News and Intraday Retail Investor Order Flow in Foreign Exchange Markets

Abstract

This paper examines the trading behavior of individual investors using a proprietary intraday dataset of a large pool of retail investor aggregate (minute by minute) long and short positions in EUR/USD for the period July 2014 to April 2016. Standard event study analysis shows no significant adjustment in trading ahead of scheduled macro news announcements and trading contrary to the announcement surprise after the event. A panel regression analysis shows that the contrarian trading behavior is mostly driven by lagged returns rather than fundamental macro news. Further intraday time series analysis shows that the lagged overall news sentiment also significantly affects investor net order flow. Finally, we show that simple cross-over trading strategies that exploit retail investor order flow imbalance are profitable. Overall our results suggest that retail investors in currency markets follow the news and are influenced by news sentiment and past returns, but do not possess the skills to extract fundamental information from public news. Their trading behavior contributes to the slowdown of the price discovery process. Our findings support the differential abilities of market participants to interpret public information as an explanation for the importance of order flow in price formation in currency markets.

Full text

1
News and Intraday Retail Investor Order Flow in Foreign Exchange Markets
1
Theofilia Kaourma, Andreas Milidonis, George Nishiotis, Marios Panayides
April 2024
Abstract
This paper examines the trading behavior of individual investors using a proprietary intraday dataset
of a large pool of retail investor aggregate (minute by minute) long and short positions in EUR/USD.
Standard event study analysis shows no significant adjustment in trading ahead of scheduled macro
news announcements and trading contrary to the announcement surprise after the event. A panel
regression analysis shows that such contrarian trading behavior is mostly driven by lagged returns
rather than fundamental macro news. Further intraday time series analysis shows that the lagged
overall news sentiment also significantly affects retail investor trading. Finally, to verify the
uninformed nature of retail trading, we show that simple cross-over trading strategies that exploit
retail investors’ order flow could be profitable. Overall, our results suggest that retail investors in
currency markets are influenced by news sentiment and past returns, but do not possess the skills to
extract fundamental information from public news. Our findings support the differential abilities of
market participants to interpret public information as an explanation for the importance of order
flow in price formation in currency markets.
Keywords: Foreign Exchange, Behavioral Finance, Retail Investors, Intraday Order Flow, TRMI, News
Sentiment, Scheduled Macro News Announcements
JEL classifications: F31, G12, G14
We thank, Elena Andreou, Irene Karamanou, George Korniotis, Andrei Simonov, Lenos Trigeorgis, Stavros Zenios and
seminar participants at the School of Economics and Management of University of Cyprus for helpful comments.
Kaourma, Nishiotis and Milidonis are from the University of Cyprus with a contact address: Department of Accounting
& Finance, School of Economics & Management, University of Cyprus, P.O. Box 20537, CY-1678 Nicosia, Cyprus. The
authors contact emails and phone numbers are kaourma.theofilia@ucy.ac.cy (+357 22893655),and
andreas.milidonis@ucy.ac.cy (+357 22893626). Panayides is from the University of Oklahoma and the University of
Cyprus and can be reached at panayides.marios@ucy.ac.cy and marios.panayides@ou.edu.

2
1. Introduction
In recent years retail investors’ participation in foreign exchange (FX) markets has become
substantial. The Bank of International Settlements (BIS) Triennial Central Bank Survey (2022) reports
that in April 2022 the retail-driven average daily over the counter (OTC) foreign exchange turnover is
US$193 billion
2
, which corresponds to 22.8% of the turnover by institutional investors (US$847
billion) and to 37.5% of the turnover by hedge funds and proprietary trading firms (US$514 billion).
The increased retail investor participation in FX markets is facilitated by low transaction costs and
the availability of substantial leverage through financial intermediaries acting as retail aggregators
(see King et al., 2012). Retail investors may be unsophisticated-noise traders (Kyle, 1985)
3
, but an
aggregation of their trading activity may significantly influence financial markets (Barber et al., 2009).
The recent short squeeze inflicted by retail investors on hedge fund managers provides an extreme
firsthand example of the impact that retail investors can have on financial markets.
4
This paper addresses three questions regarding the intra-day trading behavior of retail investors in
FX markets. First, how do retail investors trade around scheduled macro news announcements? For
example, do they trade with or against the direction of the news and do they possess the skills
necessary to extract the fundamental information from these announcements? Second, how does
the broader intraday news sentiment affect retail investor intraday order flow?
5
We extend our
analysis to go beyond the scheduled macro news announcements and we consider the overall
2
The retail-driven average daily OTC foreign exchange turnover is approximately equal to the daily average global equities
trading volume, which according to a NASDAQ article is approximately $200 billion
(https://www.nasdaq.com/articles/forex-market-overview-2019-06-07).
3
Kaniel et al. (2012) show that individual investors trading activity before earnings announcements positively predict
stock returns.
4
https://www.bloomberg.com/opinion/articles/2021-01-30/gamestop-gme-short-squeeze-who-will-surrender-first
5
Our proprietary dataset consists of intraday retail investors FX long and short positions built by aggressive trading, that
is, usage of market orders. We use this information to capture order flow at different intervals (one-minute and five-
minutes intervals) by looking at differences between net purchases and net sales over the relevant periods.

3
sentiment in the news that includes all unscheduled news with or without fundamental information.
Third, does intraday retail investor order flow predict future FX returns; or, asked differently, can
simple trading strategies exploit this order flow to provide profitable opportunities? Answering these
questions is useful for understanding the motives and mechanisms behind retail investors’ trading in
FX markets and can provide insights into how they react to fundamental information as well as to
news sentiment. Collectively the answers to these questions will shed light on the role of retail
investors in the functioning of FX markets.
The motivation to investigate the order flow of retail investors and their role in the price discovery
process around macro announcements stems from the findings of Evans and Lyons (2008), and Love
and Payne (2008). Evans and Lyons show that interdealer order flow variations contribute more to
currency price dynamics following the arrival of public macro news than in other times, and that
roughly two-thirds of the total effect of macro news on the Deutsche Mark/USD exchange rate is
transmitted via order flow. Only one third of the total effect of macro news is impounded into prices
instantaneously without any trading. Love and Payne (2008) also raise the question of why order
flow is so important in the formation of currency prices around releases of public information. Their
examination of interdealer order flow does not provide conclusive support for the idea that
differential interpretations of public information by market participants is the explanation. Our study,
by focusing on the order flow of the end-user segment of retail investors in currency markets,
provides important new evidence of the heterogenous reaction by market agents to public signals.
We use a proprietary intraday dataset that allows us to capture a high-frequency order flow measure
(both at one minute and five minute intervals) of a large pool of retail investors in EUR/USD for the
period July 2014 to April 2016 from an anonymous European regulated financial services firm. For
fundamental news, we focus on the surprise component of scheduled macro news announcements

4
in both the US and the Eurozone. To capture the broader news sentiment in all news, scheduled and
unscheduled, fundamental and non-fundamental, we use an intraday news sentiment measure
provided by Thomson Reuters Marketpsych Indices (TRMI). Standard event study analysis shows no
significant adjustment in retail investor behavior ahead of scheduled macro news announcements
and indicates a contrarian trading behavior in the first ten minutes after large announcement
surprises. A more comprehensive dynamic panel regression analysis that addresses
heteroscedasticity and autocorrelation concerns and controls for various fixed effects like day of the
week, time of the day, and event origin, shows that the contrarian trading behavior is mostly driven
by lagged returns rather than fundamental macro news. Further, a long-run time series analysis
confirms the intraday retail investor trading contrary to recent returns.It also reveals that retail
investor order flow is positively associated with the lagged overall news sentiment. The lack of trading
activity before macro-news, the contrarian activity after, and the attraction to news sentiment, all
suggest that retail traders in the FX market are uninformed. Indeed, we show that simple cross-over
trading strategies that exploit retail investors’ order flow could be profitable. Overall, our evidence
suggests that the lack of ability to correctly extract fundamental information from public news, the
fast reaction to news sentiment, and the return contrarian trading behavior, trigger retail investor
trading against informed traders.
To assess the retail investor behavior around macroeconomic announcements we utilise our intraday
dataset that allows the estimation of five-minute retail investor order flow, net long and net short
open interest, and overall unsigned trading volume for a relatively long period of twenty-two months
for the EUR/USD currency pair.
6
The use of intraday retail investor order flow in our analysis is
6
Several papers examine the effects of Inter-dealer order flow on FX prices. See for example, Cai et al. (2001), Lyons
(2001), Evans and Lyons (2002a,b,c), Evans (2002), Rime (2003) , Andersen et al. (2003), Osler (2005), Evans and Lyons
(2005, 2006) and Danielsson & Love (2006), Dominguez and Panthaki (2006), Evans & Lyons (2008), Berger et al. (2008),
Love and Payne (2008), Phylaktis and Chen (2010) , Rime et al. (2010). Flows from individual end-user segments in

5
necessary and important because typically only intraday interdealer order flow is studied in the
literature, such as Andersen, Bollerslev, Diebold and Vega (2003), who find price effects from
scheduled macro announcements. These price effects are difficult to detect at the daily frequency as
they are dominated by other factors.
7
We identify scheduled macroeconomic announcements for both the US and the Eurozone area
during our sample period and calculate the surprise component of the announcement—based on
analysts’ expected value before the announcement. Our sample contains 162 negative and 168
positive surprise events.
8
We use standard event study methodology as well as panel regression
analysis around these events. We find that retail investors do not exhibit any abnormal trading
behavior ahead of scheduled macro announcements. The also find strong evidence that retail
investors trade as contrarians after the announcement surprise. Further panel regression analysis
shows that this contrarian trading behavior is mostly driven by lagged returns rather than the news.
The contrarian trading behavior after the announcement, coupled with the lack of abnormal trading
before the announcement point to an uninformed status for individual investors in currency markets,
suggesting that they lack the skills to extract and interpret the fundamental information from macro-
news announcements. Our results differ from the findings of Kaniel et al. (2012) who explain the
contrarian trading of equity retail investors after corporate earnings announcements, as closures of
profitable positions they enter into ahead of the announcement indicating informed trading by
individual investors.
currency markets are addressed in Froot and Ramadorai (2005), Evans and Lyons (2012) and Menkhoff et al. (2016), with
only the latter examining individual investors as a separate segment at a daily frequency.
7
See also for example, Engel and Frankel (1984), Ito and Roley (1987), Ederington and Lee (1995), Almeida et al. (1998)
and Rosa (2013).
8
We delete from our sample both macroeconomic announcements that do not constitute surprises either because they
match analysts expectations or because we do not have any analysts expectations for those announcements.

