Agent-Based Model in Directional-Change Intrinsic Time

Agent-Based Modelling in Directional-Change
Intrinsic Time
V. Petrov∗ †, A. Golub‡ and R. Olsen§
†University of Zurich, 15 Andreasstrasse, 8050 Zurich, Switzerland

‡flov technologies AG, 26 Gotthardstrasse, 6300 Zug, Switzerland
§Lykke Corp., 2 Baarerstrasse, 6300 Zug, Switzerland
We describe an agent-based model where trades happen in event-based time called directional-change
intrinsic time. Events are defined as the reversal price moves of a directional-change threshold from a
local extreme. The price impact of traded volumes is modelled according to the empirically observed
squared root impact function. The time series generated by the agents is characterised by statistical
properties typical for foreign exchange rates: low auto-correlation of returns, fat-tailed distribution of
returns, aggregated normality, and the price jump scaling law. Furthermore, we introduce and use as a
benchmark, the overshoot scaling law, which is an omnipresent feature of liquid markets and relates the
expected length of price overshoots to the length of the corresponding directional-change threshold.
Keywords: Agent-based model; Stylised facts; Forex; Directional-change; Intrinsic time; Scaling laws
1. Introduction
According to the Bank for International Settlements, the daily trading volume in Foreign Exchange
(FX) markets averaged $5.1 trillion per day in April 20161 . Such substantial volume can have a
noticeable impact on the stability of the overall financial system. This volume is generated by the
enormous number of transactions made by individual and institutional traders. Traders behaviour
is strictly connected to the flow of news related to the financial system (Ederington and Lee 1993,
Bauwens et al. 2005). Risk management tools capable of foreseeing and efficiently cope with the
market impact of any political, environmental or technical change have to be built considering the
characteristics and the origin of the market changes initiated by market participants. Due to the
large trading volumes and the multitude of trading agents, the FX market is one of the biggest
financial systems where agent-based models were extensively applied for its analysis (Aloud et al.
2017).
Market participants act on behalf of their perception and interpretation of the available informa-
tion. Their consolidated behaviour is fully reflected in the prices of the traded financial instrument.
This phenomenon is also known as the efficient-market hypothesis (Fama 1970). It encouraged
scientists to look into historical time series to study the attributes of the aggregated activity of
market participants involved in the trading. Multiple works have been done on the analysis of large
amount of historical data in the search for fundamental properties of various financial markets. For
example, Bollerslev and Melvin (1994) used more than 300 000 quotes for an empirical analysis of
∗ Corresponding author. Email: [email protected]

1 https://www.bis.org/publ/rpfx16fx.pdf
Electroniccopy
Electronic copy available
available at:
at: https://ssrn.com/abstract=3240456
https://ssrn.com/abstract=3240456
the bid-ask spread and its relation to the exchange rate uncertainty. Danielsson and de Vries (1997)
and Dacorogna et al. (2001) used high-frequency data to estimate the fat tails of exchange rate
returns. Kozhan and Salmon (2012) employed the dataset of submitted market and limit orders
to examine how the information contained in order books could be exploited in simple trading
schemes.
A wide range of agent-based models was proposed to explore the observed changes in the markets
dynamic. The changes are assumed to be the markets reaction to the actions of individual and
aggregated groups of traders inhabiting the market. The models attempt to replicate the evolving
behaviour of real market participants by creating artificial agents who impact the market in the
same way as it happens in reality. Most of the models represent complex systems, populated by
a large number of independent and heterogeneous actors competing with each other (Naciri and
Tkiouat 2016). Agent-based models are aimed to reproduce and explain the phenomena of real
markets, such as bubbles, crashes, and regime switches (Samanidou et al. 2007, Ehrentreich 2007).
There is a wide range of various designs of agent-based models characterised by specific proper-
ties intended to imitate phenomenon observed in reality. Despite the diversity of the models, one
structural element is always present in their description: the definition of time and how the agents
dynamically adapt themselves to its flow. Scientists used to rely on physical time where hours, days
or even seasons are selected to measure elapsed periods and intervals between various events (see
LeBaron (2006) for the surveys on agent-based models used in finance). Physical time is tradition-
ally employed to describe the interaction between agents and the impact of their trading activity on
the market prices. Equally spaced timestamps are assumed to be the points where the actions have
to be considered. However, the real market is a complex system with its endogenous non-constant
time stream. The activity of that stream is dependent on the inhomogeneous frequency of political,
social or environmental events (Guillaume et al. 1997). The non-constant volatility typical for the
FX market is the main aspect which helps to understand the disadvantages of physical time in
agent-based modelling. The volatility of high-frequency markets is the proxy for the market par-
ticipants activity (Schwert 1989). At the same time, samples of equidistant time intervals of given
lengths used to compute volatility can significantly affect its results (Müller et al. 1997). Therefore,
financial instruments and agent-based models built upon physical time and designed to measure
or replicate statistical properties of real time series are naturally limited by the stiffness of selected
equidistant intervals.
