Content uploaded by Jonathan Benchimol
Author content
All content in this area was uploaded by Jonathan Benchimol on Jul 30, 2023
Content may be subject to copyright.
Forecast Performance in Times of Terrorism
Jonathan Benchimol†and Makram El-Shagi‡
September 2020
Abstract
Governments, central banks, and private companies make extensive use of
expert and market-based forecasts in their decision-making processes. These
forecasts can be affected by terrorism, a factor that should be considered by
decision-makers. We focus on terrorism as a mostly endogenously driven
form of political uncertainty and assess the forecasting performance of market-
based and professional inflation and exchange rate forecasts in Israel. We
show that expert forecasts are better than market-based forecasts, particularly
during periods of terrorism. However, the performance of both market-based
and expert forecasts is significantly worse during such periods. Thus, policy-
makers should be particularly attentive to terrorism when considering infla-
tion and exchange rate forecasts.
Keywords: Inflation, Exchange rate, Forecast performance, Terrorism, Market
forecast, Expert forecast.
JEL Codes: C53, E37, F37, F51.
This paper does not necessarily reflect the views of the Bank of Israel. The authors thank
the associate editor, Joakim Westerlund, the anonymous referees, Itamar Caspi, Wagner Piazza
Gaglianone, Dan Galai, Eleonora Granziera, Rudy Malka, Ariel Mansura, Benzion Schreiber, Yoav
Soffer, Michel Strawczynski, Harald Uhlig, Noam Zussman, and participants at the 34th Israel
Economic Association, 49th Money, Macro and Finance Research Group, and 34th CIRET annual
conferences, as well as participants of the Romanian Academy, University of Macau and Bank of
Israel’s research seminars for their useful comments.
†Bank of Israel, Jerusalem, Israel. Email: jonathan.benchimol@boi.org.il
‡Center for Financial Development and Stability, Henan University, Kaifeng, China, and
Halle Institute for Economic Research (IWH). Corresponding author. Email: makram.el-
shagi@cfds.henuecon.education
1
1 Introduction
In recent years, more and more researchers have shown interest in the economic
consequences of terrorism, fuelled by the increase in terrorist attacks in highly
developed countries that have traditionally been considered safe.1
Although this has led to greater understanding of the real economic impact of
terrorism, little attention has been paid to the impact of terrorism on expectations.
Given that terrorism by its very definition aims to "intimidate or create panic",
omission of the more direct psychological impact that can be seen in the expec-
tations seems problematic.2Expectations play a pivotal role in the mechanism of
modern macroeconomic models. Central banks, other policy makers, and other
public and private institutions rely heavily on professional forecasts and market-
implied expectations in their decision-making. Understanding how a climate of
fear generated by terrorism affects those expectations and forecasts is useful for
proper policy making.
This study aims to fill this gap by using data on Israel, the developed country
that has been by far the greatest target of terrorist activity (per capita). There seems
to be consensus in the literature that rare and infrequent terrorist attacks3have a
limited direct and immediate economic impact on developed economies (Abadie
and Gardeazabal, 2003).
However, persistent terrorism, as observed in Israel, might have a different
effect.4Eckstein and Tsiddon (2004) show that in the absence of (regular) terrorist
attacks, such as the Second Intifada, Israel’s per capita GDP would be higher than
its actual GDP. For example, home prices in Israel are lower in regions that are
1See, for example, Blomberg et al. (2004), Crain and Crain (2006), Dorsett (2013), Gerlach and
Yook (2016), and Ruiz Estrada and Koutronas (2016) among others.
2For instance, Wallace and Wild (2010) defines terrorism as “the threat or actual use of vio-
lence in order to intimidate or create panic, especially when utilized as a means of attempting to
influence political conduct.”
3It is important to distinguish different types of terrorism: frequent at low intensity (Israel
during the last decade), frequent at high intensity (Iraq and Syria), rare at low intensity (the United
States during the last decade), and rare at high intensity (the United States on 9/11, the United
Kingdom during the London subway bombings, and France at Paris and Nice).
4Although our paper focuses on Israel, there are some general implications. It has been argued
that the impact of terrorism in the West is different than it is in Israel because of the profound dif-
ference in the nature of terrorism in these countries. As terrorism is more frequent in Israel, it can
be anticipated and modeled, while terrorist attacks remain profoundly unpredictable (black swan)
in other Western countries. However, in recent years this is no longer true, as the Global Terrorism
Index readings of France, the United States, and the United Kingdom (Institute for Economics and
Peace) have become very close to that of Israel. Moreover, fatalities caused by a single terrorist
attack have been higher in European countries than in Israel over the last decade. Therefore, we
explicitly distinguish between the frequency and magnitude of terrorist attacks. Other differences
also, of course, remain. Most importantly, Israel is much smaller than the aforementioned coun-
tries, so per capita terrorism is still unusually high for a developed country. As such, although the
nature of terrorism in the West has changed, application of our results to other countries should be
taken with a grain of salt.
3
more prone to terrorist attacks (Elster et al., 2017). Fielding (2003a) demonstrates
that the First Intifada—which he interprets as a measure of political uncertainty—
contributed substantially to Israel’s low rate of investment, and Fielding (2003b)
shows a decline in the amount of investment.
Other contributions highlight the impact of terrorism on GDP and tourism
(Ruiz Estrada and Koutronas, 2016), inflation (Shahbaz, 2013) and the exchange
rate (Gerlach and Yook, 2016). Local firms’ behavior changes with respect to the
local environment, leading to changes, for instance, in the inflation of nondurable
goods prices in order to sell perishable stocks. Change in inflation is associated
with foreign investment reallocation leading to changes in the exchange rate.
In our paper, we focus on the impact on inflation and exchange rate fore-
casts using both expert and market expectations (implied by the price of inflation-
indexed bonds and the forward exchange rate) in Israel. We restrict ourselves
to these indicators due to a data availability constraint, since these are the only
indicators for which both market-implied and expert forecasts are available at
a monthly frequency over a sufficiently long period to provide meaningful esti-
mates.5
Before conducting a predictive ability analysis, we will lay the foundation by
performing a dynamic analysis of rationality and (relative) forecast performance.
Our assessment relies on the tests for (relative) forecast performance and forecast
rationality proposed by Giacomini and Rossi (2010) and Rossi and Sekhposyan
(2016), respectively, for unstable environments. While neither of these tests allow
for a formal assessment of our key hypothesis that terrorism affects forecast per-
formance, they are extremely helpful because (a) they allow a visual inspection of
forecast performance embedded into a structured framework, and (b) they allow
a broader look at performance that accounts for both relative errors and encom-
passing. This will allow us to show the relation between terrorism and forecast
performance on a more intuitive level.
Based on preliminary evidence gathered by the results of these tests and Gia-
comini and White (2006), we conduct an explicit analysis of the causes of forecast
performance. We control for a range of other aspects of uncertainty and instabil-
ity to ensure our results are not driven by an omitted variable bias. Specifically,
we control for financial instability, commodity prices (particularly oil and gas), ex-
change rate fluctuations, and an econometric forecast of inflation (exchange rate)
uncertainty. Because the conditional relative performance test by Giacomini and
White (2006) does not allow for control variables, we propose a slight modifica-
tion, turning the original correlation-based Wald test into a regression-based Wald
test.
To the best of our knowledge, this study is the first to conduct a broad analysis
5Which is not the case of other economic variables such as consumption or investment.
4
of how terrorism affects forecast performance and, particularly, the first to com-
pare several types of forecasts through different terrorism measures reflecting the
media coverage, nationality and geographic dimensions of the attacks. We find
that terrorism affects market participants much more than professional forecast-
ers. At least in the case of Israel, the low average performance of market partici-
pants seems to be driven mostly by terrorism. In addition, we find that terrorist
attacks affect forecasting performance controlling for the risk premium.
Recent studies about macroeconomic uncertainty, including Farhi and Gabaix
(2016) and Scotti (2016) do not deal with terrorism or short-term warfare. Our
paper is novel in this respect. Although our key research question is in regard
to the impact of terrorism on expectations, we also contribute to the growing lit-
erature comparing market-implied forecasts and professional forecasts in general
(Adeney et al., 2017). Contrary to most of the literature, we do so while fully
accounting for the dynamics of forecast performance, linking us to the literature
on forecast performance in unstable environments. Most notably this literature
includes the papers by Giacomini and White (2006), Giacomini and Rossi (2010),
and Rossi and Sekhposyan (2016), whose methods we borrow. However, the lit-
erature has grown far beyond those original papers, including an abundance of
applications such as Barnett et al. (2014) and El-Shagi et al. (2016), to name just a
few.
The remainder of the paper is organized as follows. Section 2 describes the
stylized facts and related economic forecasts that are analyzed. Section 3 develops
our methodology and the econometric techniques used to quantitatively assess
the impact of terrorist attacks on economic forecasts. The results are presented
in Section 4 and interpreted in Section 5. In Section 6, we outline some policy
implications of our findings. Section 7 concludes.
2 Background
Although there is reason to believe that terrorism might affect agents’ psychology
and expectations, it seems that the literature and institutions making regular use
of forecasts ignore this channel. While many previous studies analyze the impact
of terrorism on current economic activity, few, if any, analyze the impact of ter-
rorism on these essential economic forecasts. Unfortunately, Israel is a relevant
laboratory in which to study this impact. Section 2.1 presents some stylized facts
about terrorism, Section 2.2 describes the market-based and expert forecasts used
in this study, and Section 2.3 details further control variables used for the analysis.
Our analysis spans from 2000 to 2017, using CPI inflation data at a monthly
frequency and terrorism and exchange rate data at a daily frequency. The sources
and detailed transformations are presented below.
5
2.1 Terrorism and uncertainty
Major terrorist attacks (Keefer and Loayza, 2008; Roberts, 2009), or frequent small
and medium-sized terrorist attacks (Sandler and Enders, 2008; Benchimol, 2016),
could affect the economy. The negative impact of terrorism on short-term activity
is a result of the reallocation of internal demand for public consumption (such as
insurance, security forces, and investments) to the detriment of more productive
investments, causing growth to decrease (Blomberg et al., 2004). Our purpose is
to assess how terrorist incidents affect the bias and predictive abilities of experts
as well as market-based forecasts.6
In the aftermath of the September 11 attacks and with the rise of terrorist at-
tacks in the European Union, there has been increasing interest in the economic
consequences of terrorism (Crain and Crain, 2006; Dorsett, 2013; Gerlach and Yook,
2016; Ruiz Estrada and Koutronas, 2016) and the reasons behind it (Dreher and
Gassebner, 2008; Dreher and Fischer, 2010). In addition, linkages between ter-
rorism and economic policy have been extensively analyzed (Dreher et al., 2010;
Dreher and Fuchs, 2011).
In September 2001, when terrorists attacked the United States, the US econ-
omy was already in recession, but it reached positive growth only two months
later. This led to the conclusion that even a major terrorist attack, such as the
destruction of the World Trade Center, would have fairly limited economic conse-
quences. Similarly, after the terrorist attacks in Madrid (2004) and London (2005),
GDP growth trends in Spain and the United Kingdom were not affected. Even
the attacks in Paris (2015) did not show a measurable impact on French consump-
tion. However, in almost all such cases, although the country is geographically or
demographically large in regard to the consequences of terrorism, expectations—
including forecasts—were strongly affected.
The emergence of a small, developed market economy in parallel with a wave
of terrorist attacks provides an interesting economic example (Eldor and Melnick,
2004; Caruso and Klor, 2012). Israel is a perfect case study of a (small) developed
country facing terrorism and war at different levels and frequencies (Eckstein and
Tsiddon, 2004; Larocque et al., 2010).