6
We proceed to examine whether the broader news sentiment affects the trading behavior of retail
FX investors. The literature examining the effects of news on currency prices not only considers the
impact on prices of scheduled macro news announcements, but extends its investigation to the
impact of unscheduled fundamental and non-fundamental news. For example, Evans and Lyons
(2008) shows the impact on FX prices of both scheduled and unscheduled news appearing on the
Reuters trading screen and Domininguez and Panthaki (2006) find that unscheduled news not related
to fundamentals significantly influence intra-day FX returns and volatility.
9
To capture the sentiment
in all news scheduled and unscheduled, fundamental and non-fundamental we utilize a novel intra-
day news sentiment index provided by Thomson Reuters Marketpsych Indices.
10
The indicators are
updated every minute and they represent the 24-hour rolling average sentiment in news and social
media references to a broad spectrum of economy related topics. We calculate a 5-minute change in
relative news sentiment between the EU and the US to test the impact of news arrivals on FX retail
investors order flow. Using time series analysis, we confirm the return contrarian trading behavior
of retail investors, but we also find that the change in relative news sentiment has a positive
statistically significant effect on their order-flow. These results complement the findings of our
previous analysis and suggest that while retail investors do not trade in the direction of the
fundamental news after scheduled macroeconomic announcements (calculated surprise in the
news), they do trade in the direction of the overall sentiment in the broader public news.
Finally, the empirical evidence from simple cross-over trading strategies conditional on retail investor
order flow further verify the uninformed nature of retail trading. More specifically, the trading
9
See also Edrington and Lee (1996), Almeida et al. (1998) and Bauwens et al. (2005).
10
Sun, Najand, and Shen (2016) use TRMI intraday sentiment index data and find substantial evidence that intraday S&P
500 index returns are predictable using lagged half-hour sentiment, that their results are not driven by macroeconomic
news announcement effects and that the TRMI sentiment measure is significantly correlated with alternative sentiment
measures such as the Michigan consumer sentiment index.

7
strategy generates a sell signal when the short term retail order flow moving average crosses above
the long term moving average and a buy signal when the short term retail order flow moving average
crosses below the long term moving average. The mean and median strategy returns for holding
periods ranging from four to twenty hours, are positive, economically, and statistically significant for
both, in-sample and out-of-sample analysis. The intraday predictability of FX returns on the opposite
site of retail traders’ order flow shows not only the uninformed nature of retail traders, but also
highlight the effect of retail investors in the intraday price discovery process. The profitable
opportunities that emerge due to their contrarian trading suggests that their behavior contributes to
the slowdown of the price discovery process. Recent work by Luo, Ravina, Sammon and Viceira (2023)
on retail investors contrarian trading in equities market provide such evidence by showing that this
behavior leads to sluggish price adjustment.
Our work relates to the broader FX literature that examines the determinants of FX returns and how
information is incorporated into currency prices. In traditional textbook models of currency pricing,
public news is common knowledge and is reflected into prices directly leaving no role for trading
around the arrival of public news to cause price changes. In microeconomic asset pricing models,
trading has a causal effect on asset prices (see for example Glosten and Milgrom, 1985 and Kyle,
1985) because transactions convey information that is not common knowledge. The role of order
flow and the effect of scheduled macro news announcements in determining currency prices is
investigated extensively in the foreign exchange literature, typically using intraday interdealer data (
see footnote 4). Evans and Lyons (2008) show that while order flow contributes significantly to
changing currency prices at all times, it contributes more immediately after news announcements.
They argue that macro news trigger trading that reveals dispersed information, which in turn affects
currency prices. This highlights the importance of focusing on trading by individual end-user

8
segments around news events to identify differential interpretation of public information by currency
traders.
Our paper contributes to this literature in two important ways. First, we provide the first examination
of the FM market retail investors’ role in the trading-induced price discovery process around news
arrivals. Only a few studies examine individual end-user segments in currency markets (Froot and
Ramadorai, 2005; Evans and Lyons, 2012; Menkhoff et al., 2016) and even fewer focus on the intraday
trading activity of FX retail investors (Nolte and Nolte 2012, 2016).
11
Second, we conduct our 5-
minute order flow analysis around a broad set of scheduled macro news announcements, but we
also use a news analytics dataset
12
of the same intraday frequency that captures the full spectrum of
news. Our novel analysis of intraday retail investor order flow leads to some interesting conclusions.
We show that lack of skills to separate fundamental information from noise in public news along with
the strong tendency for return-contrarian trading leads retail investors to trade against informed
traders after the arrival of public news. More importantly, we show that retail investor intraday
trading has the potential of slowing down the adjustment of prices to fundamental news in the most
heavily traded currency pair, the EUR/USD. Hence, our results provide direct evidence that
differential interpretation of public news triggers trading that affects currency prices. This is
consistent with Love and Payne (2008) conjecture that the differential abilities of market participants
to interpret information could be an explanation for the importance of order flow in price formation.
The remainder of the paper proceeds as follows. Section 2 describes the dataset and provides
variable definitions and descriptive statistics. Section 3 investigates the retail investor trading
11
A small number of studies examine behavioral biases of retail investors in FX market (see Heimer, 2016; Ben-David et
al.,2018; Heimer and Imas, 2018).
12
We use the intraday US and Eurozone area TRMI country sentiment indices. Michaelides, Milidonis, Nishiotis and
Papakyriakou (2015) and Michaelides, Milidonis and Nishiotis (2019) use the daily TRMI country sentiment indices to
capture all public news about the country of reference that could affect equity and FX markets, respectively.

9
behavior around scheduled macroeconomic news announcements. It describes the event study and
panel regression methodologies we use and discusses the empirical results. Section 4 investigates
the effects of the overall news sentiment on retail investor order flow. It describes the intraday time
series analysis we use and discusses the empirical results. Section 5 describes simple cross-over
trading strategies conditional on retail investor trading and discusses the empirical results. Finally,
Section 6 summarizes the major conclusions.
2. Data and Descriptive Statistics.
2.1. Data
Our dataset consists of three levels of data. The first level contains aggregate retail customer minute
by minute aggressive (market orders) buy and sell open interest volume from the 10th July 2014
18:25 (EET) to 30th April 2016 23:59 (EET), for one of the major currency pairs, the EUR/USD. The
source of this proprietary dataset is an anonymous European Regulated Financial Services Firm that
provides online trading services to retail investors.
13
The company provided us with overall
demographic statistics, which indicate that on average their investors are younger, with lower
average income and net worth, but with comparable education level and gender split compared with
samples of equity retail investors used in prior literature (see for example Barber and Odean, 2008
and Graham et al., 2009).
14
The second level of our data contains scheduled macro news
announcements from both the US and the Eurozone area. All relevant information comes from
Datastream (Thomson Reuters). Finally, our third level of data comprises the innovative intraday
TRMI.
13
More than 1.5 million retail investors from more than 150 countries have used this financial services firm for their
active trading since its foundation. There are approximately five thousand active investors per day.
14
According to foreign exchange contact group of European Central Bank (ECB), the median age of retail investors in FX
market is 35, which is analogous to estimates in our sample
(https://www.ecb.europa.eu/paym/groups/pdf/fxcg/2301/Retail_FX.pdf?8b9766f1bbf56797757c4c2cb391f305).

10
The main analysis is conducted using 5-minute intervals starting from midnight every day amounting
to 288 intervals per day. The forex market is open 24 hours per day; therefore, we have 190,147 5-
minute observations in our sample period. While technically the forex market is open 24 hours a day,
7 days a week, the majority of dealers are choosing to close operations on weekends, leading to very
thin weekend liquidity. Normally, dealers providing trading services to retail investors fall into this
category. We therefore remove weekends from our data, which leaves us with 135,715
observations.
15
2.1.1. Buy and Sell Open Interest
Our proprietary dataset contains aggregate retail customer, minute-by-minute, long and short open
interest. These positions are the result of aggressive retail trading—usage of market orders.
16
The
positions are given in euro-money terms. For example, when an investor places a long (short) order
for EUR/USD, the size of the corresponding trade in euro money terms, is incorporated in the firm’s
database, as a long-initiated position (short initiated position) at the date-time that the position
opens. When a long (short) position is closed, the trade is recorded as a decrease in the long (short)
aggregate open interest.
2.1.2. Scheduled Macro News Announcements
The scheduled macro news announcements are from Datastream (Thomson Reuters). The Database
contains real time data for the macro announcements. Particularly, it includes for each country and
each classification, the event name, the exact date-time of the release, the measurement unit for
each event, as well as the actual, the previous, and the expected value of the announcement.
15
Our results are qualitatively the same after removing from our sample holidays like Christmas (24th – 26th of December),
New Year (31st of December – 2nd of January) and Easter (Friday before and Monday after Easter).
16
Our proprietary dataset provider did not provide retail positions based on passive retail trading—usage of limit orders—
but informed us that this activity is quite limited –less than 5%.

11
Expected value is given a few days before the announcement, and it is denoted as “Reuters Poll”. It
is calculated as the median of Reuters analyst forecast values. Reuters’ analysts are Economic
Research Houses, Credit Rating Agencies, Brokers, Banks and other specialist contributors around the
world. The expected value is not available for all announcements.
We filter the database for the news announcements from the US and the Eurozone area and use the
announcement categories used in earlier studies (see, for example, Andersen et al., 2003, Dominguez
and Panthaki, 2006, and Bernile et al., 2016). Table 1 lists all the news categories included in our
sample for the US and the Eurozone area, separately. The last column of Table 1, gives the sign of
the relationship of the corresponding announcement with the domestic currency. For example, a
positive sign for GDP growth indicates that a positive surprise in GDP growth is positive news for the
domestic currency and a negative sign for unemployment indicates that a surprise increase in
unemployment is negative news for the domestic currency. In general, events of which a positive
surprise signals higher (lower) economic growth are defined as positive (negative) events for the
domestic currency. We assign a positive sign for inflation because of the predominantly deflation
concerns during our sample period in both the US and the Eurozone area.
17
We classify scheduled
macro news announcements into positive and negative surprise events, in terms of their expected
effect on the movement of the EUR/USD exchange rate.
2.1.3. Thompson Reuters Marketpsych Indices (TRMI)
We use TRMI country sentiment indices to capture all public news relating to a country including
both scheduled and unscheduled news. The indices use a complex and sophisticated algorithm, which
overcomes the lexical ambiguity problem, by scoring the content of scanning text, based on its overall
diction and not on the diction of words and phrases in isolation. The TRMI country indices reflect
17
See Rosa (2013) and Love and Payne (2008).