The limitation of equidistant time periods applied to the financial data analysis can be overcome
by applying the concept of directional-change intrinsic time proposed by Guillaume et al. (1997). In
this intrinsic time representation, events are endogenously defined as reversal price moves measured
from local price extremes. All reversals have to be of the size equal to the given directional-change
threshold. The proposed measure dissects the real price curve into the sequence of alternating
ascending and descending trends. Each elementary trend ends once a new price curve reversal is
observed. Continuous price moves in the same direction of the latest reversal are called overshoots.
Overshoots represent trend components of the price curve. Only price moves indicating the begin-
ning of the new trend allow the systems clock tick. Thus, this intrinsic time is capable of dealing
with the essential properties of price curves such as trends independently of the time intervals
chosen to observe them. In other words, intrinsic time is agnostic of the scale and speed of the
price evolution registered in physical time.
Trend changes are the primary indicators of the prevailed side of the aggregated behaviour of
all market participants (Muth 1961, Cohen et al. 2007). The agent-based model, proposed in our
work, employs the directional-change intrinsic time framework to monitor trends and dissect the
price curve into a collection of directional-change points. These points and price overshoots are
used by the agents as signals for their trading activity. Agents, operating in directional-change
intrinsic time, judge the market conditions and make decisions on their actions employing prices
as the only source of information.
A successful agent-based model should be able to reproduce statistical properties of real financial
markets. These properties are mostly known as ‘stylised facts’ and are copiously discussed in
Electroniccopy
available at:
numerous papers (Kaldor 1961, Pagan 1996, Kullmann et al. 1999, Gençay et al. 2001, Cont
2001, Chakraborti et al. 2009). In this work, we test the time series synthetically created by
the model against a set of benchmarks. The main target of the test is to confirm that the created
time series are characterised by the set of selected statistical properties. The following stylised
facts are among the chosen tests: low autocorrelation of returns, fat-tailed distribution of returns,
aggregational normality, the price jumps scaling law1 and the overshoot scaling law. The latter
statistical property was chosen in addition to the four stylised facts usually adopted as benchmarks
of agent-based models. This scaling law was recently found in a wide range of real high-frequency
time series from the FX market and even in arithmetic Brownian motion (Glattfelder et al. 2011).
It establishes a relation between the average length of overshoot and the corresponding size of the
directional-change threshold. The relation states that on average, a directional change is followed by
an overshoot of the same magnitude. The sovereignty of the directional-change concept of the flow
of physical time makes the overshoot scaling law a very convenient tool for testing the performance
of agent-based models. To the extent of our knowledge, it is the first research work where a stylised
fact based on the directional-change intrinsic time is used to evaluate the time series generated by
a group of heterogeneous agents.
The outline of the remaining paper is as follows. Section 2 illustrates the details of the directional-
change intrinsic time framework. It also provides an example of a real price curve dissected into a
collection of intrinsic events. Section 3 describes two main components of the agent-based model:
the set of artificial agents and the market impact function. In Section 4 the collection of benchmarks
used to validate the properties of the generated time series is discussed. In Section 5 we present
all obtained results and observed statistical properties of generated time series. In Appendix A the
average length of overshoots is derived for the case of Brownian motion with the constant trend.
In Appendix B the pseudo-code of the directional-change intrinsic time algorithm is provided.
Appendix C concludes the paper with a collection of graphics describing all auxiliary experiments.
2. Intrinsic Time
The number of transactions as well as traded volumes in liquid markets is much lower during
holidays or weekends than during working days or after some unexpected but significant news
(Chordia et al. 2001). This evidence directly contributes to the continuous changes in the financial
time series volatility over time (Blattberg and Gonedes 1974, Christie 1982, Scott 1987). Thanks
to this non-homogeneous nature of financial markets, one can find drastically distinct price curve
evolution within two separate historical time intervals of equal length. One period could be char-
acterised by an instant price drop by several percents and not less instant recovery to the same
level (for example, a flash crash) when the other could represent an absolute standstill. The latter
often happens when the market is completely inactive due to holidays. Despite this well-known
fact, the historical dynamics of financial markets have been mostly analysed using snapshots of
market states equally spaced in time. Such snapshots are typically either too infrequent and do not
allow capturing of all the available high-frequency information (Zhou 1996) or are too numerous
which cause the excessive noisy details of the created set of data (Zhang et al. 2005).
Mandelbrot and Taylor (1967) were one of the first researchers to propose an alternative event-
based paradigm for modelling and analysing financial time series. Price changes over a fixed number
of transactions were studied and compared to the price changes over fixed periods of time. Later,
Guillaume et al. (1997) extended the early set of alternative data analysis techniques by introducing
the directional-change intrinsic time concept. According to that concept, all ticks happen as a result
of alternating rising and falling prices (returns) of a certain size. The desired for the observer size
of the returns can be arbitrary chosen This allows independent monitoring of the trend changes
at various scales. It is especially important taking into account that same price curve can be
1 Scaling law (power law): the mathematical relationship between two variables that holds true over multiple orders of magnitude.