In the past two decades, there have been five episodes of intense violence in-
volving Israel: the Second Intifada (September 2000 to February 2005), the Second
Lebanon War (July–August 2006), Operation Cast Lead (December 2008 to Janu-
ary 2009), Operation Pillar of Defense (November 2012), and Operation Protective
Edge (July–August 2014).
Unlike most terrorist attacks occurring in Europe or the United States, terrorist
6We conducted different event studies without robust results. Daily variance in market-based
and expert forecasts cannot be explained using terrorism. The effect of terrorism or financial un-
certainty on expectations does not seem to be immediate but takes time to come about.
6
attacks in Israel (Fig. 1) have not involved substantial destruction of property or
infrastructure, except during the First and Second Intifada, but they have some-
times led to substantial casualties (Fig. 2).
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
0
20
40
60
80
100
120
140
160
Quantity
Figure 1: Number of terrorist attacks in Israel between 2000 and 2016. Source: Na-
tional Consortium for the Study of Terrorism and Responses to Terrorism (START),
Global Terrorism Database.
Nevertheless, terrorist attacks affect consumer and investor behavior and, in
turn, stock market prices (Shoham et al., 2011; Kollias et al., 2011). When terror-
ists strike at regular intervals and fear and insecurity win minds and begin to
change agents’ economic behavior, the quality of economic (expert and market-
based) forecasts would be affected by these transitory events.7The psychological
effects on expectations and feelings of uncertainty might be considerable, giving
our study the unique ability to assess the impact of uncertainty, rather than the
unforeseen effects of negative shocks (Romanov et al., 2012).
In this study, we use three sources of statistics on terrorist attacks to measure
terrorism in Israel, each including four different indicators: number of people
killed during terrorist attacks, number of people wounded during terrorist attacks,
total number of casualties (killed and wounded) during terrorist attacks, and total
number of terrorist attacks. We use one academic source (Global Terrorism Data-
7Indeed, forecasters expect the cost of security policies to increase, thus increasing the expected
cost of economic activities and transactions, while large companies are expected to cancel or delay
investments.
7
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
0
100
200
300
400
500
600
Quantity
Figure 2: Number of killed and wounded during terrorist attacks in Israel between
2000 and 2016. Source: National Consortium for the Study of Terrorism and Re-
sponses to Terrorism (START), Global Terrorism Database.
base,8hereinafter GTD) and two government sources (Ministry of Foreign Affairs,
MFA; and the National Insurance Institute,9NII). The number of terrorist attacks
is not available in the MFA data.
These sources have their own methodology to account for terrorist attacks and
casualties. For robustness purposes, and to capture the different dimensions and
psychological components underlying each terrorism measurement methodology,
fear instilled by the media (GTD), fear affecting a specific nationality (NII), or ge-
ography (MFA), we use these different data sources.10
GTD data are generated by isolating an initial pool of potentially relevant arti-
cles and then using sophisticated natural language processing (NLP) and machine
learning techniques to further refine the results, remove duplicate articles, and
identify possibly relevant articles. The GTD team manually reviews this second
subset of articles to identify unique events that satisfy the GTD inclusion criteria,
and then studies and codes them according to GTD specifications. GTD data are
8Database supported by the University of Maryland and maintained by the National Consor-
tium for the Study of Terrorism and Responses to Terrorism (START).
9Which provides social security services in Israel, among other things.
10Terrorism data are also collected by the Israel Defense Forces and B’tselem. However, those
data are not sufficiently recognized in the academic literature, and are potentially less objective
than the three databases we consider.
8
reported separately for pre-1967 Israel, and Gaza and West Bank. To achieve bet-
ter comparability, in particular with the MFA data, and to account for the fact that
most attacks in Gaza and West Bank target Israelis, we aggregate the two datasets.
However, tests conducted with only terrorism data on pre-1967 Israel lead to sim-
ilar results.
MFA data are from the chronology of terrorist attacks in Israel published by
the Israeli Ministry of Foreign Affairs and collected by Johnston (2016). MFA data
include West Bank and Gaza in their definition of Israel.
NII data are from the National Insurance Institute (the State’s social security
agency). These data include terrorist attacks involving Israelis all over the world,
without geographical distinction. That is, in contrast with our other sources (GTD
and MFA), this database contains terrorist attacks outside Israel as well.
The most precise database is NII, as the records are double checked–government
and social security services–and not based on text mining (GTD) or geography
only (MFA). However, we believe GTD has the crucial advantage, due to its re-
liance on media coverage, of being related more closely to terrorism "perceptions",
which is essential for the psychological transmission channel to work.
Although during periods of frequent terrorist attacks, there were months with
very few or no terrorist casualties or attacks (e.g., there were roughly 10 people
injured in February 2003 and in July 2003, at a time when figures in most months
were in the hundreds), we use a backward-looking 12-month moving average for
all terrorism indicators because financial data and turmoil are inherently persis-
tent while the effects of terrorism could last a long time (Marsden, 2012; Bandy-
opadhyay et al., 2014; Karl et al., 2017). The moving average accounts for high
volatility by correctly identifying the respective months as part of periods with
high instability. Regressions with 3- and 6-month moving averages lead to weaker
but qualitatively similar results.
2.2 Market and expert forecasts
Israel is one of the rare developed countries with sound economic institutions
and developed financial markets (and thus access to very detailed economic in-
formation) to have experienced a long history of frequent terrorism. We are par-
ticularly interested in market expectations and professional forecasts. Since Israel
issues both inflation-indexed and unindexed bonds, it is straightforward to com-
pute market expectations of inflation (breakeven inflation rates).
As a measure of professional forecasts, we use the combined professional fore-
cast assembled by the Bank of Israel from different professional sources.11
11The Bank of Israel publishes an average of forecasts provided by several financial institutions.
There are roughly 11 providers of inflation forecasts and 6 providers of exchange rate forecasts (on
average, mainly commercial banks) over our sample.
9
Expert as well as market-based inflation and exchange rate forecasts are useful
when formulating the inflation-targeting monetary policy of a small open econ-
omy. Thus, the Bank of Israel collects a set of forecasts that are updated regularly.12
As per the Bank of Israel’s practice, we use the forecasts as given without fur-
ther risk adjustment. Although this implies that our measures are not perfect mea-
sures of expectations, it guarantees that they are perfect measures of policymakers’
perception of expectations, which is more important. However, we do control for
inflation risk and exchange rate risk to ensure that our results are not driven by
lack of risk adjustment.13
In this study, we use two types of market-based 1-year (1Y) Consumer Price
Index (CPI) inflation forecasts used by the Bank of Israel’s Monetary Policy Com-
mittee (MPC): the 1Y forward (contract) implied inflation forecast and the 1Y
breakeven (zero-coupon bond implied) inflation forecast. The first is the instanta-
neous 1Y forward inflation rate, and the second is reported by the Bank of Israel
as the official 1Y market-based inflation forecast.14 These time series are not trans-
formed and are used as is by the MPC.
In addition, the Bank of Israel collects a series of forecasts provided by profes-
sional forecasters, giving an overview of the professional inflation expectations.
Contrary to many other surveys, individual forecasters are not asked for their
opinions at a given point in time, but they are able to update their forecasts at will,
thereby giving us daily data on professional forecasts.15 The expert forecasts used
in our study are computed as the simple arithmetic mean of the inflation forecasts
of commercial banks and economic consulting firms. This measure together with
the 1Y breakeven inflation forecast are reported in official publications of the Bank
of Israel.16
The situation is equally good for exchange rates. There is an active ILS/USD
future market that allows us to infer market expectations for the exchange rate.
12See, for example, publicly available minutes related to interest rate policy decisions. The first
section, related to inflation, as well as almost all staff forecasts, mention expert and market-based
forecasts.
13See below and Section 4.1 for more details about the inflation risk premium.
14This measure is assumed to deal with several inherent breakeven inflation problems. It consid-
ers the small number of real bond series, bias derived from the indexation mechanism (indexation
lags and other mechanisms impacting the calculation of the yield to maturity of the CPI-indexed
bonds), and CPI seasonality affecting the pricing of CPI-indexed bonds. However, inflation risk
premiums as well as bias derived from differences in taxation and liquidity between different bond
types are not considered.
15Strictly speaking, expert forecasts are not at a daily frequency. The Bank of Israel collects its
own expert forecasts based on a system allowing non-costly and private updating ability of their
forecasts in an unlimited way. If the forecast is not updated, the previous value is considered.
However, because we consider an average of these expert forecasts, and almost all experts update
their forecast every week on average over the reviewed period, this time series frequently changes
during the month.
16Every month, the Bank of Israel publishes a press release. Its section on monetary policy and
inflation (data and reports) details the expected rate of inflation derived from various sources.
10
Given their importance, exchange rates are covered by the Bank of Israel’s in-
house survey of expert (banks) forecasters.17
In this study, we use the 1Y forward (contract) implied USD/ILS exchange rate
forecast as our market-based exchange rate forecast. As practiced in the Bank of
Israel the implied exchange rate is not transformed or adjusted in any further way.
The forecasts are obtained for the last day of the month. Variables used to
explain forecast performance are from the month when the forecast is made. Thus,
they can affect the forecast, and do not just appear as a forecast error by occurring
after the forecast is made. One year is the only forecast horizon where market and
expert forecasts match in Israel.
Detailed descriptions of the performance of the inflation and exchange rate
forecasts are provided in Sections 4.1 and 4.2, respectively.
2.3 Further control variables
Since we are particularly interested in conditional forecast performance, it is im-
portant for us to guarantee that a significant conditionality of forecast performance
on terrorism is not due to omitted variable bias. There are other variables that are
strongly linked to uncertainty that might be correlated with terrorism. Therefore,
our analysis considers a range of control variables explained in the following para-
graphs.
Most importantly, market expectations and professional forecasters respond
to financial uncertainty.18 It is well established that asset prices can predict in-
flation (Stock and Watson, 2003) and foreshadow tail risks in inflation (de Haan
and van den End, 2018), and stock, bond and foreign exchange market commove
(Pavlova and Rigobon, 2007). Correspondingly, financial market uncertainty can
drive inflation uncertainty.
Additionally, our financial control variables allow us, to some extent, to ac-
count for the possibility that the impact of terrorism on forecasts is indeed driven
by its impact on financial markets. For example, terrorism can affect market liq-
uidity depending on the size of the incident (Chen and Siems, 2004). However,
rare large-scale terrorist attacks, combined with improvements in market resilience
17The final variable for which implicit market forecasts exist is interest rates, whose expectations
can be computed from the yield curve. However, the expert forecasters cover the policy rate by the
Bank of Israel. Nonetheless, the expectation on interest rates from Israeli treasury bills, implied
by the term structure, constitutes an implicit forecast for the treasury bill rate. While close to
each other, the two interest rates (policy rate and treasury bill rate) are not precisely the same. In
addition, the nominal interest rate did not change significantly since 2014, while other indicators
(terrorism and control variables detailed in the next section) changed drastically. Therefore, we
exclude the nominal interest rate from our study.
18The global financial crisis and subsequent recovery period provided new insights about fore-
cast evaluation during the period of data instability in both the euro area (Benchimol and Fourçans,
2017) and the United States (Caraiani, 2016), as well as in Israel (Benchimol, 2016).
11
as well as financial stability during the last decade, make liquidity issues less likely
to affect our analysis (Peleg et al., 2011).