12
scans of the content of over 2 million articles daily for references to 25 specific topics
18
and they
reflect the news content including positive and negative scores on specific topics and specific
emotions like optimism, fear, joy, trust, violence, conflict, urgency and uncertainty.
19
The TRMI
country sentiment index is normalized to take values between -1 and 1. The minute-by-minute
country sentiment index represents a simple average of the reported information in the past 24 hours
and is available for both the US and for the Eurozone as a whole.
2.2. Variable Definition and Measurement
In order to evaluate the behavior of retail investors around scheduled macro announcements and
examine the effects of lag returns and news sentiment on their trading activity, we employ three
main variables: Order Flow, Overall Unsigned Volume and Changes in Sentiment. Their detailed
description is given below.
2.2.1. Order Flow
We first calculate the Net Long and Net Short positions per five minutes as the change in the long
and short open interest per 5-minute interval, respectively
20
. That is,
Net Longt = Longt – Longt-1 (1)
Net Shortt = Shortt – Shortt-1 (2)
18
These topics are Market Risk, Budget Deficit, Central Bank, Consumer Sentiment, Credit Conditions, Default, Economic
Conflict, Economic Growth, Economic Uncertainty, Election Sentiment, Financial System Instability, Fiscal Policy,
Government Anger, Government Corruption, Government Instability, Inflation, Interest Rate, Investment Flows,
Monetary Policy, Natural Disasters, Regime Change, Social Inequality, Social Unrest, Trade Balance, and Unemployment.
19
The psychology literature provides evidence of the effects of emotions on economic transactions (Lerner, Small and
Loewenstein, 2004, Lerner and Keltner, 2001).
20
Long and short open interest is the result of aggressive retail trading (usage of market orders).

13
The Net Longt and Net Shortt variables represent net buying (new openings of long positions net of
closures of existing long positions) and net selling activity (new openings of short positions net of
closures of short positions) in the corresponding 5-minute interval, respectively.
Order Flowt is a signed trading volume and is calculated as the difference between the Net Long and
Net Short initiated positions, in terms of volume of the base currency (Euro).
Order Flowt = Net Longt – Net Shortt (3)
2.2.2. Overall Unsigned Trading Volume
We define Overall Unsigned Volume as the sum of the absolute value of Net Long and the absolute
value of Net Short positions.
Overall Unsigned Volumet =|Net Longt| + |Net Shortt| (4)
We use Overall Unsigned Volume as a proxy of the total trading volume. However, it is important to
note that it understates the total trading (turnover) because it only reflects the net change in long
and short interest. It does not capture new long (short) positions that are cancelled out by closures
of existing long (short) positions.
2.2.3. Sentiment in public news
In order to capture the arrival of public news we calculate a rolling 5-minute change in relative news
sentiment between the Eurozone (EZC) and the US. We describe the process for this calculation
below. We first calculate the change of each TRMI country index as follows:
ΔTRMI_EZC t = TRMI_EZC t - TRMI_EZC t-1 (5)
ΔTRMI_US t = TRMI_US t - TRMI_US t-1 (6)

14
We use the 5-minute TRMI sentiment indices which represent the rolling average news sentiment
over the past 24-hours, so the 5- minute ΔTRMI variables above capture the rolling sentiment in news
arrivals in the most recent 5-minute period relative to the 5-minute period 24 hours ago.
21
The news
sentiment in both the US and Eurozone could affect the EUR/USD exchange rate so we proceed to
calculate the relative news sentiment index between the Eurozone and the US by taking the
difference between the two-ΔTRMI variables as follows
22
:
RelΔTRMIt = ΔTRMI_EZC t - ΔTRMI_US t (7)
where RelΔTRMIt is the change in the Eurozone sentiment index relative to the change in the US index
at time t. For example, a positive value for the 5-minute RelΔTRMI
23
indicates the arrival of net
positive news sentiment for the euro or net negative news sentiment for the dollar at time t.
2.3. Descriptive Statistics
Table 2 reports descriptive statistics for the Long and Short initiated positions, the Net Long and Net
Short initiated positions, Order Flow, the Overall Unsigned Trading Volume, the rolling 5-minute
RelΔTRMI, the rolling 30-minute RelΔTRMI and the EUR/USD Return. All variables are measured at a
5-minute frequency. On average, the volume of open long initiated positions is around €102 million,
ranging from €24 to €316 million, while the volume of open short initiated positions is around €125
million ranging from €26 to €363 million. The negative value of the mean Net Long volume and the
positive mean Net Short volume, indicate that on average, investors exhibit a euro selling pressure
within our sample period. The Order Flow also captures this tendency indicated by its negative mean
21
Using the changes of TRMI eliminates the high first order autocorrelation present in the levels time series.
22
Defining RelΔTRMIt as the ratio of the two variables instead of the difference does not alter our results.
23
We also estimate the rolling 30-minute RelΔTRMI. For the estimation of the 30-minute RelΔTRMI we use the
ΔTRMI_EZC t = TRMI_EZC t - TRMI_EZC t-6 and the ΔTRMI_US t = TRMI_US t - TRMI_US t-6 , where the 30-minute ΔTRMI
variables above capture the rolling news arrival in the most recent 30-minute period relative to the 30-minute period 24
hours ago.

15
value of €4,418. The mean of the Overall Unsigned Volume is close to €2 million, which provides an
indication of the average total trading per 5-minute interval in our sample period. The rolling 5-
minute (30-minute) RelΔTRMI takes values from -0.028 to 0.043 (-0.038 to 0.46) with zero mean and
median.
The 5-minute EUR/USD return ranges between -1.51% and 1.66%. Mean, median and skewness are
close to zero (skewness is -0.3325 and is not tabulated) pointing to a symmetric EUR/USD return
distribution.
2.4. Seasonality
Intraday FX data exhibit strong seasonal patterns
24
. Omission of seasonality adjustment can lead to
misleading statistical inference. Figure 1, shows the time of the day seasonal pattern while Figure 2,
shows the day of the week seasonal pattern for the sample period July 2014 to April 2016. To identify
the seasonal patterns, we use the average overall trading volume. Overall trading volume is defined
as the summation of the long and short open interest using 5-minute intraday observations. To
identify the time of the day seasonal pattern we average the overall trading volume for each 5-minute
interval of all days over our sample period while to identify the day of the week seasonal pattern we
average the overall trading volume for each day of the week.
25
We control for seasonality in our
empirical analysis.
3. Trading Behavior around Scheduled Macro Announcements
3.1 Event Study Analysis
We begin our analysis of retail investor trading behavior around scheduled macro news
announcements using an event study methodology. The main variables of interest are retail investor
24
See Melvin and Yin (2000).
25
The average long trading volume and the average short trading volume, exhibit similar patterns.

16
Order Flow and its individual components (Net Long volume and Net Short volume) as well the
Overall Unsigned Volume. We examine the abnormal behavior of these variables around positive
and negative surprise events, separately.
An abnormal value is defined as the actual value of the variable of interest minus the expected value
over the event window. For announcement i and date-time t the abnormal value is given by:
AVit = Vit - E(Vi) (8)
where, AVit, Vit, and E(Vi) are the abnormal, actual, and expected values respectively at time t.
As in Chae (2005), we use the constant mean method for estimating the expected value of the
variable under examination. The expected value equals to the mean value of the variable of interest,
over the estimation window. Therefore,
󰇛󰇜  


 (9)
where [t1,t2] is the estimation window and n is the number of observations included in that window.
The event window is defined as the two-hour window around the event
26
(i.e. twenty-five 5-minute
observations). The estimation window is defined as the period starting three hours before the event
up to one hour before the event. To get accurate and interpretable results, we drop from our sample
events that overlap either in the event or estimation window. The cumulative abnormal value (CAV)
over an interval [t1,t2] is calculated as the sum of all abnormal values over the interval given by
  

 (10)
26
“Event” is the exact date-time of the news announcement.

17
We apply a non-parametric statistical test (Wilcoxon signed-rank test), to test the null hypothesis
that the Median CAVi =0
Table 3, reports the event study results for the abnormal cumulative Order Flow, the abnormal
cumulative Net Long and Net Short, and the abnormal Overall Unsigned Volume. Panels A and B,
report the results for negative and positive surprise scheduled macroeconomic announcements,
respectively. The table reports results for the event window [0, 5], short ten minute windows before
and after the event ([-10, -5] and [+10, +15]) and longer pre and post event windows ([-60, -15] and
[+20, +60]).
The results for the pre-event windows of the order flow and its components reveal no directional
abnormal investor trading ahead of the scheduled news announcement. More specifically, there is
no significant abnormal reaction for Order Flow, Net Long and Net Short in the pre-event windows
for both negative and positive surprise events. We do however find some evidence of a drop in overall
trading in the 10-minute window before the event only for positive surprise events. This is indicated
by the negative and statistically significant (10% level) cumulative abnormal overall trading variable
for the [-10, -5] window in Panel B.
Overall, the pre-event results indicate that retail investors in FX markets do not seem to adjust their
trading behavior ahead of the scheduled macro news announcements. Chae (2005) argues that
because of information asymmetry, discretionary liquidity traders decrease their cumulative trading
volume ahead of scheduled corporate earnings announcements. The information environment
ahead of scheduled macro announcements is likely to be very different in FX markets. Ito, Lyons and

18
Melvin (1998) argue that in FX markets there is no plausible analog to an insight information
argument, which is common in equity markets.
27
Tuning to the event and post event results, we find a statistically significant increase in cumulative
abnormal overall trading in all event and post event windows for both negative and positive surprise
events. The results point to increased overall trading activity in the 60-minute period immediately
following the scheduled macro announcements, which raises the question in which direction do
investors trade in this period? To answer this question we turn to the results for oOrder Flow and
the Net Long and Net Short variables.
First, we turn to the findings for the event window and observe that the results differ across negative
and positive surprise events. For the negative surprise events, we do not observe any significant
order flow reaction at the event, consistent with the finding of no-reaction in the net long and net
short variables. However, for the positive surprise events we find a significant positive reaction in
the cumulative abnormal order flow [0, +5] window, which is associated with a significant negative
cumulative abnormal Net Short variable and an insignificant cumulative abnormal Net Long variable
in the corresponding window. Therefore, the significant increase in abnormal order flow at positive
surprise announcements is driven by significant closures in losing short positions and not by
increased investor buying. The closures of losing positions at the announcement are likely involuntary
due to high leverage and margin calls.
28
Second, we turn to the short 10-minute period after the event, where for both positive and negative
surprise events we observe a significant abnormal reaction of retail investor order flow in the
27
Lucca and Moench (2015) and Bernile et al. (2016) identify informed trading in equity markets ahead of FOMC
announcements, but not ahead of other types of macro news announcements.
28
Leverage levels upwards of 800:1 are common during our sample period and the maintenance margin is 20%. For the
effects of leverage in the US retail foreign exchange market, see Heimer and Simsek (2019).