Electroniccopy
available at:
Directional Change Event point (DC) Overshoot Move, 𝜔
End of an Overshoot (OS) Directional Change, 𝛿
Overshoot Move, ω
Price
Directional Change, 𝛿
Size of Directional
Change, δ %
Physical Time
Intrinsic Time
Figure 1. Example of a price curve dissected into a collection of directional changes using a threshold δ. Two
types of sections are typical for each consecutive pair of intrinsic events: directional-change (marked by the solid
line) and overshoot (dashed line). The direction of the initial mode is chosen to be downward. The first (left) grey
square marks the first directional-change event which occurs when the price moves downward by δ percent from
the local extreme. The local extreme for the given mode, coincides with the highest observed price (the first grey
circle). The mode alternates (upward) as soon as a new directional-change point becomes registered. After this
step, the local extreme coincides with the smallest observed price since the time of the latest directional-change.
The next upward event is registered when a positive return of the size δ happens measured from the local extreme.
After this, the mode alternates again and the dissection process continues.
characterised by different directions of low- and high-scale trends. The phenomenon has been
frequently exploited by traders. One specific trading strategy called ‘Triple Screen Trading System’
was proposed by Elder (2014).
The purpose of the directional-change intrinsic time is to register moments at which the price
curve alternates its trend of the given scale. It also finds extreme prices which correspond to local
maximum or minimum between two consecutive trend flips. The next paragraph contains a detailed
description of the dissection procedure of tick-by-tick data by the directional-change algorithm. An
example of a real price curve dissected into a collection of intrinsic events is shown in figure 1.
The latest observed tick price should be taken as the starting point to initialise the dissection
of a price curve into a set of directional changes. Then, the relative size of the directional-change
thresholds δ and the initial direction of the trend should be selected. There are only two possible
states of the trend which correspond to each given threshold δ: mode up for the upward or mode
down for the downward trend. The initial value of the extreme price Sext coincides with the first
price used to initialise the algorithm. Each new tick has to be compared to the latest registered
extreme. If the current mode is up (down) and the newest price Stick is higher (lower) than the
extreme price Sext , then the Sext takes the value of Stick . Alternatively, the distance between the
latest price Stick and the current local extreme Sext should be compared to the size of the threshold
δ. If the distance is bigger or equal to δ then the current price is a new directional-change point. At
this moment, the mode should be changed to the opposite one and the local extreme reinitialised by
Sext ← Stick . The overshoot part ω of the price curve corresponds to the price trajectory between
the latest directional-change and the local extreme points.
Overshoot intrinsic events (overshoots events) are registered every time whenever the size of
Electroniccopy
available at:
the overshoot ω becomes a multiple of the dissection threshold δ (Golub et al. 2017). We employ
this category of intrinsic events in addition to the initially proposed in Guillaume et al. (1997)
directional changes. The overshoot events will be connected to the empirically observed length of
the overshoot section with the decision-making mechanism of the agent-based model. As the reader
can see from figure 1, the minimum size of overshoot is equal to zero. That can be observed when
the price does not follow the trend and makes a reversal right after a new directional-change point.
At the same time, there is no the upper bound on the size of the overshoot section. Therefore,
there is no a limit on the number of overshoot intrinsic events between two consequent directional
changes. The details on the expected length of ω will be provided in Section 4.2 where the overshoot
scaling law is described. The pseudo-code of the dissection algorithm is provided in Appendix B.
The recording of the price activity in directional-change intrinsic time does not rely on the
exogenous evolution of physical time. Only endogenous price moves define the steps at which a
measure should be taken. Interesting properties of the financial time series emerge in this case.
Guillaume et al. (1997) presented a scaling-law discovered with the help of their new event-based
concept. The scaling law establishes the relationship between the chosen thresholds δ and the
number of the corresponding trend changes N (δ) registered within the given time period:
E
δ
N (δ) = , (1)
C
where C ∈ (0, +∞) and E ∈ (−∞, +∞) are scaling law parameters. This scaling law states that
the number of directional changes observed in a data sample of the fixed length relates to the size
of the chosen threshold as a power law function. It is worth noting that although the size of the
threshold δ can range from zero to infinity, the practically selected values depend on the volatility
of the analysed process and the frequency of price changes. For example, Glattfelder et al. (2011)
selected the set of thresholds ranging from 0.01% to 5% to analyse scaling laws of foreign exchange
rates. The typical annualised volatility of FX rates is about 20% which means that the selected
range covers the majority of price trajectories observed within several years. At the same time, the
thresholds are practically bigger applied to the analysis of emerging markets such as their volatility
is substantially higher (Petrov et al. 2018).