Financial variables have the additional advantage of usually exhibiting con-
siderable comovement with the business cycle. This is crucial since we cannot
explicitly control for business cycle indicators – which are typically measured at
quarterly intervals – due to our monthly frequency.19
Then, we employ three different measures of volatility to serve as control vari-
ables. First, we use the monthly standard deviation of daily returns (approximated
as log differences) of the relevant stock market index, that is, the TA-100 index,
which is the broadest leading index in the Tel Aviv Stock Exchange.20 Second, we
include the spread between the highest and lowest levels of this index within the
month of the forecast. Third, we consider the monthly average of the correspond-
ing daily spreads. While the monthly spread reacts more strongly to major move-
ments within a month, the average daily spread implicitly gives higher weight to
intraday fluctuations.
The financial uncertainty measured as the TA-100’s one-month rolling win-
dow volatility (standard deviation) and daily spread (high-low spread) between
2000 and 2016 displays a Pearson correlation of 0.49. The highest level of finan-
cial uncertainty was reached during and around the collapse of Lehman Brothers
(2008Q3–Q4) and the European and Greek debt crises (2010Q1, 2011Q2–Q3, and
2015Q3). Volatility was also high during most of the Second Intifada period and
the (unanticipated) elections for the 18th Knesset21 held in 2009Q1.
In addition, since terrorism in Israel might be related to unrest in the Middle
East as a whole, which has serious repercussions for the price of oil, a major factor
in global economic development, we control for commodity price volatility. More
precisely, in our empirical analysis, we control for the monthly volatility (standard
19Even unemployment, which is traditionally a monthly business cycle indicator, cannot be con-
sidered for Israel. Unemployment figures have only been produced at a monthly frequency since
2012, having been produced on a quarterly basis before that.
20The TA-100 index consists of 100 shares with the highest market capitalization, and includes
the TA-25 and TA-75 indices.
21The Knesset is the unicameral national legislature of Israel.
12
deviation of daily log differences) of crude oil,22 natural gas,23 and CRB commod-
ity price index24 expressed in ILS and USD.
Finally, we control for the volatility of the USD/ILS exchange rate, again com-
puted as the standard deviation of daily log differences over one month.
3 Methodology
While early literature on forecast evaluation usually evaluated forecast perfor-
mance for the entire sample, the past decade has seen the emergence of literature
on forecast evaluation in unstable environments, which accounts for situations
such as those described in Section 2. These new tests allow us to assess time vari-
ation in both the performance of individual forecasts and relative performance of
forecasts, as well as to account for the fact that some models and/or forecasters
might do well in some situations, but not in others.
Some tests involved are essentially supremum versions of established tests
over a rolling window of forecasts. This introduces a multiple testing problem that
causes the critical values of those tests to be much higher than those of the under-
lying individual tests. If there is no fluctuation, this causes an unnecessary loss
in power. Thus, the tests are often accompanied by full sample versions. How-
ever, since most of our results indicate strong rejection, we omit reporting these
full sample tests in this paper.
3.1 Rationality in unstable environments
We start the analysis with the most fundamental question, that is, whether the
forecasts we consider are rational. Rossi and Sekhposyan (2016) suggested using
the maximum of rolling-window Wald-type rationality tests as the test statistic.
Thus, the null hypothesis is that the forecast under consideration is rational at
22West Texas Intermediate (WTI) crude oil spot price, US dollars per barrel, not seasonally ad-
justed. Because of its sulfur concentration, WTI is a better grade of crude oil for the production
of gasoline, while Brent oil is more favorable to the production of diesel fuels. Brent’s higher
sulfur content affected its price as it was more expensive to refine into gasoline. The series of anti-
government protests, uprisings, and armed rebellions known as the Arab Spring began spreading
across the Middle East in December 2010, and the civil war in Libya escalated in February 2011.
These events created concern and worries about the instability in the region, causing volatility in
the oil market. The price of the Brent crude rallied relative to WTI crude, the reason being concerns
over the availability of supply and logistical considerations, e.g., possible violence on seaway pas-
sages. Other world events affected Brent pricing, such as the agreement that Iran would increase
the daily amount of Iranian crude flowing into the market. Since Brent is the pricing benchmark
for Iranian crude, this pushed the price of Brent down relative to WTI. Last but not least, recent
literature uses WTI instead of Brent for Israel (Choi et al., 2018).
23Henry Hub natural gas spot price, US dollars per million British Thermal Units (BTU), not
seasonally adjusted.
24Thomson Reuters/CoreCommodity Commodity Research Bureau (CRB) index.
13
every point in time during the sample. Rejection does not imply that a forecast is
permanently irrational, but that it was irrational at least once during the sample
period. Following the original research, we report the entire time series of under-
lying individual test statistics to obtain a visual representation of which periods
cause the rejection. We apply the same strategy for other related tests.
The underlying test statistic is based on a standard Mincer and Zarnowitz
(1969) regression:
yt+h=α+βˆ
yt+h,t+ηt,h, (1)
where yt+his the variable of interest at t+h,ˆ
yt+h,tis the corresponding fore-
cast made at time t, and ηt,his the error term of the test regression. The tradi-
tional rationality test examines the joint hypothesis that α=0 and β=1. It is
straightforward to observe (as pointed out, for example, by West and McCracken
(1998)) that this can be rearranged so that the h-step-ahead forecast error at time t,
ˆ
υt,h=yt+hˆ
yt+h,t, becomes the left-side variable:
ˆ
υt,h=θ0+θ1ˆ
yt+h,t+ηt,h, (2)
where θ0and θ1are the regression coefficients of the adjusted test equation.25
This gives us the more approachable null hypothesis θ="θ0
θ1#=0, which
can easily be assessed with a standard Wald test using the test statistic:
W=ˆ
θˆ
Ω1ˆ
θ0, (3)
where ˆ
θis the estimator of θand ˆ
Ωis the corresponding heteroscedasticity and
autocorrelation consistent (HAC) robust estimator of the covariance matrix of ˆ
θ.26
With a sample of Pforecasts and using window length m, the proposed test
statistic takes the following form:27
max
j2fm,...,PgWj,m, (4)
where Wj,mis a Wald statistic (as defined in Eq. 3) computed on a subsample
using observations jm+1 to j.
25That is, θ0=αand θ1=β1.
26Since it is well established that ηt,hfollows an MA(h)process even for perfectly unbiased and
efficient forecasts due to overlapping unforeseeable shocks over the hperiods of the forecast, an
HAC correction using a sufficiently high lag order is of utmost importance when not explicitly
modeling this moving average behavior. We follow the suggestion of Rossi and Sekhposyan (2016)
and use a standard Newey and West (1987) estimator for the covariance matrix.
27In our study, we use a sample of forecasts of P=177 and 25-month windows (m=25) in the
kernel for our HAC variance estimators (12 months on both sides), which is above what rules-of-
thumb usually suggest. This should suffice to correct the degree of MA introduced by overlapping
forecasts, taking into account the serial correlation.
14
The distribution of this test statistic depends on whether the uncertainty in the
parameter estimates in the forecasting model itself should be accounted for (i.e.,
not the uncertainty concerning θbut the degree of uncertainty in the parameters
used to generate ˆ
y). In the case of expert or market-based forecasts, the so-called
model-free forecasts, the distribution under the null collapses to its most simple
form, depending asymptotically only on m/Pand not the sample size used to
produce forecasts.
However, using the asymptotic critical values can create fairly sizable distor-
tions in finite samples. For typically available sample sizes (i.e., Pbetween 100
and 200) the number of observations for an individual window quickly becomes
extremely small. We therefore use the finite sample adjusted critical values pro-
vided by El-Shagi (2019).
3.2 Stability of relative forecast performance
3.2.1 Fluctuation test
Much like the rationality test in unstable environments is a maximum of individ-
ual rationality test statistics over a rolling window, the test for relative forecast
performance in unstable environments proposed by Giacomini and Rossi (2010)
is the maximum of traditional relative forecast performance tests over a rolling
window.
Similar to the previous test, the null hypothesis is that the forecasts under con-
sideration perform equally well at every point in time. Exceeding the critical value
does not imply that one model constantly outperforms the other, but merely that
there is a meaningful difference in predictive ability for a subsample.
More precisely, the test statistic is the maximum of local Diebold and Mariano
(1995) test statistics, in which the variance estimator is based on the full sample
of forecasts, rather than the individual window for which the mean difference
in predictive ability is computed. Since the full sample estimator is used in this
approach, mcan be much smaller than in the previously outlined test. This is
because the rationality test needs mto be sufficiently large to allow meaningful
estimation of Ω, which corresponds to ˆ
σin this test, within the rolling window,
which is not relevant here. Denoting the loss function for the two forecasts under
consideration at time tby L1,t,hand L2,t,hand the corresponding loss difference by
∆Lt,h=L1,t,hL2,t,h, we can write the test statistic as
max
j2fm,...,Pgˆ
σ1m1/2
j
∑
t=jm+1
∆Lt,h, (5)
where ˆ
σis the HAC robust estimator of the standard error of the mean of ∆Lt,h.
15
Since the finite sample bias in the rationality test under instability is mostly in-
troduced by the uncertainty in the estimation of Ωover mobservations, the finite
sample problems are far less pronounced in this test, and we use the asymptotic
critical values provided by Giacomini and Rossi (2010).
The test we use is two sided, because there is no valid prior assumption of the
superiority of one forecast over another. For the visual interpretation, we report
that ˆ
σ1m1/2 ∑j
t=jm+1∆Lt,h, rather than the corresponding absolute, which is
part of the test-statistic, to observe whether the rejection is driven by forecast 1 or
forecast 2 to be superior during a subsample.
Tests reported use squared forecast errors as a loss function. Performance dif-
ferences are defined as market-based loss less expert-based (survey of bank’s fore-
casts) loss. Thus, high values of the test statistic indicate worse performance of
market-based forecasts, and corresponding superiority of the expert forecasts.
3.2.2 One-time reversals in forecast performance
Often, a potential change in forecast performance is due to a single structural break
(e.g., introduction of a new forecasting model or policy that is not well understood
by one forecasting agent), rather than fluctuations over time. In this case, the
very flexible framework outlined above still creates an unnecessary loss in power,
compared with a test that explicitly models a single structural break.
Thus, Giacomini and Rossi (2010) proposed a so called one-time reversal test,
which follows the spirit of the supremum structural break tests introduced by
Hawkins (1987).
Technically, the test includes a testing procedure composed of three separate
tests.
The first test statistic is a straightforward full sample test:
LM1=ˆ
σ2P1"P
∑
t=1
∆Lt,h#2
. (6)
The second is the actual structural break statistic based on the loss differences
in various subsamples:
LM2=max
j2f0.15P,...,0.85PgLM2(j), (7)
where
LM2(j)=ˆ
σ2P1(j/P)1(1j/P)1"j
∑
t=1
∆Lt,h(j/P)
P
∑
t=1
∆Lt,h#2
. (8)
16
Finally, the joint test-statistic with the null hypothesis of equal performance at
any point in time is as follows:
φ=LM1+LM2. (9)
Correspondingly, if the third test statistic is rejected, we can reject equal perfor-
mance at every point in time. Only then do we assess the individual underlying
statistics LM1and LM2. If only LM1is rejected, this indicates the permanent su-
periority of one model. If only LM2is rejected, this indicates the reverse, in which
one model is superior only for a certain subsample. If both tests are rejected,
then the interpretation is not as clear cut, but it generally implies some change
in relative performance that is not strong enough to affect the relative order of
forecasts. If there is evidence of a structural break, the most likely breakpoint is
j=argmax
j2f0.15P,...,0.85Pg
LM2(j).
3.3 Encompassing in unstable environments
Even if one forecast is permanently or temporarily better, this does not necessar-
ily imply that the superior forecast fully exploits the available information at all
times. Thus, we move to the next step and test forecast encompassing in the same
framework, proposed by Rossi and Sekhposyan (2016), which we used to assess
rationality.