19
opposite direction of the surprise in the news announcement. For the negative surprise events, the
cumulative abnormal order flow in the window [+10, +15] is positive and statistically significant at
the 5% level, driven by a positive and highly significant (1% level) increase in the cumulative abnormal
net long variable. The corresponding abnormal net short variable is insignificant. For the positive
surprise events, the cumulative abnormal order flow in the window [+10, +15] is negative and
statistically significant at the 5% level, driven by a positive and highly significant (1% level) increase
in the cumulative abnormal net short variable. The corresponding abnormal net long variable is
insignificant. These results indicate that retail investors in FX markets exhibit a contrarian behavior
over the surprise of the announcement. Finally, for the longer post event window [+20, +60] we
observe a positive and statistically significant (10% level) reaction for the negative surprise events
and an insignificant reaction for the positive surprise events.
Overall, our event study analysis reveals that retail investors do not seem to adjust their trading
behavior ahead of the scheduled macro news announcements indicating that they do not seem to
be concerned that they could be trading against informed investors. At the event, we observe some
evidence of an increase in overall trading consistent with closures of losing positions. The overall
trading increases significantly after the event, which indicates that the arrival of new information
and/or the ensuing price adjustment triggers retail investor trading. These investors trade in the
opposite direction to the surprise in the news during the short window after the announcement. In
the next section, we test the robustness of our event study findings to a more rigorous panel
regression analysis.
3.2. Panel Regressions Analysis
In this section, we go beyond the univariate event study to conduct a dynamic panel regression
analysis in order to address heteroscedasticity and autocorrelation concerns and to control for

20
various fixed effects like day of the week, time of the week, and event origin.
29
We estimate panel
regressions using 5-minute frequency data, starting three hours preceding each announcement and
ending one hour following each announcement. We create dummy variables for each of the following
pre, at, and post event windows [-60, -15], [-10, -5], [0, +5], [+10, +15] and [+20, +60], which take the
value of one in the respective window and zero otherwise.
We run four different panel data models using Order Flow, Net Long, Net Short, and the Overall
Trading volume as dependent variables. The independent variables are the dummy variables
capturing the relative time of interest as well as lags of the dependent variable. We determine the
optimal number of lags using the Schwartz and Akaike information criteria. We also control for event
origin and for seasonality (day of the week and hour of the day). Finally, we estimate robust (White)
standard errors clustered by event.
30
Table 4, in the columns titled ‘Model 1’, reports the results of the aforementioned panel regression
analysis for the order flow, its components net long and net short, and the overall trading volume.
Panels A and B of Table 4, report the results for the negative and positive surprise event samples,
respectively.
31
The variables of interest are the dummy variables capturing the event and pre, and
post event windows. The main results of the event study hold up to the more rigorous panel
regression analysis and the introduction of additional controls. The results in the pre-event period
29
Andersen et al. (2003), show that the surprise component of US announcements, has a greater impact on the US$/CN$
exchange rate than that of Canadian announcements.
30
We also conduct robustness checks using longer estimation windows; 4 hours before the event to 1 hour before the
event, 8 hours before the event to 1 hour before the event and 12 hours before the event to 1 hour before the event.
The bigger the estimation window, the less events are included in our analysis since we are excluding overlapping events
within the event and estimation periods. We also winsorize our order flow measure at the 5% level. Further, we use a
standardization method to estimate the standardize measure of abnormal value of the variable of interest, by subtracting
the rolling 1-day-mean value of the corresponding variable and dividing by its 1-day-standard deviation,󰇛

󰇜. Instead
of the rolling 1-day-mean and the 1-day-standard deviation, we also use the rolling 1-week-mean and 1-week-standard
deviation respectively. The results remain robust.
31
Results are qualitatively the same if positive and negative surprise events are included in the same model with a dummy
variable separating the positive surprise events.

21
indicate that retail investors do not appear to react to the upcoming scheduled macro news
announcements. All the pre-event windows for all dependent variables are statistically insignificant.
At the event (window [0, +5]) we observe a significant closure of losing long positions in the negative
surprise sample and losing short positions in the positive surprise sample. The order flow at the event
is positive and significant in the positive surprise sample. In the negative surprise sample however, it
is statistically insignificant because in addition to the closures of losing long positions we also observe
significant closures of profitable short positions evening out the effect on order flow. The overall
trading volume at the event window is positive and statistically significant in both samples, but it is
statistically insignificant in the post event windows.
The results in the short post event window for the directional order flow variables remain the same
as the event study results.
32
We observe a significant abnormal reaction of retail investor order flow
in the opposite direction of the surprise in the news announcement. For the negative surprise events,
the order flow in the window [+10, +15] is positive and statistically significant at the 5% level. This
result is driven by a positive and significant (5% level) increase in the net long variable indicating that
retail investors intensify their buying activity (opening new long positions) after bad news. The net
short variable is not statistically significant indicating no significant selling activity or significant
closure of short positions after bad news. For the positive surprise events, the order flow in the
window [+10, +15] is negative and statistically significant at the 1% level. This result is driven by a
positive and significant (1% level) increase in the net short variable indicating that retail investors
intensify their selling activity (opening new short positions) after good news. The net long variable is
32
Existing literature examining the linkage between exchange rate returns and macroeconomic news at an intraday basis
find that the main impact occurs within 20 minutes (Ederington and Lee (1996), Andersen and Bollerslev (1998), Almeida
et al. (1998), Andersen et al. (2003), Dominiquez and Panthaki (2006)).

22
not statistically significant indicating no significant buying activity or significant closure of long
positions after good news.
These results indicate that shortly after the event retail investors in FX markets exhibit a contrarian
behavior relative to the sign of the surprise of the scheduled macro news announcement. They open
new long positions after bad news and new short positions after good news. This contrarian trading
behavior after the announcement, coupled with the lack of abnormal trading before the
announcement point to an uninformed status for individual investors in currency markets even after
the release of the announcement, suggesting that they lack the skills to extract and interpret the
fundamental information from macro-news announcements. Our results differ from the findings of
Kaniel et al. (2012) who explain the contrarian trading of retail investors in the equities market after
corporate earnings announcements as closures of profitable positions they enter into ahead of the
announcement suggesting informed trading by individual investors.
Kaniel et al. (2012) using daily data, further find that this contrarian trading behavior by equity market
retail investors is explained by both past returns and the sign of the earnings surprise. Menkhoff et
al. (2016), also using daily data, find a return contrarian behavior by retail investors in currency
markets. To shed more light on what is driving retail contrarian trading in the FX market, and
following Kaniel et al. (2012), we proceed to test whether this behavior documented in our event
study and panel regression analyses is news-driven or return driven.
To address this question, we rerun the panel regression analysis including lag returns in our models
and report the results in the columns labeled ‘Model 2’ in Table 4. The results for both negative and
positive surprise samples indicate a strong intraday return-contrarian trading behavior by retail
investors. Positive (negative) past returns have a strong negative (positive) effect on order flow,
which is driven by significant closures (new openings) of long positions and a significant increase

23
(closure) of short positions.
33
Interestingly the event and pre event trading behavior results
documented in ‘Model 1’ columns remain the same for all directional trading variables and the overall
trading volume in both samples. The post event trading behavior results in the positive surprise
sample also persist after the inclusion of lag returns in the model indicating a news-contrarian
behavior after positive surprise events even after we incorporate lag returns in the analysis.
However, the coefficient estimates on the post event dummy variables for Order Flow and Net Long
variables in the negative surprise sample are no longer statistically significant, indicating a pure
return-contrarian trading behavior after negative surprise events.
34
4. News Sentiment and Trading Behavior: A Time Series Analysis.
In this section, we analyse how retail investors react to the overall sentiment in the news. To achieve
this we use the TRMI country sentiment index described in detail in the data section of the paper.
The TRMI country sentiment indices reflect references to specific topics and specific emotions
contained in news articles and social media and what they measure is potentially very different from
the pure fundamental information contained in the surprise component of scheduled
macroeconomic announcements used in our previous analysis. For the latter, it requires special skills
to interpret the news and derive the pure surprise component of the announcement (fundamental
information) that we use in our analysis; while in contrast, the TRMI country indices capture the
intensity and content of news articles and social media in the form that it reaches investors.
To test the impact of news sentiment and lag returns on retail investors’ behavior, we use a time
series analysis, where the dependent variable is the 5-minute Order Flow. As independent variables,
following predictive methodology, we include lagged values of the EUR/USD exchange rate return
33
A drop in the Net Long (Short) variable indicates more closures of previously opened long (short) positions than new
openings of long (short) positions.
34
Andersen et al. (2003) provide evidence of a stronger impact of bad news on FX prices relative to good news.