The intrinsic time ticks more often during periods of high volatility and less often when the
market is relatively quiet. The latter insures that no important information will be lost as well as
no noisy points will enter the created data sample. The visualisation of the statement is provided
in figure 1 where the last two directional-change events are registered after a weekend (right part
of the plot where the exchange rate does not move). There were no trades therefore no price
moves and therefore no information which could enrich the collected recorded. Additionally, the
directional-change intrinsic time allows capturing the most valuable moments represented by the
local maximums and minimums indicating the end of each trend. All price ticks between the
registered intrinsic events are considered as noisy at the scale given by the size of the threshold
and become ignored. The prices, neglected by one threshold can at the same time play the role of
tipping points at another scale. This multi-scale property of the directional-change intrinsic time
can be used to decrease the signal to noise ratio in each specific experimental scenario sensitive to
the scale of market moves.
There are various advantages of the event-based approach employed for high-frequency data
analysis. They have been successfully applied and described in multiple papers. Among others:
(i) The directional-change algorithm was used to discover numerous scaling laws hidden in
the price curves from the FX market. The found properties are expected to improve the
inferences we make on the price evolution through analysing liquid markets (Glattfelder
et al. 2011);
(ii) Several directional-change thresholds of various sizes can be initiated at the same time. The
evolution of such multiscale system can be employed to describe prices of the given financial
Electroniccopy
available at:
time series. This interpretation of the market dynamic leads to an efficient estimator of the
price trajectory unlikeliness. This unlikeliness is later used as the indicator of the current
and the forecasted liquidity of the market (Golub et al. 2014);
(iii) The number of directional-changes grows fast in periods of high volatility. The number stays
small when the volatility is close to zero. Thus, the approach can be used as a volatility
estimator of the given time series. The number of events as well as the magnitude of the
trend components are the proxy for the volatility size (Petrov et al. 2018).
A large number of directional changes can be registered if the size of the threshold δ is significantly
smaller than the volatility of the analysed time series. The average numbers of overshoot intrinsic
events observed in the upward and downward local trends is approximately equal to each other if
the overall trend of the time series is equal to zero. Alternatively, if there is a persistent trend, the
price curve tends to overshoot more often when the type of the directional-change mode coincides
with the direction of the overall trend. Thus, the average numbers of overshoot events in upward
and downward trends will not be equal to each other. However, it is possible to modify the original
algorithm in such a way that the trend will have no impact on the balance of the overshoots
number. A couple of distinct thresholds can be used to monitor two classes of directional-change
modes independently: δup to register events which happen within the upward trend and δdown for
events within the downward trend. The theoretical analysis in Appendix A demonstrates how the
trend size affects the expected length of the overshoot sections. As it can be seen from equations
(A6) and (A6), the expected size of an overshoot depends on the trend of the market µ and on
its volatility σ. For example, if the trend is negative, then the expected upward overshoots ω(δup )
are equal to downward overshoots ω(δdown ) only if δup < δdown . It is important to note that the
length of the overshoot section contributes to the number of overshoot intrinsic events. Therefore,
the equations from Appendix A can be used as the reference to the expected quantity of intrinsic
events per given period of time.
Most of the analytical tools developed in finance were initially built using the assumption that
the analysed market does not exhibit any particular trend. However, real price curves are collections
of alternating trends of various scales. Therefore, parameters of some trading or risk-management
algorithms require adjustments which can compensate nonzero trend typical for the given market
and the selected period of time. The directional-change intrinsic time concept, equipped by the
aforementioned asymmetric thresholds property, is one of the instruments capable to deal with that
trend problem. A couple of directional-change thresholds (δup , δdown ) can be selected in such a way
that the number of upward and downward overshoot intrinsic events will be always equal to each
other. However, the price trend cannot be estimated ex-ante. Thus, it is impossible to say which
pair of thresholds should be chosen to neutralise the drift expected in the near feature. An efficient
solution of the problem was proposed in Golub et al. (2017). The authors describe an automated
trading algorithm utilising a set of statistical properties accompanying the directional-change in-
trinsic time. The algorithm efficiently generates profit by capturing the price curve coastline and
trading at the moments of expected alternating trends. The equal number of buy and sell trades
should happen within periods of the growing and the falling price sections to ensure stable positive
returns. In other words, at any moment of time, the algorithm should have precisely tuned pair of
upward and downward thresholds which corresponds to the currently present trend. The authors
used the accumulated inventory of the trading bot as the proxy for the trend. The selected thresh-
olds should be changed every time when the absolute size of the inventory crosses selected limits.
In our work, we are creating a trivial agent-based model where the artificial agents do not keep
track of their inventory. Instead, each of them has a unique ex-ante specified pair of non-modified
thresholds. Such settings lead to the case when at any moment of time there is an agent whose
couple of thresholds captures the current price trajectory in the best possible way (registering equal
number of upward and downward overshoot events). The whole set of agents and their parameters
will be described in the following section.
Electroniccopy
available at:
3. Structure of the model
Financial markets can be schematically described as the interplay of two components. First, a
group of trading agents characterised by the diverse range of behavioural patterns. Second, the
exchange rate response on the aggregated agents trades (see, for example, Delage et al. (2010)).
