciation trend precisely around this time. Tables 3 and 4 show regression results. To make the main result easier to spot, coefficients of variables that are more relevant in the second period are in bold print.
In three out of four crosses, the number of coefficients that turns out to be signicant is substantially larger in the second period. The exception is provided by the USDCHF cross. Interestingly, it is only in terms of foreign variables that a pattern emerges between the two subsamples. In other words, there is no obvious evidence that the role of domestic variables varies in the two periods.
It is not just the number of signicant coefficients for foreign variables that changes, but also the size of these coefficients. Typically, these are several orders of magnitude bigger during the crisis than in the pre-crisis period. For instance, whereas the reaction of EURCHF to surprises in the German Real GDP is twice as big and the reaction to surprises in the US Empire State Business Conditions Index or in the Consumer Condence is some four times larger, the reaction to surprises in the US Non-Manufacturing ISM is no less than twenty times larger. In the case of the JPY, the change between the two periods is extremely impressive. Whereas domestic surprises tend to be signicant (and bear the right sign) already in the pre-crisis period, most foreign variables turn signicant only when the crisis occurs. This pattern is clearly visible in both the EURJPY and USDJPY exchange rates. The most extreme individual case is provided by surprises in the Nonfarm Payroll measure of employment: while in the pre-crisis period this variable bears the wrong sign, in the crisis a one standard deviation negative surprise generated a 0.27% JPY appreciation against the EUR and a 0.37% appreciation against the USD.
All in all, we thus nd strong evidence of a much more pronounced impact of foreign macroeconomic surprises on our two safe-haven currencies during the recent crisis. The increased impact of macroeconomic surprises is also reected in higher  2 in the crisis. These results complement the evidence in Goldberg and Grisse (2013) on time-varying responses of asset prices to macroeconomic surprises. Table 3 here  Table 4 here In a separate set of regressions we look at whether macroeconomic surprises also have more persistent effects in the crisis, by using one to several hour windows instead of a ve-minute window to compute the change in the exchange rates. We nd no evidence of a change in patterns in this respect. This result stands in no contradiction to the increased intensity of reaction in ve-minute windows, as of course the crisis is a time of much higher volatility in exchange rates.

Asymmetries
One open question is whether safe-haven currencies respond symmetrically to macroeconomic surprises. Evidence in this respect diverges.  (2007) argue on the contrary that no evidence of asymmetric reactions can be found in the dollar-deutsche mark and the dollar-yen exchange rates. They use data spanning between 1986 and 1996 and surprises for a much smaller number of variables than we do here.
We collected all surprises that have the potential to generate an appreciation in an APP_S  vector and all surprises potentially leading to a depreciation in a DEP_S  vector. Hence, in the case of the EURCHF for instance, APP_S  contains all negative foreign surprises as well as the domestic positive surprises, multiplied by -1. DEP_S  contains all positive foreign surprises and all negative domestic surprises, again with the inverted sign. We then estimate the following regression for the pre-crisis and the crisis samples: We expect both   and   to have a positive sign. Asymmetry would request   6 =   . Table 5 displays the results.
In all but one case, coefficients bear the expected sign. Whereas in the pre-crisis sample no clear pattern emerges, in the crisis period the coefficient of   _S  is larger than the  _S  coefficient in all four regressions. This suggests that macroeconomic surprises have a somewhat larger impact on both currencies when they induce an appreciation than when they have weakening effects. However, t-tests do not allow for a formal rejection of the the   =   hypothesis. We should recall, however, that the crisis sample also contains extended periods in which monetary policy counteracted appreciation tendencies in the two currencies. Still, we interpret these results as a signal that to some degree asymmetric reactions might be at work, adding evidence to the ndings of Fatum et al. (2010).
Note also, that this parsimonious specication conrms the qualitative results of tables 3 and 4: the surprise coefficients tend to be larger and the  2 higher in the crisis period than in the pre-crisis period. The only exception is again the USDCHF cross.

Non-linearity
Ranaldo and Söderlind suggest that safe-haven currencies may behave in a non-linear fashion. In particular, they seem to respond non-linearly to 13 varying levels of their preferred risk measure, observed volatility in the FX markets. In the following step we analyze whether evidence of non-linear responses to macroeconomic surprises can be found. We estimate the following regressions The non-linearity hypothesis is conrmed if    0 and    0. As results in Table 6 show, we nd no evidence at all of a non-linear impact of large surprises.

Impact of the monetary policy regime
One might wonder to what extent the link between macroeconomic surprises and exchange rate reactions is affected by the attitude of monetary authorities. In a further step, we therefore investigate whether the sensitivity of the CHF and the JPY to macroeconomic surprises was modied by the change in the SNB and BoJ monetary policy regimes. The landmarks are provided by the adoption of the minimum-exchange-rate policy by the SNB in September 2011 and by the implementation of a more aggressive quantitative-easing policy by the BoJ as of December 2012.
Because of the reduced number of observations in the sample we refer again to the parsimonious specication (3). Results are displayed in table 7 and shall be compared to those in the lower panel of table 5.
If one is looking for nuances, then one might read estimates in table 7 as suggesting that the size of the   _S  coefficients declined after the new policy regimes were adopted, particularly in the case of EURCHF and EU-RJPY. Globally, however, coefficients remain highly signicant, suggesting an unchanged relevance of macroeconomic surprises. We thus conclude that macroeconomic surprises maintained their impact on the CHF and JPY even under these two specic policy regimes.