24
and the lagged value of news sentiment change. As a proxy for news sentiment, we use the TRMI
Eurozone and US country level sentiment indices and construct the relative 5-minute (30-minute)
Eurozone to US sentiment change variable, 5-minute RelΔTRMI (rolling 30-minute RelΔTRMI), as
described in section 2.2.3. We also control for the effects of lags of the dependent variable to avoid
omitted variable bias. We apply the Schwartz and Akaike information criteria, to choose the optimal
number of lags and we correct the coefficient variance/covariance matrix for autocorrelation and
heteroscedasticity using the Newey-West method. In all versions of the model, we control for
seasonality by including hour of the day and day of the week dummy variables. Finally, we run a
version of the model that includes dummy variables capturing various windows around positive and
negative surprise macroeconomic announcements to control for the effects of scheduled macro
news.
Table 5 reports the results. The variables of interest are the relative lag news sentiment and the lag
returns. Models 1 and 2 include the rolling 5-minute RelΔTRMI while models 3 and 4 include the
rolling 30-minute RelΔTRMI . Models 2 and 4 include controls for the scheduled macro news, while
models 1 and 3 do not. All models include the first lag EUR/USD return and the lag return for the
period [t-2, t-12]. In all models, the coefficients of lagged EUR/USD exchange rate returns are
negative and statistically significant at the 1% level indicating the strong causal effect on the retail
investors’ contrarian trading behavior consistent with our panel regression analysis.
35
Translating to
an economic effect, in model 1, for example, a one standard deviation increase in the first 5-minute
35
Results are qualitatively the same when we use the winsorized order flow measure at the 5 percent level or if we use
a standardization method to estimate the standardize measure of abnormal value of order flow. We estimate the
standardize measure of abnormal value of order flow by subtracting the rolling 3-hours-mean value of the corresponding
variable and dividing by its 3-hours-standard deviation,󰇛

󰇜. Instead of the rolling 3-hours-mean and the 3-hours-
standard deviation, we also use the rolling 1-day-mean (1-week-mean) and 1-day-standard deviation (1-week-standard
deviation) respectively.

25
lagged return, Rt-1, is associated with a decrease in the 5-minute order flow in the magnitude of
€667,011.6 (16.812* 0.0393=0.6670116 million).
We now turn to the lag change in news sentiment. We find a positive and statistically significant
coefficient estimate at the 5% level for the one lag relative news sentiment change (models 1 and 2),
while the statistical significance drops to the 10% level for the sixth lag relative change (models 3 and
4). These results indicate that retail FX investors pay attention to the news and respond fast in the
direction of the news sentiment change. In model 1 for example, a one standard deviation increase
in the one lag change in relative news sentiment (euro positive change) is associated with an increase
in Order Flow (Euro buying) by €19,014.00 (31.690*0.0006 = 0.019014 million) within a 5-minute
interval.
36
This result is consistent with the uninformed status of retail investors in FX markets implied by the
absence of abnormal trading before the announcement and the contrarian trading behavior after the
announcement shown in Section 3. Overall our findings suggest that even though retail investor are
attentive to the news and easily decipher the direction of the sentiment in the news content, they
collectively lack the interpretation skills necessary to extract the fundamental information contained
in the surprise component of scheduled news announcements. The following question then arises.
Does the trading behavior of retail investors, being return contrarians and driven by news sentiment,
collectively puts them at the losing site of trades. To answer this question we proceed to test whether
taking the opposite site of the aggregate retail investor order flow constitutes a profitable trading
strategy.
36
Results remain quantitatively similar after controlling for the level of sentiment.

26
5. Cross-Over Trading Strategies
Our previous analysis provides evidence that retail investors exhibit return contrarian trading
behavior, lack the skills to interpret the fundamental information in scheduled macro news
announcements, and rather follow the sentiment in the news. Given these findings, in this section
we investigate directly the uninformed nature of retail trading. We employ simple cross-over trading
strategies conditional on retail investor aggregate order flow. The aim of this analysis is to validate
the importance of retail investors trading activity and highlight their role in the price discovery
process in FX markets. Our objective is not to investigate the relative performance of various trading
rules and strategies and propose a trading strategy that maximize profits.
We deploy a simple cross-over trading strategy that generates buy and sell signals opposite to those
indicated by retail investor Order Flow. More specifically, we sell EUR/USD when the short-term
moving average of Order Flow crosses above the long-term moving average of Order Flow and buy
EUR/USD when the short-term moving average of Order Flow crosses below the long-term moving
average of Order Flow. In particular, each time we receive a buy (sell) signal, we take a long (short)
position in EUR/USD currency pair and we calculate the mean and median log return series statistics,
for non-overlapping signals, for holding periods from 4 hours up to 20 hours after the signal. We use
three different trading cross-over strategies, using 3-hour, 4-hour and 6-hour aggregation periods.
All the intraday aggregation periods are compared to the daily moving average, leading to the
following three cross-overs: 3-hour moving average vs daily moving average, 4-hour moving average
vs daily moving average and 6-hour moving average vs daily moving average. Each of the three
trading strategies is evaluated both in-sample and out-of-sample. In order to have sufficient number
of signals in both samples for all holding period windows, we split the sample into two equal sample

27
periods. Our results from the three different cross-over strategies are qualitative the same, so we
only report the results for cross-over strategy 3-hour moving average vs daily moving average.
Table 6, Panels A and B, report the in-sample and out-of-sample results, respectively. Each panel
presents the mean and median log returns separately for long and short positions taken after long
and short trading signals for 4, 8, 16 and 20 hours holding periods. The strategy provides a short
trading signal when the short-term (3 hours) Order Flow moving average crosses above the long-term
(daily) moving average. A long trading signal occurs when the short-term (3 hours) Order Flow
moving average crosses below the long-term (daily) moving average of Order Flow. The column
labeled ‘N’ reports the number of long and short trading signals received in each sample period.
The results reveal strong positive statistically and economically significant returns for long and short
positions across all holding periods for both the in-sample and out-of-sample tests. For example, the
20-hour holding period average (median) return out-of-sample is 0.348% (0.414%) for long positions
and 0.295% (0.295%) for short positions with 31 long signals and 36 short signals. Our results indicate
the presence of intraday predictability of FX returns in the opposite direction of retail investors’ order
flow.
37
The fact that we find this evidence in the EUR/USD currency market, by far the most heavily
traded currency pair,
38
further highlights the role retail investors can play in financial markets.
The FX return predictability is consistent with our findings on the trading behavior of retail investors
around the arrival of news, and considering all of our findings together, we provide significant insights
into the retail investor’s role in the price discovery process. The lack of ability to correctly extract
fundamental information from public news, the fast reaction to the sentiment of the news, and the
37
Since our strategy is merely to uncover the uninformed nature of retail trading we do include in our cross-over strategy
transaction costs.
38
According to BIS the USD/EUR is the most heavily traded currency pair representing 24% of all currency trading in April
2019 with the USD/JPY second at 13.2%.

28
return contrarian trading behavior trigger retail investor trading against informed traders. Their
trading behavior implicitly provides liquidity to informed traders that possess the skills to interpret
the fundamental information in the macro news. They do not seem to receive any benefits from this
service because they do not trade in the direction of the news before its release and their contrarian
trading after the news can delay the price adjustment to the news. The return predictability and the
evidence of lack of skill to interpret the fundamental information in the news by a significant segment
of traders, provides important evidence of the differential abilities of FX market participants to
interpret public information. It can, therefore, explain why trading—and order flow—after the public
signal affects prices (Love and Payne, 2008).
Our result of intraday return predictability by individual investors and their return contrarian
behavior square well with the findings of Menkhoff et al. (2016). Using daily data, they find that
individuals tend to be return “contrarians” and that long-short currency portfolios mimicking
individual investor order flow exhibit significant negative returns. Kaniel et al. (2008) and Kaniel et al.
(2012) also find contrarian-trading behavior by individual investors in equity markets, but different
from our findings they show that individual investors benefit from their implicit role as liquidity
providers to institutional investors.
6. Conclusions
We provide significant insights into the effects of news on the intraday trading behavior of retail
investors in FX markets. We examine their trading reaction to fundamental information contained in
scheduled macro-news announcements and analyze the effects of the overall intraday news
sentiment on their trading.

29
Overall our results suggest that retail investors in FX markets are attentive to the news and are
influenced by the overall news sentiment and past returns, but do not possess the skills to extract
and interpret the fundamental information contained in macro news announcements. More
specifically, standard event study analysis shows no significant adjustment in retail investor’s
behavior ahead of scheduled macro news announcements and indicates a contrarian trading
behavior to the announcement surprise after the first five minutes of the event. A more
comprehensive panel regression analysis shows that the contrarian trading behavior is mostly driven
by lagged returns rather than the surprise in the news announcement. Further time series analysis
confirms the intraday retail investor return contrarian trading behavior, but also shows that retail
investor order flow is positively associated with the 5-minute lagged overall news sentiment.
Finally, we find that simple cross-over trading strategies based on signals opposite to the direction of
retail investors’ order flow are profitable, highlighting the role of retail investors in the price discovery
process in currency markets. The lack of ability to extract fundamental information from public news,
the fast reaction to news sentiment, and the return contrarian trading behavior suggest that they
contribute to the slowdown of the price discovery process. Overall, our findings support the
differential abilities of market participants to interpret information as an explanation for the
importance of order flow in price formation around releases of public information in FX markets. `
References
Almeida, A., Goodhart, C. and Payne, R., 1998. The Effects of Macroeconomic News on High
Frequency Exchange Rate Behavior. Journal of Financial and Quantitative Analysis, 33, 383-408.
Andersen, T.G., Bollerslev, T., 1998. Deutsche mark-dollar volatility: intraday volatility patterns,
macroeconomic announcements and longer run dependencies. Journal of Finance 53, 219–265.

30
Andersen, T., Bollerslev, T., Diebold, F., Vega, C., 2003. Micro effects of macro announcements: real-
time price discovery in foreign exchange. American Economic Review 9, 38-62.
Andersen, T.G., Bollerslev, T., Diebold, F., Vega, C. 2007. Real-time price discovery in global stock,
bond and foreign exchange markets. Journal of International Economics 73, 251–277.
Azar, P.D. and Lo, A.W., 2016. The wisdom of twitter crowds: Predicting stock market reactions to
fomc meetings via twitter feeds. Journal of Portfolio Management, 42(5), pp.123-134.
Baker, M. and Stein, J.C., 2004. Market liquidity as a sentiment indicator. Journal of Financial Markets,
7(3), pp.271-299.
Baker, M. and Wurgler, J., 2006. Investor sentiment and the cross‐section of stock returns. Journal of
Finance, 61(4), pp.1645-1680.
Baker, M. and Wurgler, J., 2007. Investor sentiment in the stock market. Journal of Economic
Perspectives, 21(2), pp.129-151.
Baker, M., Wurgler, J. and Yuan, Y., 2012. Global, local, and contagious investor sentiment. Journal of
Financial Economics, 104(2), pp.272-287.
Barber, B.M. and Odean, T., 2008. All that glitters: The effect of attention and news on the buying
behavior of individual and institutional investors. Review of Financial Studies, 21(2), pp.785-818.
Barber, B.M., Odean, T. and Zhu, N., 2009. Do retail trades move markets? Review of Financial
Studies, 22(1), pp.151-186.
Bauwens, L., Ben Omrane, W., Giot, P., 2005. News announcements, market activity and volatility in
the euro/dollar foreign exchange market. Journal of International Money and Finance 24, 1108–1125.
Ben-David, I., Birru, J. and Prokopenya, V., 2018. Uninformative feedback and risk taking: Evidence
from retail forex trading. Review of Finance, 22(6), pp.2009-2036.
Ben Omrane, W. Hafner, C., 2015. Macroeconomic news surprises and volatility spillover in foreign
exchange market. Empirical Economics 48, 577-607.