We stick to this generalised representation to keep our model as simple as possible. Therefore, the
agent-based model has: a set of agents buying and selling a fixed volume only at the moment when
they register an intrinsic event; the volume impact function, a special algorithm which generates
the next price move (return) as a reaction to the aggregated imbalance of the number of buyers
and sellers. The latter will be described a few paragraphs below.
The demand and supply create equilibrium in any efficient financial system. In other words,
the total buy and sell volumes over given period of time are equal to each other. Nevertheless,
our agent-based model does not assume the complete equilibrium. Instead, we postulate that the
excess volumes formed by all trading agents can be endlessly bought and sold to and from some
external counterparties. The excess volume will be later connected to the price returns through the
volume impact function.
The whole life cycle of each agent consists of only two actions: opening a long or short position
and waiting for a moment to reverse it. The agents do not exploit each possible opportunity to
perform trades (in the same way how it happens in the real world): they flip their positions with
probability Pf lip which makes the performance of the model more realistic. The moments when the
agents try to open or flip the positions coincide with the occasion when intrinsic events become
observed. The sequence of such moments is unique for each agent since it is determined by the
assigned directional-change thresholds δup and δdown .
For simplicity, all agents trade identical volumes equal to one lot. To flip a position from long
to short an agent shorts one unit to close its long position and shorts an additional unit to open a
short position. In total two units should be sold. A similar procedure is in place for flipping from a
short to a long position, whereas the agent buys two units. Thus, at any iteration there are always
Nlong agents who decided to flip their position from short to long after analysing the latest return
and Nshort agents who decided to become short instead of long. The value ∆Nn = Nlong − Nshort
indicates the excess demand or supply at the step n. The variable ∆Nn , also called the net volume,
is used by the model to determine the subsequent price change using the volume impact function.
A simplified example of a sequence of intrinsic events registered by the trader with parameters
δup = 3 and δdown = 2 is visualised in figure 2.
Real market participants have a diverse set of trading strategies: trading in working days or
weekends, using technical or fundamental analysis, preferring high-frequency trading or holding
long-term positions (see, for example, the survey of US market provided by Cheung and Chinn
(2001)). Good agent-based models aimed to mimic real liquid markets should also be characterised
by such varied behaviour. We emulate a complex system by creating a group of intrinsic event agents
each of which has a unique set of parameters δup and δdown which impact their trading patterns.
The model does not have a pair of agents with completely the same settings. Each agent interprets
a price curve from the point of view of the directional-change intrinsic time. The interpretation
depends on the size of the assigned thresholds and their dissimilarity. Some of the simulated traders
can register intrinsic events at each new price tick by operating with small enough thresholds (like
real high-frequency traders). At the same time, traders with large thresholds register intrinsic time
events significantly less frequently which makes their behaviour similar to the one of the long-
term investors. Therefore, the selected diverse composition of the model makes the behaviour of
individual traders exclusive and the whole range of strategies manifold. As a result, various trading
activity patterns are reproduced.
The goal of the agent-based model is to generate a set of returns characterised by the same
statistical properties found in the FX market. Each produced return is assumed to be defined in
logarithmic terms (log-returns). This assumption makes it possible to directly compare aggregated
returns and the size of the directional-change thresholds typically expressed in the percentage
Electroniccopy
available at:
15 Intrinsic Events
14 Local extreme
13
12
11
Price
10
9
8
7
6
5 A B C D
4
0 1 2 3 4 5 6 7 8 9 10 11 12 13
Step
Figure 2. An example of a price curve dissected by a set of intrinsic events at which the agent with parameters
δup = 3 and δdown = 2 opens a position and makes trades. For simplicity, the chosen thresholds are expressed in
the absolute price moves instead of the relative ones. The initial price is equal to 10 and the agents initial mode
is down. Red arrows mark the distance measured between local extremes and the following directional change
points. Green arrows label price moves which lead to the next overshoot intrinsic event. Letters A, B, C and D
are put here to tag four intervals of alternating modes. The agent registers its first directional-change intrinsic
event as soon as the price goes down from the local extreme by at least two points (labelled by a circle). Since
this is the first intrinsic event observed by the agent, it opens its first position at this point (step number 2). Then
it is waiting for the next event which happens at step 3 after a big price move by four points up from the latest
local extreme (coincides with the previous directional change intrinsic event, a grey square). Independently on its
previous trading decision, the agent keeps analysing the price curve. At step 5 the price has made an overshoot
move measured from the preceding directional-change point the size of which is equal to up (a grey triangle).
This point marks the first overshoot intrinsic event. Though the price continues its overshoot move up, it does
not go far enough to trigger another overshoot intrinsic event and the next tipping point becomes a directional
change at step 8. The next two overshoot intrinsic events happen at steps 9 and 11. The last directional change
concludes the example at step 13.
terms. In other words, returns between two given prices at steps m and n (xn and xm ) are defined
as r(n, m) ≈ log(xm )−log(xn ) = Sm −Sn where Sm and Sn are sums of returns accumulated by the
steps m and n. Thus, a new directional-change intrinsic event happens when the absolute distance
between the latest price and the local extreme becomes bigger or equal to δ. This simplification
significantly facilitates all computations and will be used in the rest of the article.