The key difference is that the equation underlying the Wald test changes to a
standard encompassing equation given by
ˆ
υ1,t,h=θ0+θ1(ˆ
υ1,t,hˆ
υ2,t,h)+ηt,h, (10)
where ˆ
υ1,t,hand ˆ
υ2,t,hare the forecast errors of models 1 and 2, respectively.
Contrary to the rationality test, we are merely interested in θ1. Thus, the indi-
vidual Wald statistics collapse to
W=ˆ
θ1ˆ
ω1ˆ
θ1, (11)
where ˆ
ωis the lower right element of ˆ
Ωand ˆ
θ1is the estimator of θ1.
3.4 Conditional relative performance
After exploring time variation in forecast performance and, more importantly, rel-
ative performance, we now assess the reasons for the variation. To that end, we
employ the test for conditional forecast performance proposed by Giacomini and
White (2006).
17
Denoting the set of conditions that potentially explain the difference in perfor-
mance at time tby row vector ht, the test statistic is given by
T=P P1P
∑
t=1
ht∆Lt,h!ˆ
Ω1 P1P
∑
t=1
ht∆Lt,h!0
, (12)
where Tis a standard Wald statistic based on pairwise correlations between the
elements of htand ∆Lt,h.ˆ
Ωis the corresponding HAC robust estimator of the
covariance matrix. The test statistic follows a simple χ2
qdistribution, in which qis
the number of elements in ht.
The null hypothesis is that forecast performance is not related to any indicator
collected in ht. The explanatory variables are usually, and in our example, mea-
sured at time trather than at t+h; that is, we do not assess what kind of shock
at t+his unforeseeable for certain forecasters, but we assess conditions at twhen
the forecast is made. To a certain degree, this allows us to choose the preferred
forecast, ex ante, that is, when the forecast is made, rather than later when the
realization is known.
It must be noted that including several indicators in htdoes not “control” for
indicators in the sense of regression analysis, since the elements of htare simple
pairwise correlations rather than regression coefficients. However, we also want to
assess whether terrorism truly has an impact that is not related to financial market
uncertainty. Nonetheless, due to the aforementioned construction, performing a
Wald test on only one coefficient estimated jointly with others merely yields a
result that is obtained when testing this individual explanatory variable without
controlling for further indicators.
Therefore, we also run an ad hoc variation of this test, in which we use re-
gression coefficients rather than correlation coefficients and the corresponding co-
variance matrix, and we run a Wald test on the coefficient(s) of interest only. The
underlying regression is
∆Lt,h=φ0+φ1terrort+φ2controlt+ηt,h, (13)
where terrortcan be any of our terrorism indicators discussed in Subsection 2.1
and controltany of the covariates introduced in 2.3.
4 Results
This section presents the results for tests related to the CPI inflation and the USD/ILS
exchange rate, as well as expert and market-based forecasts.
Our regressions are monthly mainly because we chose the forecasts that best fit
reported inflation in terms of timing, and to avoid useless daily noise from finan-
18
cial variables. This also makes sure that the forecasts can correctly be interpreted
and compared as one-year-ahead forecasts.
4.1 Inflation
Rationality Our rationality test in unstable environments strongly rejects for
both expert forecasts and market-implied inflation expectations rationality (Fig.
3).
Since the null hypothesis of the underlying test is that the forecast is always
rational, this does not imply irrationality on average or even for a majority of the
periods. A visual inspection of the time series of the individual Wald statistics that
underlie the (supremum) test statistic used indicates that, in both cases, the rejec-
tion is primarily driven by a strong bias following 2012Q3, which is mainly related
to warfare instability (Operation Pillar of Defense, 2012Q4, and Operation Protec-
tive Edge, 2014Q3). Furthermore, the underlying Wald statistics pick up some
movement during the highest violence level of the Second Intifada, Operation
Cast Lead (2008Q4–2009Q1), unanticipated Israeli legislative elections (2009Q1)
and financial uncertainty until 2010Q3.
(a) Expert forecast (b) Market-based forecast
Figure 3: Rationality test statistic (solid) and the critical value (dashed) for CPI
inflation. Note: Significance level: 5%. Short warfare (dark gray): Second Intifada (A), Second
Lebanon war (B), Operation Cast Lead (C), Operation Pillar of Defense (D), Operation Protective
Edge (E). Relatively long terrorism period (light gray): rockets from Gaza and Lebanon (T1), mor-
tars and rockets from Gaza, Flotilla episode and terrorist attacks (T2), terrorist attacks (including
abroad), mortar and rockets from Gaza (T3), and Stabbing Intifada mixed with periods of rockets
and mortars from Gaza (T4).
Even this preliminary evidence, based on rationality tests, lends quite strong
support to our main hypothesis; that is, uncertainty and instability of any form,
19
whether financial or—as in the unfortunate case of Israel—caused by terrorism,
strongly affects expectations and thereby forecasts.
Relative forecast performance and encompassing Both expert and market-based
inflation forecasts are strongly affected by uncertainty, and the order of magnitude
seems similar. Although a fluctuation test rejects the null hypothesis of equal fore-
cast performance at each point in time, this rejection is exclusively driven by the
early period of the sample, when experts hugely outperformed market expecta-
tions. Fig. 4 shows development of the underlying individual test statistics over
time, indicating another extended period of superior expert forecasts in the sec-
ond half of the 2000s. However, the test statistics are far below the critical value.
At other times, the performance of both forecasts is almost identical.
(a) Fluctuation test (b) One-time reversal test
Figure 4: Fluctuation and one-time reversal test statistics (solid), and the critical
value and average loss differences (dashed) in (a) and (b), respectively, for CPI
inflation. Note: Significance level: 5%. See Fig. 3 for an explanation of the labels.
These prolonged but insignificant fluctuations in relative performance corre-
spond to periods of fear (Second Lebanon War and victory of Hamas in the Pales-
tinian legislative elections of 2006, as well as Palestinian Gaza-based mortar at-
tacks and Israeli missile strikes during 2011–2013). The magnitude of the fluctua-
tions is small enough that we cannot rule out a single structural break in relative
forecast performance around 2004, as indicated by the one-time reversal test.
The encompassing test (Fig. 5) paints a clearer picture.
Again, we clearly reject that the typically superior expert forecasts encompass
market forecasts. These rejections are mainly driven by the very same periods
that drive the results of our rationality test. That is, while both experts and market
20
Figure 5: Encompassing test statistic (solid) and the critical value (dashed) for CPI
inflation. Note: Significance level: 5%. See Fig. 3 for an explanation of the labels.
participants are strongly affected by uncertainty, they are affected in very differ-
ent ways, which leads to the natural follow-up question: Can uncertainty indeed
explain the differences in forecast performance?
Conditional relative forecast performance Our core question is whether terror-
ist attacks affect the (relative) performance of expert and market-based forecasts.
All indicators of terrorism we use turn out to be significant at the 5% level.28
To assess the robustness of these results, beyond the inclusion of a range of
different indicators, we test specifications controlling for the alternative reasons
for time-varying forecast performance that might be correlated with terrorism.29
First, we include several indicators of financial uncertainty (Section 2.3). Terror-
ism might cause market turbulence that in turn affects economic forecasts, just
as financial turmoil might do when caused by any other source. Second, we in-
clude commodity price volatility as well as USD/ILS volatility. Since terrorism
in Israel might be related to instability in the Middle East, where several main
28In addition to the standard model, we run an extended model in which we account for po-
tential nonlinearity by including the square of the terrorism term. In cases in which we find a
significant square term, it is combined with a highly significant coefficient on the level. That is,
there is some evidence that the impact of terrorism on forecast performance declines quickly. In
other words, it is the existence of terrorism, rather than its degree, that affects forecast performance.
However, these results are far from robust and strongly depend on the selection of the terrorism
indicator. Therefore, we opt to exclude said nonlinearity from our baseline specification. More
detailed results are available upon request.
29As a robustness check, we also conducted this exercise with terrorism data excluding terrorism
in Gaza and West Bank, thus using only GTD measures (data on Israel without Gaza and West
Bank). Our results lead to similar conclusions as presented below. These results are available upon
request.
21
oil producers are located, there might be a relationship between inflation forecasts
and commodity price volatility. Commodity prices are known to have a significant
influence on the world economy and, specifically, on Israel, which is a small open
economy, importing oil, and was a gas importer before developing its natural gas
resources.
Tables 1 and 2 present detailed results of conditional predictive ability tests
(CPA) when considering the joint test of both market and expert inflation forecasts
against control variables. In our analysis we focus on significance. To maintain in-
formation on the direction of the effect, our tables report t-statistics rather than
p-values. Terrorism has a significant effect on relative forecast performance inde-
pendent of the control variables included in the model. In all cases, the impact
of terrorism on the loss difference is positive, implying that the quality of market
forecasts deteriorates more strongly in times of high terrorism than that of expert
forecasts.
Our modified test relies on regression coefficients rather than correlations for
the underlying Wald test statistics, and still finds indicators for terrorism signifi-
cant after including financial uncertainty, USD/ILS, or commodity price changes
as control variables. The only exception is the number of attacks, as measured by
GTD, since in most cases, we find significance for neither the number of attacks
nor the additional control, but we do find individual significance for the number
of attacks, without additional controls, and joint significance. We achieve the best
results when using NII indicators, particularly—and surprisingly—the number of
attacks as measured by NII.
Most alternative indicators of uncertainty, that is, financial indicators and com-
modity prices, seem to play a role when assessed individually, but they are not
robust. While we find that most indicators still seem significant for some terror-
ism indicators, those results are not stable and strongly depend on the terrorism
indicator chosen.
The only exception is gas price volatility, which remains significant in all re-
gressions. However, in all those cases, the corresponding terrorism indicator is
still significant, again with the exception of the number of attacks, as reported by
GTD. Based on the overwhelming robustness of our findings for 10 out of 11 ter-
rorism indicators, we find fairly strong evidence that terrorism has an effect on
relative forecast performance.30
These detailed results highlight that forecasters are insignificantly better than
market-based forecasts for 1Y breakeven inflation, but significantly better for 1Y
forward inflation.31
30Other results are available upon request. These tests were conducted with variables expressed
in ILS and USD, and considered a 12-month moving average for the control variable. All results
confirm that terrorism is the best explanatory variable of inflation forecast errors.
31This result is mostly driven by the earliest part of our sample (Second Intifada).