31
Berger, D.W., Chaboud, A.P., Chernenko, S.V., Howorka, E. and Wright, J.H., 2008. Order flow and
exchange rate dynamics in electronic brokerage system data. Journal of international Economics,
75(1), pp.93-109.
Bernile, G., Hu, J. and Tang, Y., 2016. Can information be locked up? Informed trading ahead of macro-
news announcements. Journal of Financial Economics, 121(3), pp.496-520.
Borovkova, S. and Mahakena, D., 2015. News, volatility and jumps: the case of natural gas futures.
Quantitative Finance, 15(7), pp.1217-1242.
Brown, G.W. and Cliff, M.T., 2005. Investor sentiment and asset valuation. Journal of Business, 78(2),
pp.405-440.
Cai, J., Cheung, Y.L., Lee, R.S.K., Melvin, M., 2001. Once in a generation yen volatility in 1998:
fundamentals, intervention and order flow. Journal of International Money and Finance 20, 327–347.
Chae, J., 2005. Trading volume, information asymmetry, and timing information. Journal of Finance
60, 413-442.
Cheung, Y.W., Chinn, M.D. and Pascual, A.G., 2005. Empirical exchange rate models of the nineties:
Are any fit to survive?. Journal of international money and finance, 24(7), 1150-1175.
Danielsson, J. and Love, R., 2006. Feedback trading. International Journal of Finance & Economics,
11(1), 35-53.
DeGennaro, R.P., Shrieves R.E., 1997. Public information releases, private information arrival and
volatility in the foreign exchange market. Journal of Empirical Finance 4, 295–315.
DeLong, J. Bradford, Andrei Shleifer, Lawrence H. Summers, and Robert J. Waldmann, 1990, Noise
trader risk in financial markets, Journal of Political Economy 98, 703-738.
Dominguez K.M.E., F. Panthaki. 2006. What defines ‘news’ in foreign exchange markets, Journal of
International Money and Finance 25, 168-198.
Ederington, L.H., Lee, J.H., 1996. The creation and resolution of market uncertainty: the impact of
information releases on implied volatility. Journal of Financial and Quantitative Analysis 31, 513–539.

32
Ehrmann, M. and Fratzscher, M., 2005. Exchange rates and fundamentals: new evidence from real-
time data. Journal of International Money and Finance, 24(2), 317-341.
Engel, C. and Frankel, J., 1984. Why interest rates react to money announcements: An explanation
from the foreign exchange market. Journal of Monetary Economics, 13(1), 31-39.
Evans, M.D., 2002. FX Trading and Exchange Rate Dynamics. Journal of Finance 57, 2405–2447.
Evans, M.D., Lyons, R.K., 2002a. Order flow and exchange rate dynamics. Journal of Political Economy
110, 170-180.
Evans, M.D., Lyons, R.K., 2002b. Time-Varying Liquidity in Foreign Exchange. Journal of Monetary
Economics 49, 1025-1051.
Evans, M.D., Lyons, R.K., 2002c. Informational integration and FX trading. Journal of International
Money and Finance, 21(6), 807-831.
Evans, M.D. and Lyons, R.K., 2005. Do currency markets absorb news quickly?. Journal of
International Money and Finance, 24(2), pp.197-217.
Evans, M.D. and Lyons, R.K., 2006. Understanding order flow. International Journal of Finance &
Economics, 11(1), pp.3-23.
Evans, M.D., Lyons, R.K., 2008. How is macro news transmitted to exchange rates? Journal of
Financial Economics 88, 26–50.
Evans, M.D. and Lyons, R.K., 2012. Exchange rate fundamentals and order flow. Quarterly Journal of
Finance, 2.
Fama, E.F., 1970. Efficient capital markets: A review of theory and empirical work. Journal of Finance,
25(2), 383-417.
Flood, R.P. and Rose, A.K., 1995. Fixing exchange rates a virtual quest for fundamentals. Journal of
Monetary Economics, 36(1), .3-37.
Glosten, L.R. and Milgrom, P.R., 1985. Bid, ask and transaction prices in a specialist market with
heterogeneously informed traders. Journal of Financial Economics, 14(1), pp.71-100.

33
Goodhart, C.A., Hall, S.G., Henry, S.B. and Pesaran, B., 1993. News effects in a high‐frequency model
of the sterling‐dollar exchange rate. Journal of Applied Econometrics, 8(1), pp.1-13.
Hafez, P.A. and Xie, J., 2013. Sentiment Derived from News Predicts EUR/USD Movements. Available
at SSRN.
Heimer, R.Z., 2016. Peer pressure: Social interaction and the disposition effect. Review of Financial
Studies, 29(11), pp.3177-3209.
Heimer, R. and Imas, A., 2018, December. Biased By Choice: How Financial Constraints Can Reduce
Financial Mistakes. Review of Financial Studies, Forthcoming.
Heimer, R. and Simsek, A., 2019. Should retail investors’ leverage be limited?. Journal of Financial
Economics, 132(3), pp.1-21.
Hendershott, T., Livdan, D. and Schürhoff, N., 2015. Are institutions informed about news?. Journal
of Financial Economics, 117(2), pp.249-287.
Huang, D., Jiang, F., Tu, J. and Zhou, G., 2015. Investor sentiment aligned: A powerful predictor of
stock returns. Review of Financial Studies, 28(3), pp.791-837.
Ito, T. and Roley, V.V., 1987. News from the US and Japan: which moves the yen/dollar exchange
rate?. Journal of Monetary Economics, 19(2), 255-277.
Kaniel, R., Liu, S., Saar, G. and Titman, S., 2012. Individual investor trading and return patterns around
earnings announcements. Journal of Finance, 67(2), pp.639-680.
Kaniel, R., Saar, G. and Titman, S., 2008. Individual investor trading and stock returns. Journal of
Finance, 63(1), pp.273-310.
Karnaukh, N., 2016. The dollar ahead of FOMC target rate changes. Unpublished working paper.
Kelly, B. and Pruitt, S., 2013. Market expectations in the cross‐section of present values. Journal of
Finance, 68(5), pp.1721-1756.
Kim, O. and Verrecchia, R.E., 1994. Market liquidity and volume around earnings announcements.
Journal of Accounting and Economics, 17(1-2), pp.41-67.

34
King, M.R., Osler, C.L. and Rime, D., 2012. Foreign exchange market structure, players and evolution.
Marsh James, Sarno (Eds.), The Handbook of Exchange Rates, Wiley, pp. 3-44
Kumar, A. and Lee, C., 2006. Retail investor sentiment and return comovements. Journal of Finance,
61(5), pp.2451-2486.
Kuttner, K.N., 2001. Monetary policy surprises and interest rates: Evidence from the Fed funds
futures market. Journal of Monetary Economics, 47(3), pp.523-544.
Kyle, A.S., 1985. Continuous auctions and insider trading. Econometrica 53, 1315-1335.
Lechner, S. and Nolte, I., 2009. Customer Trading in the Foreign Exchange Market: Empirical Evidence
from an Internet Trading Platform. Available at SSRN.
Lee, C., Shleifer, A. and Thaler, R.H., 1991. Investor sentiment and the closed‐end fund puzzle. Journal
of Finance, 46(1), pp.75-109.
Lerner, J.S. and Keltner, D., 2001. Fear, Anger, and Risk. Journal of Personality and Social Psychology,
81, pp. 146-159.
Lerner, J.S., Small, D.A. and Loewenstein, G., 2004. Heart Strings and Purse Strings: Carry-over Effects
of Emotions on Economic Transactions. Psychological Science, 15, pp. 337-341.
Love, R. and Payne, R., 2008. Macroeconomic News, Order Flows, and Exchange Rates. Journal of
Financial and Quantitative Analysis 43, 467-488.
Lucca, D.O. and Moench, E., 2015. The Pre‐FOMC Announcement Drift. Journal of Finance, 70(1),
pp.329-371.
Luo, C., Ravina, E., Sammon, M. and Viceira, L. M., 2023. Retail Investors’ Contrarian Behavior Around
News, Attention, and the Momentum Effect. Working Paper. Available at SSRN:
https://ssrn.com/abstract=3544949.
Lyons R., 2001. The microstructure approach to exchange rates. MIT press.
Meese, R., Rogoff, K., 1983. Empirical exchange rate models of the seventies. Journal of International
Economics 14.

35
Melvin, M. and Yin, X., 2000. Public information arrival, exchange rate volatility, and quote frequency.
The Economic Journal, 110(465), pp.644-661.
Menkhoff, L., Sarno, L., Schmeling, M. and Schrimpf, A., 2016. Information flows in foreign exchange
markets: Dissecting customer currency trades. Journal of Finance, 71(2), 601-634.
Michaelides, A., Milidonis, A., Nishiotis, G.P. and Papakyriakou, P., 2015. The adverse effects of
systematic leakage ahead of official sovereign debt rating announcements. Journal of Financial
Economics, 116(3), pp.526-547.
Michaelides, A., Milidonis, A. and Nishiotis, G.P., 2019. Private information in currency
markets. Journal of Financial Economics, 131(3), pp.643-665.
Neal, R. and Wheatley, S.M., 1998. Do measures of investor sentiment predict returns?. Journal of
Financial and Quantitative Analysis, 33(4), pp.523-547.
Nolte, I. and Nolte S., 2012. How do individual investors trade? European Journal of Finance, 18(9-
10), pp. 921-947.
Nolte, I. and Nolte S., 2016. The information content of retail investors order flow. European Journal
of Finance, 22(2), pp. 80-104.
Osler, C.L., 2005. Stop-loss orders and price cascades in currency markets. Journal of International
Money and Finance, 24(2), pp.219-241.
Payne, R., 2003. Informed trade in spot foreign exchange markets: an empirical investigation. Journal
of International Economics 61, 307-329.
Peterson, R.L., 2016. Trading on sentiment: the power of minds over markets. John Wiley & Sons.
Phylaktis, K. and Chen, L., 2010. Asymmetric information, price discovery and macroeconomic
announcements in FX market: do top trading banks know more? International Journal of Finance &
Economics, 15(3), 228-246.
Plakandaras, V., Papadimitriou, T., Gogas, P. and Diamantaras, K., 2015. Market sentiment and
exchange rate directional forecasting. Algorithmic Finance 4, 69-79.