To simplify the model, we represent the whole set of agents as a collection of separate points on
a square grid (figure 3). Each point corresponds to a couple of thresholds assigned to one unique
trader. Values δdown are located along the horizontal side of the grid (x-axis) and all included in
the model thresholds δup are placed along the vertical side (y-axis). It is worth mentioning that the
given setup is built on the assumption that any price return should become a signal for a trade.
Thus, the distance between two consecutive values on the same axis was selected to be equal to one
price tick. This guarantees that any cumulative return will serve as a new signal for the agent-based
model.
The geometrical size of the grid defines the extent to which the agents cover the diversity of
various trading patterns. We split the grid by L points horizontally and vertically to assure the
symmetry of the trading strategies. Figure 1(a) in Appendix C demonstrates the number of trades
performed by all individual agents from the entire grid. As expected, agents with the smallest
thresholds located in the left bottom corner make the biggest number of trades (high-frequency
traders). The right top corner represents the rarely trading of agents who perform using the biggest
thresholds (long-term holders).
Electroniccopy
available at:
𝛿𝑢𝑝
I
3
II
2
III
1
1 2 3 𝛿𝑑𝑜𝑤𝑛
Figure 3. A part of the grid representing the collection of trading agents. Each point corresponds to an agent
defined by a set of unique parameters (δup , δdown ). δup and δdown are the size of upward and downward directional-
change thresholds correspondingly. We mark regions with specific properties by numbers I, II and III. In the
region I there are traders with the upward directional change thresholds larger than the downward ones (δup >
δdown ). The region II contains ‘symmetric’ agents (δup = δdown ). Region III labels all agents with downward
thresholds larger than upward ones (δup < δdown ).).
The grid can visually be divided into three separate sections according to the types of agents
activity (figure 3). Traders from section I have upward directional-change thresholds larger than
the downward ones (δup > δdown ). These agents register the equal number of overshoot intrinsic
events within local trends upward and downwards only when there is a persistent positive trend.
Otherwise, the average number of overshoot intrinsic events registered within the directional-change
mode downwards prevail over of the average number within the upward sections. These agents tend
to exploit more trading opportunities when the trend is negative or equal to zero. Agents from
the region II have equal upward and downward thresholds and are called diagonal agents. The
absence of any global trend is the most stable condition for them. In that case, they witness the
equal number overshoot intrinsic events within local trends in both directions. Region III marks all
agents with upward directional-change thresholds smaller than the downward ones (δup < δdown ).
Their behaviour is the opposite to the behaviour of traders from the region I. For each trader from
the region I there is a corresponding opposite agent from the region III. Therefore, the complete
set is fully symmetric and thus imitates the counterbalance of different traders in the real financial
world. We will show later that one can simulate steady trends of given sizes by deactivating agents
from specific parts of the grid.
The lack of demand motivates the supply side to reduce prices and the lack of supply affects the
price rises (Walras 2013). This empirical observation was used in our work to define the volume
impact function for the developed agent-based model. The function is committed to computing
price changes caused by the imbalance between the total demand and supply. The exact shape
of the function may depend on various factors (Lillo et al. 2003). Among them: the selected time
horizons where the impact is observed; the size of the traded volumes; the types of markets where
the trades are performed and many others. Several research works have been done on this topic
and different models were proposed. A stable and linear impact function was described by Kyle
(1985). Huberman and Stanzl (2004) provided later the proof that permanent price impact must
indeed be linear while the temporary one can be of a more general form. At the same time, more
sophisticated non-linear functions were reported (Hasbrouck 1991, Hausman et al. 1992, Kempf
and Korn 1999). We decided to choose the model proposed in the relatively recent work Bouchaud
(2010) for our experiments. According to Bouchaud, the impact of trading volume is non-linear
and one of the best approximations is the square root function. Therefore, we endow the market
Electroniccopy
available at:
response to the agents aggregated trades in the following way:
j p k
rn (∆Nn ) ≡ Sn − Sn−1 = αsgn(∆Nn ) |∆Nn | , (2)
where: rn (∆Nn ) is the price change (return) at the step n; α is the parameter limiting the minimum
price shift; sgn(.) and b.c are the sign and the floor functions correspondingly.
There is a situation among all possible scenarios when the number of agents deciding to flip
short positions is less than the number of agents deciding to do the opposite just by one agent.
This case leads to the imbalanced volume equal to the sum of the volume to close a position and
the volume to open an opposite. Therefore, the smallest possible non-zero value ∆Nn is equal to
two. The parameter α was chosen to guarantee that in such situation the√market will respond to
the smallest possible imbalance by changing the price by one unit: α = 2/2. We use the floor
function to simulate only discreet price changes typically observed in real markets.