22
GTD MFA NII
Terrorism Control variable CPA terror control CPA terror control CPA terror control
Killed 0.02 0.01 0.02
TA-100 vol. 0.03 1.34 2.47 0.03 1.92 2.34 0.03 1.32 2.46
TA-100 spread 0.03 2.87 0.85 0.02 3.33 1.07 0.03 2.94 1.03
USD/ILS vol. 0.03 2.23 0.82 0.03 2.93 0.95 0.04 2.47 0.96
Oil* vol. 0.03 1.41 2.51 0.03 1.75 2.29 0.04 1.14 2.43
Gas* vol. 0.04 1.93 2.12 0.03 2.57 1.82 0.04 1.70 2.05
CRB* vol. 0.03 3.02 1.64 0.03 3.21 1.60 0.03 2.80 1.63
Oil vol. 0.03 1.54 2.22 0.03 2.32 2.14 0.04 1.48 2.21
Gas vol. 0.04 1.93 2.14 0.03 2.58 1.86 0.04 1.71 2.08
CRB vol. 0.03 2.52 1.20 0.03 2.98 1.26 0.03 2.61 1.32
Wounded 0.01 0.01 0.01
TA-100 vol. 0.03 1.86 2.32 0.03 1.92 2.34 0.02 3.01 2.72
TA-100 spread 0.03 3.18 0.99 0.02 3.33 1.07 0.02 3.02 0.84
USD/ILS vol. 0.03 2.70 0.82 0.03 2.93 0.95 0.02 2.98 0.84
Oil* vol. 0.03 2.02 2.34 0.03 1.75 2.29 0.02 1.97 2.48
Gas* vol. 0.04 2.66 2.09 0.03 2.57 1.82 0.02 2.50 1.89
CRB* vol. 0.03 3.06 1.63 0.03 3.21 1.60 0.02 3.20 1.62
Oil vol. 0.03 2.35 2.11 0.03 2.32 2.14 0.02 2.38 2.26
Gas vol. 0.04 2.68 2.11 0.03 2.58 1.86 0.02 2.49 1.91
CRB vol. 0.03 2.83 1.20 0.03 2.98 1.26 0.02 2.89 1.14
Total 0.01 0.01 0.01
TA-100 vol. 0.03 1.78 2.35 0.03 1.81 2.37 0.02 2.90 2.66
TA-100 spread 0.03 3.21 0.97 0.03 3.29 1.07 0.02 3.24 0.91
USD/ILS vol. 0.03 2.68 0.82 0.03 2.88 0.95 0.02 3.13 0.86
Oil* vol. 0.03 1.93 2.38 0.03 1.64 2.32 0.02 1.97 2.43
Gas* vol. 0.04 2.61 2.10 0.03 2.45 1.87 0.03 2.62 1.89
CRB* vol. 0.03 3.10 1.64 0.03 3.18 1.61 0.02 3.39 1.64
Oil vol. 0.03 2.23 2.13 0.03 2.16 2.15 0.02 2.42 2.22
Gas vol. 0.04 2.63 2.12 0.03 2.46 1.90 0.03 2.61 1.91
CRB vol. 0.03 2.84 1.20 0.03 2.94 1.27 0.02 3.07 1.17
Number 0.02 0.01
TA-100 vol. 0.09 1.14 2.86 0.05 3.45 2.31
TA-100 spread 0.09 0.67 0.47 0.05 3.93 1.09
USD/ILS vol. 0.08 0.34 0.77 0.05 3.52 0.66
Oil* vol. 0.08 0.06 2.84 0.05 2.68 1.63
Gas* vol. 0.04 0.69 2.21 0.04 3.68 1.82
CRB* vol. 0.09 0.94 1.52 Legend 0.05 4.10 1.73
Oil vol. 0.07 -0.38 2.52 1% 0.05 2.76 1.50
Gas vol. 0.04 0.70 2.23 5% 0.04 3.67 1.80
CRB vol. 0.08 0.40 1.12 10% 0.05 3.51 0.96
Table 1: Predictive ability tests of 1-year breakeven and expert inflation forecasts.
Note: CPA is the p-value of the conditional predictive ability test when only a constant is included.
terror is the t-test of the terrorism variable considered. control is the t-test of the control variable
considered. A positive t-test means that expert forecasts are better than market-based ones. A
negative t-test means the opposite. Total is the sum of those killed and wounded, and Number is
the quantity of terrorist attacks. Variables with are in USD.
23
GTD MFA NII
Terrorism Control variable CPA terror control CPA terror control CPA terror control
Killed 0.05 0.06 0.05
TA-100 vol. 0.01 3.01 0.50 0.01 2.61 0.68 0.01 2.76 0.58
TA-100 spread 0.00 3.05 -0.37 0.00 2.69 -0.02 0.00 2.86 0.18
USD/ILS vol. 0.00 3.08 -1.30 0.00 2.62 -0.72 0.00 2.76 -0.64
Oil* vol. 0.04 2.99 0.25 0.04 2.59 0.21 0.04 2.75 0.15
Gas* vol. 0.08 3.03 2.26 0.08 2.56 2.10 0.08 2.74 2.07
CRB* vol. 0.01 3.01 -0.62 0.01 2.63 -0.74 0.01 2.77 -0.63
Oil vol. 0.03 2.97 0.40 0.03 2.59 0.46 0.03 2.75 0.46
Gas vol. 0.08 3.02 2.29 0.07 2.55 2.14 0.07 2.73 2.12
CRB vol. 0.00 3.05 -0.88 0.00 2.64 -0.45 0.00 2.78 -0.35
Wounded 0.04 0.04 0.01
TA-100 vol. 0.01 3.35 -0.16 0.01 3.08 0.20 0.02 2.62 1.40
TA-100 spread 0.00 3.40 -0.24 0.00 3.16 -0.11 0.00 2.59 -0.96
USD/ILS vol. 0.00 3.41 -1.49 0.00 3.10 -0.90 0.01 2.54 -1.35
Oil* vol. 0.04 3.30 -0.08 0.04 3.06 -0.09 0.03 2.54 0.29
Gas* vol. 0.08 3.40 2.20 0.08 3.06 1.80 0.05 2.47 1.91
CRB* vol. 0.01 3.34 -1.10 0.01 3.10 -1.19 0.01 2.56 -1.31
Oil vol. 0.03 3.26 0.12 0.03 3.05 0.32 0.02 2.56 0.55
Gas vol. 0.08 3.40 2.26 0.08 3.05 1.86 0.05 2.45 1.96
CRB vol. 0.00 3.38 -1.18 0.00 3.11 -0.89 0.01 2.58 -1.42
Total 0.04 0.04 0.01
TA-100 vol. 0.01 3.31 -0.03 0.01 3.04 0.27 0.02 2.77 1.31
TA-100 spread 0.00 3.36 -0.25 0.00 3.12 -0.05 0.00 2.75 -0.76
USD/ILS vol. 0.00 3.37 -1.45 0.00 3.05 -0.86 0.01 2.71 -1.26
Oil* vol. 0.04 3.27 -0.01 0.04 3.01 -0.05 0.03 2.69 0.22
Gas* vol. 0.08 3.36 2.22 0.08 3.01 1.85 0.06 2.67 1.90
CRB* vol. 0.01 3.31 -0.98 0.01 3.06 -1.09 0.01 2.72 -1.15
Oil vol. 0.03 3.24 0.18 0.03 3.01 0.34 0.03 2.71 0.50
Gas vol. 0.08 3.36 2.27 0.08 3.00 1.91 0.06 2.65 1.95
CRB vol. 0.00 3.34 -1.11 0.00 3.07 -0.80 0.01 2.74 -1.29
Number 0.00 0.01
TA-100 vol. 0.01 1.55 1.36 0.02 4.79 0.11
TA-100 spread 0.00 1.42 -1.55 0.00 4.88 -1.00
USD/ILS vol. 0.01 1.63 -1.92 0.01 5.40 -2.91
Oil* vol. 0.01 1.40 0.49 0.03 4.71 -1.12
Gas* vol. 0.01 1.50 2.24 0.04 4.91 1.86
CRB* vol. 0.01 1.36 -1.72 Legend 0.01 4.82 -1.63
Oil vol. 0.01 1.32 0.53 1% 0.03 4.70 -1.00
Gas vol. 0.01 1.52 2.28 5% 0.04 4.93 1.89
CRB vol. 0.01 1.52 -1.81 10% 0.01 5.14 -2.47
Table 2: Predictive ability tests of 1-year forward and expert inflation forecasts.
Note: CPA is the p-value of the conditional predictive ability test when only a constant is included.
terror is the t-test of the terrorism variable considered. control is the t-test of the control variable
considered. A positive t-test means that expert forecasts are better than market-based ones. A
negative t-test means the opposite. Total is the sum of those killed and wounded, and Number is
the quantity of terrorist attacks. Variables with are in USD.
24
It seems plausible that the number of terrorist attacks as measured by NII is
the most powerful predictor of forecast accuracy. This result might be related to
the great importance forecasters and markets give to frequency compared with
severity of terrorist attacks (Pizam and Fleischer, 2002). Since the outcome of an
attack in terms of casualties is much more stochastic than its occurrence, it makes
perfect sense that people respond to the number of attacks, regardless of whether
the terrorists claimed many victims. Yet this seems to be contradicted by the poor
performance of the number of attacks as measured by GTD. This might be driven
by the screening methodology of GTD (text mining methodology), which could
cause minor attacks to be excluded, thereby creating a distorted measure of the
number of attacks while still being very accurate in terms of victims.
Similarly, the good performance of gas price volatility is hardly surprising,
given the role natural gas production has played for Israel in the last few years.
Interestingly, this indicates that uncertainty concerning external conditions (oil
and commodity prices) is far less important than factors that might actually affect
production in Israel directly (e.g., gas prices and terrorism).
Forecast error vs. mismeasurement of expectations Technically, neither of the
market-based measures we use is a pure measure of inflation unless we assume
risk-neutral investors. Rather, breakeven inflation corresponds to the inflation ex-
pectation plus an inflation risk premium. If the effect of terrorism were increasing
uncertainty, for example, in terms of the political response,32 this might easily cre-
ate inflation risk. If agents were risk averse, this would affect breakeven inflation.
Even if the accuracy of market expectations were to remain unchanged, the fore-
cast performance of breakeven inflation rates might thereby deteriorate, as they
are no longer an accurate measure of expectations.
In this section, we test whether this mismeasurement drives the deterioration
of the relative forecast performance of breakeven inflation from the sovereign
bond market compared with professional forecasters. To this end, we generate
a measure of inflation uncertainty that accounts for the potential impact of ter-
rorism. We estimate a recursive window, pseudo-out-of-sample, GARCH forecast
of year-over-year inflation. Both the variance of shocks and the process itself are
modeled as an autoregressive moving average model ARMA(1,1) process. We also
include terrorism as an exogenous determinant of shocks.33
Table 3 presents the results of a GARCH(1,1) model. The reported results use
the full sample. However, as mentioned above the time series for predicted volatil-
ity are based on recursive-window out-of-sample predictions.
32See Berrebi and Klor (2008) for an analysis of the causal effects of terrorism on the Israeli
electorate’s preferences.
33We use the best-performing measure from our study, that is, the number of attacks as reported
by the NII.
25
Model with terrorism Model without terrorism
Mean model
Coef. Std. Coef. Std.
Const. 0.859*** 0.281 0.867*** 0.288
AR (1)0.644*** 0.061 0.648*** 0.061
MA (1)0.952*** 0.013 0.952*** 0.013
Variance model
Const. 0.012 0.015 0.021* 0.011
AR (1)0.160** 0.070 0.169** 0.067
MA (1)0.768*** 0.092 0.749*** 0.090
terror 0.001 0.001
Table 3: GARCH model with and without terrorism. Note: Standard errors are robust.
***, **, and * indicate significance at the 1%, 5%, and 10% level, respectively.
The available data on both terrorism and inflation go back much further than
our sample of professional forecasts. Therefore, our out-of-sample predictions for
the variance of inflation span our entire sample as used for the assessment of fore-
cast performance. In the full sample, we see that the impact of terrorism on in-
flation variance is minimal at best, indicating that this is not the channel through
which terrorism affects forecast performance.
A visual inspection of the time series of predicted standard deviation inflation
including and excluding terrorism (Fig. 6) confirms this full sample result for the
recursive window estimation. Most of the time, there is no meaningful difference
between the predictions generated by the two alternative models.
While we find that the standard deviation of expected inflation as measured by
our GARCH has a robust and sizable impact on the relative forecast performance,
terrorism remains robust in explaining relative predictive ability.
In other words, our results are twofold. It seems that terrorism truly affects the
predictive ability of market participants. Even when accounting for risk, we find
that terrorism matters. Nevertheless, there is evidence that inflation risk matters.