36
Rime, D., 2003. New electronic trading systems in foreign exchange markets. New Economy
Handbook, 469504.
Rime, D., Sarno, L. and Sojli, E., 2010. Exchange rate forecasting, order flow and macroeconomic
information. Journal of International Economics, 80(1), pp.72-88.
Rosa, C., 2013. Market efficiency broadcasted live: ECB code words and euro exchange rates. Journal
of Macroeconomics 38, 167–178.
Schmeling, M., 2007. Institutional and individual sentiment: Smart money and noise trader risk?.
International Journal of Forecasting, 23(1), pp.127-145.
Schmeling, M., 2009. Investor sentiment and stock returns: Some international evidence. Journal of
Empirical Finance, 16(3), pp.394-408.
Smales, L.A., 2014. News sentiment in the gold futures market. Journal of Banking & Finance, 49,
pp.275-286.
Sun, L., Najand, M. and Shen, J., 2016. Stock return predictability and investor sentiment: A high-
frequency perspective. Journal of Banking & Finance, 73, pp.147-164.
Tetlock, C.P., 2007. Giving Content to Investor Sentiment: The Role of Media in the Stock Market.
Journal of Finance 62, 1139-1168.
Tetlock, P.C., Saar‐Tsechansky, M. and Macskassy, S., 2008. More than words: Quantifying language
to measure firms' fundamentals. Journal of Finance, 63(3), pp.1437-1467.

37
Figure 1: Time of the day retail trading.
Figure 1 presents the average retail trading volume for each 5-minute time interval during the day for the period July
2014 to April 2016. Retail trading volume is defined as the summation of the retail long and retail short open interest
using 5-minute intraday observations.
205
210
215
220
225
230
235
240
245
00:00
00:55
01:50
02:45
03:40
04:35
05:30
06:25
07:20
08:15
09:10
10:05
11:00
11:55
12:50
13:45
14:40
15:35
16:30
17:25
18:20
19:15
20:10
21:05
22:00
22:55
23:50
Avg Open Interest (in Millions)
Time of the Day
Time of the Day Seasonal Pattern

38
Figure 2: Day of the week retail trading.
Figure 2 presents the average retail trading volume for each day during the week for the period July 2014 to April 2016.
Retail trading volume is defined as the summation of the long and short open interest using 5-minute intraday
observations.
215
220
225
230
235
240
Monday Tuesday Wednesday Thursday Friday
Avg Open Interest (in Millions)
Millions
Weekday Name
Day of the Week Seasonal Pattern

39
Table 1: Summary Statistics of scheduled macro announcements.
This table presents the Thomson Reuters scheduled macro news announcements, separately for the US and the Eurozone.
For each announcement, the table specifies the announcement’s name, the announcement’s reporting frequency, as well
as the classification and the sign of the announcement. Sign is defined as positive (+) if an increase on the corresponding
announcement surprise component, represents more economic growth in terms of domestic economy and negative (-)
if it represents less.
Country | Frequency
Announcements
Classifications
Sign
United States
Monthly
Capacity Utilization
Industry Sector
+
Construction Spending
Industry Sector
+
Consumer Confidence Index
Surveys & Cyclical
+
Consumer Credit
Government Sector
+
Consumer Price Index - CPI
Prices
+
Durable Goods
Industry Sector
+
Factory Orders
Industry Sector
+
Gross Domestic Product - GDP*
National Account
+
Government Budget Deficit
Government Sector
+
Housing Starts
Industry Sector
+
ISM Index
Surveys & Cyclical
+
Index of leading indicators
Surveys & Cyclical
+
Industrial Production
Industry Sector
+
New Home Sales
Industry Sector
+
Non - Farm Payrolls
Labour Market
+
Personal Income
National Account
+
Producer Price Index - PPI
Prices
+
Retail Sales
Consumer Sector
+
Trade Balance
External Sector
+
Unemployment Rate
Labour Market
-
Vehicle Sales
Consumer Sector
+
Whole Sales
Consumer Sector
-
Every Six Week
FOMC - Monetary Policy Meeting
Other
+
Weekly
Initial Unemployment Claims
Labour Market
-
Eurozone
Monthly
Consumer Confidence Index
Surveys & Cyclical
+
Euribor Futures
Other
+
Eurostat Trade
External Sector
+
Gross Domestic Product - GDP
National Account
+
Industrial Production
Industry Sector
+
M3 - Money Supply
Government Sector
+
PMI Index
Surveys & Cyclical
+
Producer Price Index - PPI
Prices
+
Retail Sales
Consumer Sector
+
Unemployment Rate
Labour Market
-
Every Six Week
ECB - Monetary Policy Meeting
Other
+
* There are three types of GDP announcements, namely, GDP advance, GDP preliminary and GDP final. Each type is
announced at a quarterly basis. However, the overall frequency of announcements is at a monthly basis since each
type of GDP is announced at a different month of a quarter.

40
Table 2: Summary Statistics of Long initiated positions, Short initiated positions, Net Long, Net Short, Order Flow, Overall
Unsigned Volume, 5 and 30-minute, EU vs US sentient change and EUR/USD Return.
We use Order Flow, Net Long, Net Short and Overall Unsigned Volume as proxies for retail investors overall behavior.
Order Flow is a signed trading volume and is calculated as the difference between the Net Long and Net Short positions,
where the Net Long and Net Short positions are the changes in the long and short initiated positions (long and short open
interest) per 5-minutes, respectively. For example, Net Long i,t = Long i,t – Long i,t-1. The sum of absolure Net Long and
the absolute Net Short positions provide a measure of overall trading volume/intensity and we denote it as Overall
Unsigned Volume. 5-minute RelΔTRMI (30-minute RelΔTRMI), refers to the 5-minute (30-minute) EU vs US TRMI
sentiment change (see detailed description on Section 2.2.3.). The exchange rate return at time t is measured as the
percentage of the difference in the log of exchange rate prices between time t and t-1 multiplied by 100.
Variable
Mean
Median
Minimum
Maximum
Std Dev
Long Initiated Positions
102,343,005
99,350,120
24,355,180
316,000,000
42,535,073
Short Initiated Positions
125,373,052
120,269,500
26,352,690
363,000,000
48,419,837
Net Long
-1,254
0.00
-98,000,000
58,000,000
1,995,204
Net Short
3,188
0.00
-69,000,000
49,000,000
2,178,731
Order Flow
-4,418
0.00
-113,000,000
102,000,000
2,988,562
Overall Unsigned Volume
1,855,923
1,000,000
0.00
145,000,000
3,039,106
5-minute RelΔTRMI
-0.00001
0.00001
-0.02816
0.04328
0.00057
30-minute RelΔTRMI
-0.00003
0.00006
-0.03838
0.04637
0.00198
Return
-0.00009
0.00000
-1.50889
1.65991
0.03930

41
Table 3: Event study results on the behavior of retail investors around macroeconomic news announcements.
This table presents event study results of how retail investors behave around macroeconomic news announcements. We
use Order Flow, Net Long, Net Short and Overall Unsigned Volume as proxies for retail investors overall behavior. Using
the announcements’ surprise component, we classified announcements into two main categories, positive and negative
surprise events (detailed description on Section 2.1.2.) therefore results are reported separately for each category. We
then measure cumulative abnormal values over different sub-periods of the event window [-60min, +60min], capturing
retail investors behavior in period before, during and after the announcement. Panel A (Panel B) reports, cumulative
abnormal values, for negative (positive) surprise events along with the level of their statistical significance (SS). ***, **
and * denote statistical significance at the 1%, 5%, 10% level, respectively.
Panel A: Event study results for Negative Surprise Events
Variables of
Interest
Order Flow
Net Long
Net Short
Overall Unsigned
Volume
Event Window
Median
SS
Median
SS
Median
SS
Median
SS
[-60, -15]
0.465
0.180
-0.434
0.359
[-10, -5]
-0.062
0.199
0.098
-0.309
[0, +5]
0.080
0.036
-0.140
0.963
***
[10, 15]
0.364
**
0.456
***
-0.047
0.480
***
[20, 60]
0.774
*
0.384
-0.649
**
1.587
*
Panel B: Event study results for Positive Surprise Events
Variables of
Interest
Order Flow
Net Long
Net Short
Overall Unsigned
Volume
Event Window
Median
SS
Median
SS
Median
SS
Median
SS
[-60, -15]
-0.713
-0.300
0.093
0.620
[-10, -5]
-0.160
0.113
0.204
-0.540
*
[0, +5]
0.494
**
0.400
-0.369
***
0.895
***
[10, 15]
-0.517
**
-0.029
0.575
***
0.414
**
[20, 60]
-0.917
-0.190
0.538
2.437
***