Periods of economic equilibrium happen when demand meets supply (Debreu 1987). In this case,
the sequence of trades happened at the market does not cause any price change, so the evolution
of the price curve becomes temporarily stable. Such states do not last notably long. Even small
market fluctuations enliven further asymmetric activity of the sell and buy sides. Eventually, the
disturbance becomes portrayed in a sequence of substantial price moves. Similar to the reality,
the proposed agent-based model is capable of producing zero difference between the number of
all buyers and sellers which entails zero net volume. This volume being put into equation 2 does
not cause any price change. The agents do not receive any new information and thus do not have
a chance to make decisions on further trades without price changes. As a result, this puts the
following evolution of the price curve on hold. To reactivate the trading, we add a small random
price shift upward or downward with equal probability 0.5. The shift happens whenever the net
volume ∆Nn is equal to zero. The size of the random move was chosen to be big enough to trigger
a new intrinsic event for agents who have one or both thresholds equal to the size of the smallest
price tick: δ = 1. This way, these agents receive a new piece of information and try to flip their
opened positions creating new demand or supply which leads to the consequent price change.
We used the next parameters√ for the main simulations: initial price level S0 = 0; minimum price
step (a tick) rmin = 1; α = 2/2; smallest and biggest thresholds δmin = 1 and δmax = 50; the step
between two consecutive thresholds on the grid equal to 1; total number of trading agents is 2500.
Prices, produced by the model, are aggregated returns computed as responses to all imbalanced
volumes that happened in the past. Therefore, the zero initial price simply indicates that no returns
happened before the model was initialised. The smallest threshold δmin = 1 guarantees that any
elementary price move will trigger a new intrinsic event for agents operating at the minimal scale.
The size of the probability to flip position Pf lip coincides with the empirically and theoretically
found probability to register a new overshoot event before the directional-change one (Golub et al.
2014):
Pf lip = e−1 = P(ω(δ) ≥ δ). (3)
The maximum possible price move rmax can be observed only when all agents defined in the model
decide to flip their positions simultaneously. Thus, the largest absolute price change is determined
by the number of agents on the grid and is directly connected to its length L:
√
rmax = αL 2. (4)
To generate such big return, all agents should register intrinsic events at the same moment of time
and their opened positions have to be of the same type. In the real markets, such significant returns
with negative sign are usually interpreted as crashes. They do not happen on a daily basis and
are observed when large numbers of market actors accidentally make similar trading decisions in
10
Electroniccopy
available at:
Net Volume
Δ𝑁𝑁 = 0 d)
Random Return b) c) Generated Return

a)
𝑟𝑟 = ±1 𝑟𝑟(Δ𝑁𝑁)
Net Volume
e) Δ𝑁𝑁 ≠ 0
Figure 4. The schematic summary of the algorithm used to generate the synthetic time series. a) The model
generates a random return in order to create the first signal for the group of agents represented by a grid of
thresholds. b) All agents consider the latest price return as a signal in their local directional-change intrinsic time
defined by the pairs of thresholds (δup , δdown ). Some group of the agents flip their opened positions reacting to the
observed return. The generated net volume ∆N is expressed as the overall difference between the total number of
agents decided to flip long and short positions. The net volume can be: d) equal to zero which means that no new
signal is generated and a random price shift should happen to provide a new signal for the agents; e) not equal
to zero which means the volume impact function (2) can generate a new return as the reaction to the aggregated
traders behaviour. The new return is the input for the grid of agents at the next step of the simulation.
one critical instant. In our model, the probability to observe the maximum price move of the size
rmax is likewise possible but minuscule. It is guaranteed by the condition that the decision to flip
positions should happen at the same time for all agents. The latter happens with the probability
2
PfLlip which rapidly tends to zero with increasing the size of the grid L.
The schematic summary of the algorithm used to generate the synthetic time series is provided
in figure 4.
4. Benchmarks
The main goal of this research work is to check whether the agent-based model operating in the
directional-change intrinsic time is capable of generating synthetic time series, statistical properties
of which are coherent with the ones typical for foreign exchange market. Several benchmarks have
been chosen to verify the accuracy of the model. The whole set of tests consists of four traditional
methods often used in research with the same intent and a new approach wholly based on the
directional-change intrinsic time. We propose that the latter is the best way of evaluating agent-
based models. Further, we describe all selected benchmarks in details.