This indicates that market participants are far from risk neutral. This makes using
breakeven inflation rates as a substitute for forward-looking measures of inflation
in policymaking even more problematic.
4.2 Exchange rate
Rationality Regarding rationality, results for the USD/ILS exchange rate are
fairly similar to those for inflation. Fig. 7 shows that our rationality test in unstable
environments strongly rejects both expert forecasts and market-implied exchange
rate (USD/ILS) expectations.
26
Apr 1999
Dec 1999
Aug 2000
Apr 2001
Jan 2002
Sep 2002
May 2003
Jan 2004
Oct 2004
Jun 2005
Feb 2006
Nov 2006
Jul 2007
Mar 2008
Nov 2008
Aug 2009
Apr 2010
Dec 2010
Sep 2011
May 2012
Jan 2013
Sep 2013
Jun 2014
Feb 2015
Oct 2015
Jul 2016
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
Predicted standard deviation of inflation
Model with terrorism
Model without terrorism
Figure 6: Predicted standard deviation of inflation from the models with and with-
out terrorism.
The rejection is driven by a strong irrationality between 2007Q2 and 2008Q1
corresponding to the political and military conflicts between Fatah and Hamas
leading to Hamas’ takeover of Gaza (Battle of Gaza, 2007Q2–Q3). Substantial
terrorism and financial uncertainty (the Subprime crisis,34 2007Q2–2008Q4, and
Greek crisis,35 2009Q4–2010Q3) increased around these periods. As Section 4.1
shows, the underlying Wald statistics pick up some movement during the highest
violence level of the Second Intifada.
These rationality tests support the conclusion that terrorism strongly affects
expectations and, thereby, expert and market-based forecasts.
Relative forecast performance and encompassing Like inflation, expert and market-
based exchange rate forecasts are affected by uncertainty of a similar order of mag-
nitude. Neither the fluctuation test (Fig. 8) nor the one-time reversal test rejects,
even at the beginning of the period (Second Intifada). This confirms that at any
point in time, expert and market-based forecast performances are fairly similar (at
34In the the United States, Cecchetti (2009) and Mishkin (2010) consider that the crisis began in
2007Q1, when several large subprime mortgage lenders started to report losses. The real trigger for
the crisis was in 2007Q3, when the French bank BNP Paribas temporarily suspended redemptions
from three of its fund holdings that had invested in assets backed by US subprime mortgage debt.
As a result, credit spreads began widening, interest rates in Europe shot up overnight, and the
European Central Bank immediately responded with the largest short-term liquidity injection in
its nine-year history (Benchimol and Fourçans, 2017).
35At the end of 2009, the three main rating agencies downgraded Greece’s credit rating, and
during the first two quarters of 2010, three austerity packages were announced by the Greek gov-
ernment. These decisions, also known as a cause of the European Debt Crisis, strongly impacted
the euro area, Israel’s main trade partner (Benchimol, 2016).
27
(a) Expert forecast (b) Market-based forecast
Figure 7: Rationality test statistic (solid) and the critical value (dashed) for
USD/ILS. Note: Significance level: 5%. See Fig. 3 for an explanation of the labels.
the 5% significance level).
Fig. 8 shows the dynamics of the underlying individual test statistics over
time. There is some indication that expert forecasts were superior at the begin-
ning of the sample, yet the test statistics are far below the critical value. Contrary
to inflation forecasts (Section 4.1), market-based forecasts were slightly superior
to expert forecasts between 2005Q1 and 2008Q4. This period corresponds to a
relatively calm period. The two transition points between superiority of market-
based and expert forecasts occurred at very specific times. The first corresponds to
strong fears involving the Hamas victory in the elections for the second Palestinian
Legislative Council, and the second corresponds to Operation Cast Lead (2008Q4–
2009Q1). While the first period heralded increasing fears in and around Israel36
(and corresponded to the switch from the superiority of expert forecasts to superi-
ority of market-based forecasts), the second event promised a period of increasing
stability37 (and corresponded to the switch from the superiority of market-based
forecasts to superiority of expert forecasts). Operation Pillar of Defense (2012Q4)
contributed to the performance deterioration of expert forecasts relative to market-
based forecasts.
The encompassing test (Fig. 9) highlights the clear rejection that the typically
superior expert forecasts encompass market forecasts. However, these rejections
36The Battle of Gaza resulted in Hamas taking control of the Gaza Strip from Fatah (2007Q2); the
Israeli military’s launch of Operation Hot Winter (2008Q1); and regular, violent terrorist attacks
until Operation Cast Lead (2008Q4–2009Q1).
37This relatively calm period lasted from 2009Q2 until Operation Returning Echo (2012Q1), car-
ried out to stop cross-border attacks, as mortars and rocket fire had started several quarters before.
28
(a) Fluctuation test (b) One-time reversal test
Figure 8: Fluctuation and one-time reversal test statistics (solid), and the criti-
cal value and average loss differences (dashed) in (a) and (b), respectively, for
USD/ILS. Note: Significance level: 5%. See Fig. 3 for an explanation of the labels.
are not driven by the same periods that drive the results of our rationality test.
Instead, we reject encompassing fairly consistently until 2008Q2. This indicates
that—for the exchange rate—markets constantly monitor and include factors that
are not properly accounted for by professionals. While this does not contradict our
hypothesis regarding the importance of terrorism, it does not provide sufficient
support for it either.
Conditional relative forecast performance Interestingly, although the initial tests
provide only weak evidence of fluctuations in relative forecast performance, the
tests for conditional forecast performance have different implications.
Generally, the results are not quite as clear as they were for inflation. While
some conditional predictive ability tests reject some terrorism indicators at the
10% level, we fail to reject them in other cases.
The picture changes when accounting for other sources of uncertainty. While
the joint test now generally fails to reject the indicators because we add a less
powerful indicator, the individual t-tests for the terrorism indicators indicate con-
sistent rejections.
Table 4 presents detailed results of the CPA tests when considering the joint
test of both market and expert USD/ILS forecasts against control variables.
We find again that both terrorism and financial uncertainty explain variations
in the exchange rate’s (USD/ILS) relative forecast performance. In particular, we
find that market forecasts deteriorate more strongly in times of uncertainty com-
pared with expert forecasts. We find evidence that terrorism truly matters for
29
GTD MFA NII
Terrorism Control variable CPA terror control CPA terror control CPA terror control
Killed 0.09 0.11 0.12
TA-100 vol. 0.23 4.68 -0.29 0.25 3.46 -0.05 0.25 3.12 -0.04
TA-100 spread 0.23 4.62 0.44 0.26 3.66 0.80 0.27 3.26 0.67
USD/ILS vol. 0.21 4.56 1.16 0.22 3.75 1.47 0.22 3.38 1.47
Oil* vol. 0.23 4.91 -1.87 0.28 3.71 -1.99 0.29 3.34 -1.92
Gas* vol. 0.25 4.74 0.12 0.30 3.59 0.08 0.31 3.24 0.11
CRB* vol. 0.24 4.63 0.66 0.27 3.49 0.51 0.28 3.15 0.46
Oil vol. 0.24 4.78 -1.57 0.28 3.57 -1.39 0.29 3.21 -1.16
Gas vol. 0.25 4.74 0.08 0.30 3.59 0.05 0.31 3.24 0.09
CRB vol. 0.23 4.62 0.79 0.24 3.66 0.99 0.25 3.30 0.97
Wounded 0.10 0.17 0.39
TA-100 vol. 0.24 4.37 -0.73 0.28 2.06 -0.09 0.16 0.24 0.65
TA-100 spread 0.25 4.26 0.15 0.35 2.11 -0.27 0.55 0.09 -1.40
USD/ILS vol. 0.21 4.16 0.97 0.26 2.22 1.20 0.24 0.24 1.09
Oil* vol. 0.26 4.69 -2.50 0.36 2.26 -2.05 0.25 0.26 -0.84
Gas* vol. 0.28 4.41 0.17 0.39 2.17 0.15 0.39 0.17 0.76
CRB* vol. 0.26 4.27 0.08 0.35 2.12 -0.21 0.31 0.22 -0.71
Oil vol. 0.27 4.48 -1.84 0.37 2.16 -0.97 0.32 0.24 -0.22
Gas vol. 0.28 4.40 0.14 0.39 2.17 0.12 0.40 0.18 0.70
CRB vol. 0.24 4.24 0.52 0.30 2.16 0.61 0.29 0.23 0.45
Total 0.10 0.16 0.33
TA-100 vol. 0.24 4.50 -0.66 0.28 2.29 -0.10 0.21 0.50 0.62
TA-100 spread 0.24 4.39 0.23 0.33 2.36 -0.07 0.52 0.37 -1.32
USD/ILS vol. 0.21 4.30 1.00 0.25 2.48 1.24 0.26 0.52 1.08
Oil* vol. 0.25 4.80 -2.37 0.35 2.52 -2.10 0.32 0.53 -0.98
Gas* vol. 0.27 4.54 0.15 0.37 2.42 0.12 0.42 0.45 0.66
CRB* vol. 0.25 4.40 0.20 0.33 2.35 -0.09 0.35 0.48 -0.56
Oil vol. 0.26 4.60 -1.82 0.35 2.40 -1.06 0.37 0.51 -0.31
Gas vol. 0.27 4.54 0.12 0.38 2.42 0.09 0.43 0.45 0.61
CRB vol. 0.23 4.37 0.58 0.29 2.40 0.67 0.32 0.50 0.47
Number 0.12 0.14
TA-100 vol. 0.28 1.11 0.63 0.29 1.45 0.22
TA-100 spread 0.26 1.08 -1.24 0.33 1.47 -1.17
USD/ILS vol. 0.27 0.88 0.93 0.28 1.30 0.62
Oil* vol. 0.27 1.12 -1.16 0.32 1.61 -1.60
Gas* vol. 0.29 1.08 0.81 0.34 1.45 0.63
CRB* vol. 0.26 1.03 -0.56 Legend 0.33 1.49 -0.67
Oil vol. 0.29 1.12 -0.59 1% 0.35 1.58 -0.84
Gas vol. 0.29 1.08 0.76 5% 0.34 1.46 0.58
CRB vol. 0.29 1.04 0.29 10% 0.33 1.46 0.01
Table 4: Predictive ability tests of 1-year forward and expert USD/ILS forecasts.
Note: CPA is the p-value of the conditional predictive ability test when only a constant is included.
terror is the t-test of the terrorism variable considered. control is the t-test of the control variable
considered. A positive t-test means that expert forecasts are better than market-based ones. A
negative t-test means the opposite. Total is the sum of those killed and wounded, and Number is
the quantity of terrorist attacks. Variables with are in USD.
30
Figure 9: Encompassing test statistic (solid) and the critical value (dashed) for
USD/ILS. Note: Significance level: 5%. See Fig. 3 for an explanation of the labels.
relative forecast performance, and the effect of financial uncertainty becomes in-
significant when controlling for terrorism. Neither oil nor gas price fluctuations
play a stronger role than terrorism.38
Given that the variation in forecast performance is quite explicable (not ran-
dom), we believe that it is the lack of power of the fluctuation test in small sam-
ples, rather than the constant relative performance, that causes us to fail to reject
the null hypothesis.
Table 4 indicates that the USD/ILS forecast performance is strongly related
to terrorism data. Surprisingly, the NII data now seem clearly less powerful in
explaining forecast performance. In particular, the number of attacks, which out-
performed all other measures for inflation, is now completely void of predictive
power. On the contrary, the number of people killed now plays a major role. The
reason might be that—by its nature—international traders are much more active
in the exchange market than they are in the Israeli bond market. Due to the large
number of attacks in Israel, low-casualty terrorism barely receives international
mention these days. This might explain why the number of casualties (i.e., people
killed in attacks) is much more important for the exchange market.