42
Table 4: Panel Regressions of individual investors’ behavior around macroeconomic announcements.
This table presents results from panel regression models where the dependent variables include the 5-minute Order
Flow, Net Long, Net Short and Overall Unsigned Volume (measured in Euro millions). For each macroeconomic
announcement for the period 10 July, 2014 to 30 April, 2016 we use the time series starting three hours before and
ending one hour after the event. We classify the announcements into positive and negative surprise events (see Section
2.1.2.) and report the results separately in Panels A and B. In Model 1, the independent variables include dummy variables
capturing the relative time of interest in minutes as well as lags of the corresponding dependent variable. The optimal
number of lags is determined by the Schwartz and Akaike information criteria. Model 2, also includes six lags of the 5-
minute return variable. ***, ** and * denote statistical significance at the 1%, 5%, 10% level, respectively. p-values are
reported in parenthesis.
Panel A: Negative Events (#162)
Order Flow
Net Long
Net Short
Overall Unsigned
Volume
Independent Variables
Model 1
Model 2
Model 1
Model 2
Model 1
Model 2
Model 1
Model 2
Relative time [-60 , -15]
0.121
0.085
0.021
-0.006
-0.114
-0.097
0.100
0.101
(0.299)
(0.440)
(0.776)
(0.936)
(0.128)
(0.167)
(0.352)
(0.349)
Relative time [-10, -5]
-0.105
-0.158
-0.137
-0.169
-0.031
-0.015
0.014
0.013
(0.576)
(0.360)
(0.250)
(0.149)
(0.828)
(0.908)
(0.933)
(0.936)
Relative time [0.+5]
0.003
-0.025
-0.676**
-0.677**
-0.704**
-0.677**
1.597***
1.600***
(0.992)
(0.931)
(0.041)
(0.044)
(0.027)
(0.035)
(0.003)
(0.003)
Relative time [+10,+15]
0.540**
0.184
0.369**
0.180
-0.201
-0.019
0.323
0.332
(0.048)
(0.485)
(0.020)
(0.235)
(0.302)
(0.919)
(0.152)
(0.135)
Relative time [+20,+60]
0.242**
0.150
0.047
-0.016
-0.237***
-0.184**
0.002
0.002
(0.047)
(0.146)
(0.527)
(0.825)
(0.009)
(0.018)
(0.980)
(0.982)
lag_return_1
-15.973***
-7.790***
8.334***
-0.566
(0.000)
(0.000)
(0.000)
(0.651)
lag_return_2
-7.974***
-4.830***
4.133***
1.914
(0.000)
(0.000)
(0.000)
(0.107)
lag_return_3
-6.209***
-4.179***
2.759***
-0.571
(0.000)
(0.000)
(0.000)
(0.599)
lag_return_4
-2.825**
-1.596**
1.179
0.619
(0.014)
(0.035)
(0.164)
(0.476)
lag_return_5
-2.023**
-0.876
1.578**
0.269
(0.040)
(0.251)
(0.013)
(0.767)
lag_return_6
-2.834
-2.639**
0.396
-1.001
(0.175)
(0.041)
(0.722)
(0.536)
Constant
-0.218
-0.255
-0.071
-0.085
0.175
0.189
0.928***
0.926***
(0.234)
(0.131)
(0.539)
(0.404)
(0.203)
(0.142)
(0.000)
(0.000)
Lags of Dependent
Variable
YES
YES
YES
YES
YES
YES
YES
YES
Time_of_the_Day
YES
YES
YES
YES
YES
YES
YES
YES
Day_of_the_Week
YES
YES
YES
YES
YES
YES
YES
YES
Observations
7,518
7,490
7,544
7,490
7,572
7,490
7,429
7,426
R-squared
0.0149
0.0804
0.0057
0.0403
0.0051
0.0354
0.1294
0.1297

43
Table 4 (continued)
Panlel B: Positive Events (#168)
Order Flow
Net Long
Net Short
Overall Unsigned
Volume
Independent Variables
Model 1
Model 2
Model 1
Model 2
Model 1
Model 2
Model 1
Model 2
Relative time [-60 , -15]
0.011
0.042
-0.047
-0.050
-0.078
-0.098
-0.020
-0.023
(0.922)
(0.688)
(0.573)
(0.516)
(0.320)
(0.195)
(0.792)
(0.771)
Relative time [-10, -5]
-0.033
-0.053
-0.154
-0.182
-0.129
-0.126
-0.201
-0.205
(0.869)
(0.768)
(0.232)
(0.134)
(0.377)
(0.374)
(0.158)
(0.147)
Relative time [0.+5]
0.818**
0.807**
-0.131
-0.155
-0.949***
-0.938***
1.429***
1.437***
(0.015)
(0.015)
(0.545)
(0.482)
(0.001)
(0.001)
(0.000)
(0.000)
Relative time [+10,+15]
-0.760***
-0.541**
-0.191
-0.068
0.587***
0.441***
0.280
0.279
(0.004)
(0.022)
(0.313)
(0.691)
(0.001)
(0.009)
(0.165)
(0.171)
Relative time [+20,+60]
-0.050
0.001
-0.055
-0.011
0.012
-0.013
0.077
0.068
(0.660)
(0.990)
(0.536)
(0.899)
(0.886)
(0.874)
(0.419)
(0.470)
lag_return_1
-19.630***
-9.925***
9.800***
1.204
(0.000)
(0.000)
(0.000)
(0.373)
lag_return_2
-10.614***
-5.388***
6.533***
-0.614
(0.000)
(0.000)
(0.000)
(0.485)
lag_return_3
-4.844***
-2.838**
3.662***
0.508
(0.001)
(0.022)
(0.003)
(0.598)
lag_return_4
-0.826
-2.121***
1.017
1.712*
(0.532)
(0.003)
(0.205)
(0.072)
lag_return_5
-3.111***
-2.478***
2.040**
1.786*
(0.007)
(0.000)
(0.038)
(0.069)
lag_return_6
0.784
0.285
-0.010
-0.438
(0.502)
(0.737)
(0.989)
(0.561)
Constant
-0.025
-0.034
0.057
0.068
0.111
0.118
1.051***
1.044***
(0.871)
(0.796)
(0.628)
(0.506)
(0.245)
(0.122)
(0.000)
(0.000)
Lags of Dependent Variable
YES
YES
YES
YES
YES
YES
YES
YES
Time_of_the_Day
YES
YES
YES
YES
YES
YES
YES
YES
Day_of_the_Week
YES
YES
YES
YES
YES
YES
YES
YES
Observations
7,660
7,641
7,776
7,654
7,641
7,641
7,525
7,525
R-squared
0.0373
0.1119
0.0014
0.0484
0.0173
0.0628
0.1633
0.1648

44
Table 5: Time Series analysis for the impact of sentiment on individual investors’ behavior.
This table presents time series regression results for the period 10 July 2014 to 30 April 2016. The dependent variable is
the 5-minute Order Flow. Models 1 & 2 (Models 3 & 4) include as independent variables the rolling 5-minute (30-minute)
lagged relative sentiment change between the EU and US (RelΔTRMI). All models include the 5-minute lagged EUR/USD
exchange rate return (Rt-1) and the lagged exchange rate return for the period t-2 to t-12 (Rt-2, t-12). Models 2 & 4 also
include dummy variables capturing several time intervals around macroeconomic news announcements. The table
reports the coefficient estimates and p-values (in parenthesis). ***, ** and * denote statistical significance at the 1%,
5%, 10% level, respectively.
Independent Variables
5-min Sentiment Change
30-min Sentiment Change
Model 1
Model 2
Model 3
Model 4
Lagged Sentiment Change (RelΔTRMI)
31.690**
32.918**
7.092*
7.435*
(0.038)
(0.031)
(0.081)
(0.067)
R t-1
-16.812***
-16.821***
-16.810***
-16.819***
(0.000)
(0.000)
(0.000)
(0.000)
R t-2, t-12
-2.383***
-2.379***
-2.383***
-2.379***
(0.000)
(0.000)
(0.000)
(0.000)
d_neg_relative time [-60 , -15]
0.003
0.002
(0.976)
(0.981)
d_neg_relative time [-10, -5]
-0.250
-0.249
(0.141)
(0.143)
d_neg_relative time [0.+5]
-0.038
-0.037
(0.896)
(0.899)
d_neg_relative time [+10,+15]
0.080
0.078
(0.743)
(0.747)
d_neg_relative time [+20,+60]
0.043
0.043
(0.618)
(0.619)
d_pos_relative time [-60 , -15]
-0.054
-0.054
(0.555)
(0.552)
d_pos_relative time [-10, -5]
-0.101
-0.100
(0.554)
(0.557)
d_pos_relative time [0.+5]
0.762**
0.762**
(0.019)
(0.019)
d_pos_relative time [+10,+15]
-0.652***
-0.652***
(0.005)
(0.005)
d_pos_relative time [+20,+60]
-0.083
-0.083
(0.459)
(0.460)
Constant
-0.017
-0.016
-0.017
-0.017
(0.503)
(0.519)
(0.488)
(0.502)
Lags of Dependent Variable
YES
YES
YES
YES
Hour_of_the_Day FE
YES
YES
YES
YES
Day_of_the_Week FE
YES
YES
YES
YES
Observations
127,251
127,109
127,251
127,109
Adjusted R-squared
0.065
0.065
0.065
0.065

45
Table 6: Mean and Median return of the in-sample and out-of-sample trading strategy.
This table presents mean and median returns from a simple cross over trading strategy that generates buy and sell signals opposite to that indicated by individual investors
Order Flow. It generates sell signals of EUR/USD when the short term (3 hours) moving average of Order Flow crosses above the long term (daily) moving average of Order
Flow and buy signals of EUR/USD when the short term (3 hours) moving average of Order Flow crosses below the long term (daily) moving average of Order Flow. We then
calculate the mean and median log returns on EUR/USD for holding period 1 to 20 hours for non-overlapping signals. Panel A (Panel B) reports mean and median results along
with the level of their statistical significance (SS), for the in-sample (out-of-sample) analysis, for 4, 8, 16 and 20 hours holding periods. ***, ** and * denote statistical significance
at the 1%, 5%, 10% level, respectively.
Panel A: Cross- Over Strategy - In-Sample
Panel B: Cross- Over Strategy - Out-of-Sample
Holding
period
Strategy
signal
N
Mean
SS
Median
SS
N
Mean
SS
Median
SS
4 hours
long
152
0.054
**
0.033
***
152
0.049
***
0.044
***
4 hours
short
156
0.033
*
0.018
156
0.065
***
0.044
***
8 hours
long
112
0.112
***
0.093
***
100
0.122
***
0.111
***
8 hours
short
115
0.099
***
0.060
***
114
0.095
***
0.084
***
12 hours
long
72
0.230
***
0.188
***
75
0.213
***
0.211
***
12 hours
short
85
0.100
**
0.069
**
77
0.192
***
0.156
***
16 hours
long
42
0.383
***
0.227
***
52
0.301
***
0.366
***
16 hours
short
64
0.177
***
0.169
***
59
0.292
***
0.248
***
20 hours
long
31
0.403
***
0.390
***
31
0.348
***
0.414
***
20 hours
short
47
0.240
***
0.108
***
36
0.295
***
0.295
***