Role of the market environment
In a next step, we introduce the variables measuring general market conditions. They will operate as daily dummies in our regressions. We regress 14 our four exchange rate crosses on all variables of Tables 1 and 2 respectively plus our market-environment measures      is alternatively measured by VIX  and BB  (jointly and separately) and a one-day lag (corresponding to 288 ve-minute spells) of the two as well as by (standardized) changes in VIX  and BB  . Table 8 provides summary statistics for these variables.
In order to better gauge the signicance of our market-environment variables, regressions are based on 1.473 million observations, i.e. including the exchange-rate returns when no macroeconomic surprises are available. The VIX index is traded from 9:15 to 16:15 US East coast time. VIX  denotes prior-day closing values. BB  refers to news disseminated between 00:05 and 23:55 on the day preceding ∆fx  +1 . Of course, VIX  and BB  are going to grow older (and possibly less relevant) as we refer to ∆fx  +1 further into the next day. Table 9 displays the results for the various versions of regression (6). For the sake of manageability, only the results for the market-environment variables are reported. Results for the surprise variables remain unchanged. Table 9 here We recall that positive values of VIX  and BB  are associated with a favorable environment, while negative values reect times where negative news dominates, and thus risk is high. VIX  and BB  bear the right sign in all four regressions. An increase in these variables is associated with a depreciation of the CHF or JPY. When introduced separately, VIX  and BB  appear to be signicant in three out of four regressions. The sign and size of the coefficient of VIX −288 and BB −288 clearly suggest that we refer to the changes in these two variables. Results for specications including the changes only are in the fourth and fth panel of table 9.
All in all, we interpret these results as clear evidence that returns in EURCHF, EURJPY and USDJPY are systematically inuenced by changes in VIX  and BB  . Coefficients are small. We should recall, however, that we are regressing ve-minute returns on daily variables. Hence a coefficient of 0.0002 for ∆BB  in the EURCHF regression implies that a change in BB  by one standard deviation moves the EURCHF exchange rate by 0.06% on a daily basis. Similarly, a one standard deviation change in VIX  is associated with a 0.99% variation in the EURCHF. These are economically very substantial numbers.
The previous results suggest that market conditions have on average a signicant impact on the CHF and JPY. Safe-haven currencies, however, are thought to respond in a particularly pronounced manner during times of extreme stress. Ranaldo and Söderlind (2010) nd 'indicative' evidence of this phenomenon by using a dummy variable for 21 extreme events between 1993 and 2007, covering natural disasters, nancial crises and acts of war or terror. We proceed by expanding regression (6) with dummy variables that reect extreme values of our market-environment conditions. Concretely, we create a dummy for extreme negative values of BB  and one for extreme positive values of BB  by setting EN _VIX  is dened in a similar manner. EP_VIX  is actually a vector composed of 0s only, as our standardized measure of VIX  never gets higher than two standard deviations.
The two lower panels of table 9 show the results. Here again we also introduce one-day lagged dummies for extreme conditions. By construction, coefficients for our extreme market conditions variables are expected to be positive. As far as EN _VIX  is concerned, coefficients are either insignicant (CHF crosses) or bear the wrong sign (JPY crosses). On the contrary, extreme market environment conditions as measured by our Bloomberg index provide signicant results. Both the CHF and JPY respond with an appreciation to extreme negative conditions as measured by EN _BB  . The response to extreme positive conditions instead is signicant and makes sense only for the USDCHF exchange rate.

Do macroeconomic surprises and the market environment interact?
We next verify whether macroeconomic surprises and the market environment interact. One could in particular expect a larger impact of macroeconomic surprises on the CHF and JPY when the market environment is particularly positive or particularly negative.
We run regression (7) using both levels and changes in VIX  and BB  as market environment measures. Across the board we nd no convincing evidence that the market environment interacts with macroeconomic surprises. They thus appear to be two separate channels.

Conclusions
We use ve-minute returns for the CHF and JPY over a thirteen-year period encompassing the post-2007 crisis to investigate the link between macroeconomic surprises, general market environment and monetary policy on the behavior of these two safe-haven currencies. We show that macroeconomic surprises and the general market environment are two distinct channels through which international factors can affect safe-haven currencies.
Both currencies are strongly affected by surprises in foreign macroeconomic data. US variables are the main drivers. However, German variables also play an important role, in particular for the CHF. The meaning of macroeconomic surprises for the two exchange rates is highly dependent on the state of the international economy. In particular, the impact of foreign surprises on CHF and JPY returns was strongly magnied by the crisis. Domestic surprises instead, maintain a relatively stable role across states of the economy. We also nd some evidence that the impact of macroeconomic surprises can be asymmetric: surprises tend to produce somewhat larger appreciation movements than depreciation movements during the crisis.
Somewhat surprisingly, we nd no signicant relationship between our safe-haven currencies and macroeconomic surprises stemming from the countries that were at the heart of the European sovereign debt crisis. It must thus be the case that information originating in these countries affected safehaven currencies via other channels. These other channels are most likely captured by our market-environment measures. Indeed, we show that safehaven currencies react in a systematic way to measures of changes in the general market environment. This result is robust to the use of two distinct measures of the market environment: the VIX and a variable based on Bloomberg wires. We also show that the reaction of safe-haven currencies is particularly pronounced when the market environment turns extremely negative. This result, however, can only be identied when our Bloomberg measure of market conditions is used. This suggests that this variable displays a signicant value added, compared to VIX, when it comes to understanding the behavior of safe-haven currencies.