As terrorism has a clear and identified impact on financial and real asset mar-
kets (Zussman et al., 2008; Elster et al., 2017), our finding adds to the literature by
showing that terrorism has a substantial impact on forward markets, involving a
significant loss in predictive ability.
Table 4 shows that the Tel Aviv Stock Exchange (TASE) spread, oil volatility
in USD and ILS, and CRB volatility in USD t-tests are generally negative over all
38As a robustness check, we also conduct this exercise with terrorism data excluding Gaza and
West Bank events (GTD). Our results are similar and available upon request.
31
our terrorism indicators, indicating that expert forecasts, relative to market fore-
casts, performance deteriorates in times of uncertainty (i.e., when the terrorism
indicator’s t-test is high).
5 Interpretation
Expert forecasts of inflation are superior to market-based forecasts during periods
of high uncertainty. The result is qualitatively similar for exchange rate forecasts.
In both cases, we find strong evidence that it is indeed terrorism, rather than re-
lated phenomena (e.g., commodity price fluctuations and financial distress), that
triggers changes in forecast performance. Moreover, we find in both cases that the
performance of market forecasts deteriorates more strongly.
These results are not surprising. As terrorist attacks affect agents’ behavior,
market and economic expectations change accordingly. Expert and market fore-
casts are then affected in several ways. First, such changes increase forecasting
bias and errors. The drastic impact of terrorism seriously affects market player and
forecaster perceptions, which in turn influence their implied forecasts. Second,
such events substantially affect the predictive ability of these forecast providers
and their updates following the events, at least for the Israeli inflation and ex-
change rate (USD/ILS) forecasts.
However, the impact on relative performance indicates that there is more to
the story–an additional fundamental variable that adds some uncertainty. It seems
that these events disturb the predictive ability of market participants considerably
more than they affect professional forecasters. Again, this is hardly surprising. It
makes sense that market participants, in a wider sense, are more affected by the
psychological impact of uncertainty than are experienced professionals.
There are many potential reasons why market forecasts respond differently
than professional forecasts. Although a large share of investors (e.g., investment
banks) pay attention to the inflation forecasts of professional forecasters, who in
turn also keep an eye on the market, the behavior of professional forecasters is
singular. Indeed, their objective functions (e.g., in terms of risk aversion) and the
time to consider updates of their forecasts lead to significant differences compared
to the market-based forecasts. An additional factor–information asymmetry–also
contributes to explaining the effect on exchange rate forecasts. Exchange-rate fore-
casts are supposedly more closely related to international markets, and foreign
participants may have different news sources or even preferences.
Interestingly, financial uncertainty is not significant when controlling for ter-
rorism, although terrorism remains robustly significant when controlling for fi-
nancial uncertainty. In other words, terrorism has a major role in the bias and
predictive ability of both expert and market-based forecasts. Interpreting these
32
forecasts without considering the current terrorism situation could lead to severe
mistakes for decision makers.
In line with the recent literature about the impact of oil prices on the US,
French, and UK inflation forecasts (Badel and McGillicuddy, 2015; Bec and De Gaye,
2016), we show that, for Israel, exchange rate volatility and commodities volatility
matter in explaining inflation forecast errors (without considering terrorism data).
However, our results suggest that in the case of Israel, terrorism has a strong ex-
planatory power for both expert and market-based forecast errors compared with
other control variables (Section 2.3).
Both terrorism and the volatility of natural gas prices have significant effects on
the exchange rate forecasts. This link is a consequence of the three ways in which
natural gas has influenced the Israeli economy since the 2000s. In the late 1990s,
the Israeli government decided to encourage the use of natural gas for several
reasons (environmental, cost, and resource diversification). From 2000 to 2016,
Israeli natural gas consumption multiplied more than 10-fold (from 0.01 to 0.12
quadrillion BTU). Thus, Israel became an important natural gas consumer. At
the same time, several natural gas reserves were discovered in Israel, making Is-
rael one of the 30 natural gas-exporting countries in the world in 2016. Before
this, Israel acquired natural gas through Egyptian pipelines in the Sinai Peninsula,
which were frequently targeted by terrorist organizations (e.g., during the August
2012 Sinai attacks). The prominent link between natural gas price volatility and
USD/ILS forecasts resides in the Bank of Israel’s program to offset the effects of
natural gas production on the foreign exchange rate by purchasing foreign cur-
rency.
Another layer was added to the Bank of Israel’s exchange rate policy in May
2013, following rapid appreciation of the nominal effective exchange rate of ILS
(about 11.5% between September 2012 and May 2013). At the beginning of the
process, the appreciation was influenced by the geopolitical situation, due to the
dissipation of tension prevalent at the beginning of the period. As the start date
for natural gas production from the Tamar site drew closer, the foreign exchange
market started to price in the expected return from decades of natural gas produc-
tion. The Bank of Israel reacted by launching its natural gas offset program.
These three factors are strongly related to imported inflation and the exchange
rate (USD/ILS). As the Bank of Israel tried to offset the appreciation of the Shekel
due to natural gas production and exports, natural gas price volatility became a
robust control variable, considering Israeli inflation as well as USD/ILS exchange
rate forecasts.
As mentioned before, the implied market forecasts are actually a combination
of market expectations and risk premia.39 While we aim to control for that effect,
39Including the liquidity premium reflecting the different liquidity of nominal and inflation-
33
we can only partly do so; that is, we can account for the inflation risk itself (Section
4.1). Other factors that might be affected in times of economic uncertainty, such as
liquidity risk, have to be omitted due to lack of data.40
6 Policy implications
Like the central banks of other developed economies, the Bank of Israel uses market-
based as well as expert forecasts to back its own forecasts used for monetary policy
decisions. These forecasts are also of prime importance because of their utilization:
they are presented and discussed by the MPC and by decision makers and institu-
tions in Israel and abroad (e.g., the International Monetary Fund, Organisation for
Economic Co-operation and Development, and the World Bank). From a policy
perspective, it is thus of minor importance whether our market forecasts are truly
forecasts in the original sense of the word. What matters is that they are treated as
such by policymakers despite their shortcomings.
Our results show that using or interpreting these forecasts (expert and market-
based), without considering terrorism in Israel can cause an incorrect perception of
their current predictive power. Underestimating forecast errors and correspond-
ingly overestimating predictive ability could lead to severe over-reliance on incor-
rect forecasts. Consequently, when establishing their inflation and exchange rate
projections, these institutions and decision makers should interpret expert and
market-based forecasts differently, conditional on the current level of terrorism.
For example, between 2008Q4–2010Q3, these institutions should have considered
a persistent and robust bias in expert and market-based forecasts of inflation.41
While prior literature and policy discussions have considered economic factors
that influence inflation expectations (and thus, the forward-looking estimates of
the inflation gap), such as financial instability and commodity prices, these factors
have not included terrorism despite its apparently far greater importance.
linked government bonds.
40Greater war risks narrow the breakevens and affect many other financial asset prices (Rigobon
and Sack, 2005). Narrowing of the US breakevens during the global financial crisis represented
an investor preference for the liquidity of nominal government bonds (Fleckenstein et al., 2014;
Pflueger and Viceira, 2016). The difference between forward exchange rates and actual physical
expectations (foreign exchange risk premium) could also represent compensation for facing disas-
ter risk (Farhi et al., 2009), reinforcing our conclusions for Israel.
41Flexible inflation targeters, such as the Bank of Israel, present their monetary policy objectives
in terms of the path of the inflation gap they are willing to tolerate following a cost-push shock
until the economy moves back to the inflation target. In practice, central banks formulate the
normative trade-off between inflation and output variability in this natural and intuitive way, thus
improving communication with the general public (Cukierman, 2015). This inflation gap makes
use of market-based and expert inflation forecasts to make a decision today on the monetary policy
rate to be implemented tomorrow. Therefore, underestimating forecast error and overestimating
predictive ability performance during periods of terrorism could result in errors/bias in critical
monetary policy decisions.
34
The same argument holds true for exchange rate forecasts. During the last
few decades, Israeli monetary policy was strongly influenced by the USD/ILS ex-
change rate. Although some new challenges have emerged, the exchange rate still
has an impact on the monetary policy decision process. Thus, when policymak-
ers evaluate the future path of the Israeli economy and consider the market-based
and expert exchange rate forecasts, they should also consider the current terrorism
situation to avoid bias or low predictive ability in the forecasts they use.
Risk matters to some degree when considering breakeven inflation as a mea-
sure of expectations.42 That said, terrorism still seems to affect the actual underly-
ing expectation. This result should be considered when the MPC assesses inflation
forecasts. For instance, during the period September 2005 to August 2006 rockets
were launched from Gaza and Lebanon (Period T1). During this time, inflation
expectations rose and as a consequence of that policy makers dealt with possible
transmission from inflation expectations to actual inflation, calling for a rise in in-
terest rates. During that period interest rates had been risen by 2 percent, while
actual inflation remained approximately constant. The lesson from this paper is
that the extent of that rise shall be discussed in real time under the light of a pos-
sible bias in inflation expectations. Terrorism may have reinforced this bias.
When policymakers evaluate the future path of inflation and the exchange
rates in Israel, they should give higher weight to expert forecasts during periods
of terrorism. However, the encompassing tests suggest that market-based fore-
casts should not be dismissed entirely, as they do provide some information. The
difference in mean difference of squared errors is 1.86 in times of high terrorism
(above median), and 0.31 in times of low terrorism (below median). In short, ex-
pert forecasts are better in both regimes. Still, there can be a benefit to including
the worse forecast if it is not encompassed. However, this seems to be mostly the
case in times of extremely high terrorism as in the early part of our sample. Given
that this period is inevitably part of the initial training period of a pseudo out of
sample forecasting exercise, we are unable to quantify the benefits using our fairly
short sample.
Our findings provide two practical solutions for the policymaker concerning
decisions using forecasts extracted after terrorist attacks. The first is to assess the
forecasts published just before the terrorist attack, i.e., the forecasts prevailing in a
period that was not subject to terrorism, and to give them greater emphasis in the
current decision process compared to the forecasts published in times of terrorism.
The second is to prefer expert-based rather than market-based forecasts in this
configuration.
42Our finding that terrorism affects market-based inflation forecasts remains robust when con-
trolling for inflation risk. Nevertheless, we find that risk measures have a substantial impact on
breakeven inflation forecasts.
35
7 Conclusion
As a small, developed, open economy implementing an inflation-targeting mone-
tary policy in the context of financial instability and terrorism, Israel remains the
best laboratory for analyzing inflation and exchange rate forecasts.
The consequences of terrorist attacks on expert as well as market-based fore-
cast performance are absent from the literature. This study fills that gap by show-
ing that terrorism had a significant impact on inflation and exchange rate forecast
errors in Israel over the last 15 years. Moreover, under all types of uncertainties
analyzed, expert inflation forecasts are generally better than breakeven (market-
based) inflation forecasts.
When considering the number of terrorist attacks, the picture is very clear:
whatever the type of inflation forecast, the number of terrorist attacks has the best
explanatory power for the relative predictive ability of the forecasts we consider.
Terrorism has a significant impact on both components of market-based inflation
forecasts (inflation forecast as well as its risk premium), even if agents are not risk
neutral.
We also show that oil and TASE-100 control variables are sometimes found to
affect inflation forecasts, a result in line with Bec and De Gaye (2016), although
this finding depends strongly on the terrorism indicator included in the model.
Exchange rate forecasts are more strongly affected by the number of fatalities
from terrorist attacks, whatever the quantitative methodology for accounting for
them. We relate this to the fact that external market participants in the exchange
market give higher weight to attacks in which human lives are lost.
Uncertainty in general, and terrorism in particular, alters the forecasting per-
formance of market participants much more than professional forecasters’ one. At
least in the case of Israel, the weak average performance of market participants
seems to be mostly driven by those episodes.
Policymakers should pay attention to market-based forecasts and prefer expert
forecasts during periods of terrorism. Forecasters’ experience and low-frequency
of updating, i.e., less subject to overreaction, could play a role in their superior
predictive ability compared with market-based forecasts.
References
Abadie, A., Gardeazabal, J., 2003. The economic costs of conflict: a case study of
the Basque Country. American Economic Review 93 (1), 113–132.
Adeney, R., Arsov, I., Evans, R., 2017. Inflation expectations in advanced
economies. RBA Bulletin March Quarter, Reserve Bank of Australia.
36
Badel, A., McGillicuddy, J., 2015. Oil prices and inflation expectations: is there a
link ? Regional Economist Federal Reserve Bank of St. Louis (July).
Bandyopadhyay, S., Sandler, T., Younas, J., 2014. Foreign direct investment, aid,
and terrorism. Oxford Economic Papers 66 (1), 25–50.
Barnett, A., Mumtaz, H., Theodoridis, K., 2014. Forecasting UK GDP growth and
inflation under structural change. A comparison of models with time-varying
parameters. International Journal of Forecasting 30 (1), 129–143.
Bec, F., De Gaye, A., 2016. How do oil price forecast errors impact inflation fore-
cast errors? An empirical analysis from US, French and UK inflation forecasts.
Economic Modelling 53 (C), 75–88.
Benchimol, J., 2016. Money and monetary policy in Israel during the last decade.
Journal of Policy Modeling 38 (1), 103–124.
Benchimol, J., Fourçans, A., 2017. Money and monetary policy in the Eurozone: an
empirical analysis during crises. Macroeconomic Dynamics 21 (3), 677–707.
Berrebi, C., Klor, E. F., 2008. Are voters sensitive to terrorism? Direct evidence
from the Israeli electorate. American Political Science Review 102 (03), 279–301.
Blomberg, S. B., Hess, G. D., Orphanides, A., 2004. The macroeconomic conse-
quences of terrorism. Journal of Monetary Economics 51 (5), 1007–1032.
Caraiani, P., 2016. The role of money in DSGE models: a forecasting perspective.
Journal of Macroeconomics 47 (B), 315–330.
Caruso, R., Klor, E. F., 2012. Political Economy Studies on the Israeli-Palestinian
Conflict: Introduction. Peace Economics, Peace Science and Public Policy 18 (2).
Cecchetti, S., 2009. Crisis and responses: the Federal Reserve in the early stages of
the financial crisis. Journal of Economic Perspectives 23 (1), 51–75.
Chen, A. H., Siems, T. F., 2004. The effects of terrorism on global capital markets.
European Journal of Political Economy 20 (2), 349–366.
Choi, S., Furceri, D., Loungani, P., Mishra, S., Poplawski-Ribeiro, M., 2018.
Oil prices and inflation dynamics: Evidence from advanced and developing
economies. Journal of International Money and Finance 82 (C), 71–96.
Crain, N. V., Crain, W. M., 2006. Terrorized economies. Public Choice 128 (1-2),
317–349.
Cukierman, A., 2015. The choice of flexibility in targeting inflation during normal
times and during disinflations. Research in Economics 69 (4), 494–502.
37
de Haan, L., van den End, J. W., 2018. The signalling content of asset prices for
inflation: Implications for quantitative easing. Economic Systems 42 (1), 45–63.
Diebold, F. X., Mariano, R. S., 1995. Comparing predictive accuracy. Journal of
Business & Economic Statistics 13 (3), 253–63.
Dorsett, R., 2013. The effect of the troubles on GDP in Northern Ireland. European
Journal of Political Economy 29 (C), 119–133.
Dreher, A., Fischer, J. A., 2010. Government decentralization as a disincentive for
transnational terror ? An empirical analysis. International Economic Review
51 (4), 981–1002.
Dreher, A., Fuchs, A., 2011. Does terror increase aid ? Public Choice 149 (3), 337–
363.
Dreher, A., Gassebner, M., 2008. Does political proximity to the U.S. cause terror ?
Economics Letters 99 (1), 27–29.
Dreher, A., Gassebner, M., Siemers, L.-H., 2010. Does terrorism threaten human
rights ? Evidence from panel data. Journal of Law and Economics 53 (1), 65–93.
Eckstein, Z., Tsiddon, D., 2004. Macroeconomic consequences of terror: theory and
the case of Israel. Journal of Monetary Economics 51 (5), 971–1002.
El-Shagi, M., 2019. Rationality tests in the presence of instabilities in finite samples.
Economic Modelling 79, 242–246.
El-Shagi, M., Giesen, S., Jung, A., 2016. Revisiting the relative forecast perfor-
mances of Fed staff and private forecasters: A dynamic approach. International
Journal of Forecasting 32 (2), 313–323.
Eldor, R., Melnick, R., 2004. Financial markets and terrorism. European Journal of
Political Economy 20 (2), 367–386.
Elster, Y., Zussman, A., Zussman, N., 2017. Rockets: the housing market effects of
a credible terrorist threat. Journal of Urban Economics 99 (C), 136–147.
Farhi, E., Fraiberger, S. P., Gabaix, X., Ranciere, R., Verdelhan, A., 2009. Crash
risk in currency markets. NBER Working Papers 15062, National Bureau of Eco-
nomic Research.
Farhi, E., Gabaix, X., 2016. Rare disasters and exchange rates. Quarterly Journal of
Economics 131 (1), 1–52.
Fielding, D., 2003a. Counting the cost of the Intifada: consumption, saving and
political instability in Israel. Public Choice 116 (3-4), 297–312.
38
Fielding, D., 2003b. Modelling political instability and economic performance: Is-
raeli investment during the Intifada. Economica 70 (277), 159–186.
Fleckenstein, M., Longstaff, F. A., Lustig, H., 2014. The TIPS-Treasury bond puzzle.
Journal of Finance 69 (5), 2151–2197.
Gerlach, J. R., Yook, Y., 2016. Political conflict and foreign portfolio investment:
Evidence from North Korean attacks. Pacific-Basin Finance Journal 39, 178–196.
Giacomini, R., Rossi, B., 2010. Forecast comparisons in unstable environments.
Journal of Applied Econometrics 25 (4), 595–620.
Giacomini, R., White, H., 2006. Tests of conditional predictive ability. Economet-
rica 74 (6), 1545–1578.
Hawkins, D., 1987. A test for a change point in a parametric model based on a
maximal Wald-type statistic. Sankhya A: The Indian Journal of Statistics 49 (3),
368–376.
Johnston, R., 2016. Terrorism, counterterrorism, and unconventional warfare.
Tech. rep., Johnston’s Archive.
URL http://www.johnstonsarchive.net/terrorism/terrisraelsum.html
Karl, M., Winder, G., Bauer, A., 2017. Terrorism and tourism in Israel: analysis of
the temporal scale. Tourism Economics 23 (6), 1343–1352.
Keefer, P., Loayza, N., 2008. Overview: terrorism, economic development, and
political openness. In: Keefer, P., (Eds.), N. L. (Eds.), Terrorism, economic devel-
opment, and political openness. Cambridge, UK: Cambridge University Press,
pp. 1–14.
Kollias, C., Papadamou, S., Stagiannis, A., 2011. Terrorism and capital markets:
The effects of the Madrid and London bomb attacks. International Review of
Economics & Finance 20 (4), 532–541.
Larocque, D., Lincourt, G., Normandin, M., 2010. Macroeconomic effects of terror-
ist shocks in Israel. Defence and Peace Economics 21 (4), 317–336.
Marsden, S. V., 2012. Successful terrorism: framework and review. Behavioral Sci-
ences of Terrorism and Political Aggression 4 (2), 134–150.
Mincer, J. A., Zarnowitz, V., 1969. The evaluation of economic forecasts. In: Eco-
nomic Forecasts and Expectations: Analysis of Forecasting Behavior and Perfor-
mance. NBER Chapters. National Bureau of Economic Research, pp. 3–46.
39
Mishkin, F. S., 2010. Over the cliff: from the Subprime to the Global Financial
Crisis. Journal of Economic Perspectives 25 (1), 49–70.
Newey, W. K., West, K. D., 1987. A simple, positive semi-definite, heteroskedastic-
ity and autocorrelation consistent covariance matrix. Econometrica 55 (3), 703–
708.
Pavlova, A., Rigobon, R., 2007. Asset prices and exchange rates. Review of Finan-
cial Studies 20 (4), 1139–1180.
Peleg, K., Regens, J. L., Gunter, J. T., Jaffe, D. H., 2011. The normalisation of ter-
ror: The response of Israel’s stock market to long periods of terrorism. Disasters
35 (1), 268–283.
Pflueger, C. E., Viceira, L. M., 2016. Return predictability in the Treasury market:
real rates, inflation, and liquidity. In: Handbook of Fixed-Income Securities.
John Wiley & Sons, pp. 191–209.
Pizam, A., Fleischer, A., 2002. Severity versus frequency of acts of terrorism: which
has a larger impact on tourism demand ? Journal of Travel Research 40 (3), 337–
339.
Rigobon, R., Sack, B., 2005. The effects of war risk on US financial markets. Journal
of Banking & Finance 29 (7), 1769–1789.
Roberts, B. W., 2009. The macroeconomic impacts of the 9/11 attack: evidence
from real-time forecasting. Peace Economics, Peace Science and Public Policy
15 (2), 1–29.
Romanov, D., Zussman, A., Zussman, N., 2012. Does terrorism demoralize? Evi-
dence from Israel. Economica 79 (313), 183–198.
Rossi, B., Sekhposyan, T., 2016. Forecast rationality tests in the presence of insta-
bilities, with applications to Federal Reserve and survey forecasts. Journal of
Applied Econometrics 31 (3), 507–532.
Ruiz Estrada, M. A., Koutronas, E., 2016. Terrorist attack assessment: Paris No-
vember 2015 and Brussels March 2016. Journal of Policy Modeling 38 (3), 553–
571.
Sandler, T., Enders, W., 2008. Economic consequences of terrorism in developed
and developing countries: an overview. In: Keefer, P., Loayza, N. (Eds.), Ter-
rorism, economic development, and political openness. Cambridge, UK: Cam-
bridge University Press, pp. 17–47.
40
Scotti, C., 2016. Surprise and uncertainty indexes: Real-time aggregation of real-
activity macro-surprises. Journal of Monetary Economics 82 (C), 1–19.
Shahbaz, M., 2013. Linkages between inflation, economic growth and terrorism in
Pakistan. Economic Modelling 32 (C), 496–506.
Shoham, A., Rosenboim, M., Malul, M., Saadon, Y., 2011. Core and periphery - the
dual effect of terror. Peace Economics, Peace Science, and Public Policy 17 (1),
1–15.
Stock, J. H., Watson, M. W., 2003. Forecasting output and inflation: the role of asset
prices. Journal of Economic Literature 41 (3), 788–829.
Wallace, J., Wild, S. E., 2010. Webster’s New World Law Dictionary. Hoboken, NJ:
Wiley.
West, K. D., McCracken, M. W., 1998. Regression-based tests of predictive ability.
International Economic Review 39 (4), 817–840.
Zussman, A., Zussman, N., Nielsen, M. O., 2008. Asset market perspectives on the
Israeli-Palestinian conflict. Economica 75 (297), 84–115.
41