4.1. Traditional benchmarks

One of the well-known facts about the market microstructures is that price returns in any liquid
market do not exhibit significant linear autocorrelation (Arneodo et al. 1996). This phenomenon is
also formulated as the ‘efficient market hypothesis’: prices instantly and fully reflect all available
information making it impossible to build a simple trading strategy based on the ‘statistical ar-
bitrage’ (Basu 1977). Only at ultra-short time interval price curves could have slightly correlated
historical returns. At this scale, the market, as a global multi-agent system, is still absorbing a new
piece of information. According to Cont et al. (1997), ‘in a few minutes’ the autocorrelation can
be safely assumed to be equal to zero. Low autocorrelation of returns is one of the most popular
stylised facts which regularly accompany liquid markets. Therefore, we selected that statistical
property to be among our benchmarks. The autocorrelation function of the series of returns M
11
Electroniccopy
available at:
with mean µ and variance σ(τ ) computed using the given lag τ is defined as
E[(Mt − µ)(Mt+τ − µ)]

R(τ ) = . (5)
σ(τ )2
The second stylised fact chosen for the analysis is related to the fat-tailed distribution of returns at
the frequency higher or equal to one day. This property, also known as excess kurtosis, is thoroughly
discussed in the book Mandelbrot and Hudson (2010). The authors point out that there is the full
range of theories build on top of the assumption that returns are normally distributed (including
the famous work of Black and Scholes (1973)). Nevertheless, the processes in the real financial
markets have always been different from Brownian motion and this assumption is a severe flaw
of any related financial model. Therefore, it is important to create the model which is able to
reproduce returns characterised by the fat-tailed distribution. We measure the excess kurtosis k in
the following way:
h(r(t, τ ) − hr(t, τ )i)4 i

k= − 3, (6)
σ(τ )4
where σ(τ ) is the variance of log-returns r(t, τ ) = S(t) − S(t − τ ) computed with the lag τ . Excess
kurtosis k = 0 means the absolutely normal distribution. Values bigger than zero indicate deviations
from it. Brown and Warner (1985) demonstrated that in the stock market excess kurtosis is usually
below 4; Cont (2001) mentioned that for S&P 500 futures the value is around 16; the kurtosis is
equal to 60 for Dollar/Swiss Franc futures; in Gençay et al. (2001) one can find that for USD FX
rates it is roughly 30 (through 10 minutes lag intervals).
Excess kurtosis is typically positive when time lag is relatively small and it tends to zero as the
lag increases. This fact is called the aggregated normality or aggregational Gaussianity and can
be accounted for by the ‘mixture of normals’ (Antypas et al. 2013). We select several increasing
time lags to demonstrate that the distribution of returns produced by the agent-based model has
distinct excess kurtosis. The values of the kurtosis are expected to be dependent on the selected
scale. Thus, the aggregated normality is the third benchmark in our set.
The scale-invariance of the absolute price change (return) to the period of time when it occurs
(Müller et al. 1990, Mantegna and Stanley 1995, Dacorogna et al. 2001) is the fourth selected
benchmark. There is no agreement on the origin of this invariance called scaling law or power
law and various assumptions can be found in the literature (Bouchaud 2001, Farmer et al. 2004,
Joulin et al. 2008). Its omnipresence in finance incentivised scientists to apply it for real financial
problems such as risk management and volatility modelling (Ghashghaie et al. 1996, Gabaix et al.
2003, Di Matteo 2007). We check whether the returns generated by the directional-change agent-
based model are also characterised by this power law.
Most of the scaling laws found in financial markets rely on intervals measured in physical time.
It is the essential fact to be aware of constricting any model of the real financial system. As it was
discussed in Section 2, measures of the markets dynamic performed in physical time are not efficient
in catching extreme events. Physical time does not adjust its flow to the speed of actions affecting
the market. The extreme events, in turn, are the most critical information for statistical analysis.
On the other hand, agent-based models are neither capable of ‘feeling’ the flow of physical time
(they are a piece of code, after all). Therefore, it is not a trivial question of how one can connect the
sequence of actions performed by simulated agents to real seconds, days and years. For example, the
scale of volatility has a direct impact on the number of observed directional-change intrinsic events
(see figure 1). Each intrinsic event is a signal for the corresponding agent who registered it. Thus,
the volatility clustering means compression and expansion of the inner clock used by the agent-
based model. The exact way of how the agents activity relates to the speed of physical time should
be algorithmically predetermined. The assumption that the agents make decisions at equidistant
moments of time, for example, every second, is the most popular approach used to bridge the gap
12
Electroniccopy
available at:
between physical and intrinsic time in agent-based modelling. In this case, for instance, 20 000 000
steps in intrinsic time would correspond to 231.5 trading days (that is, close to a full trading year
or 252 days). To validate the model by the traditional benchmarks, we follow the same principles
and postulate that price changes can be observed only over discrete moments of time.
4.2. The ultimate benchmark

Scaling laws are ubiquitous properties of our world and are present in any domain of natural and
social phenomena such as physics, biology and social sciences (Andriani and McKelvey 2007). The
goal of the agent based model is the simulation of a collection of returns statistical properties of
which coincide with the ones typical for the real financial markets. Therefore, it is also important
to consider such omnipresent scaling properties while validating artificial sets of interacting agents.
The traditional benchmarks mentioned in the previous section can be successfully applied for
describing processes and actions happening in the real world. Time is the universal measure of
actions and interactions in such real environment. Therefore, all these benchmarks account for
physical time as the factor of scale. Unfortunately, the artificial agents have no ‘feeling’ of elapsed
periods and operate using the ‘signal-reply’ logic only. Therefore, applying these benchmarks as
well as methods and tools where time plays the crucial role for interpreting the agents automated
actions is not fully correct. We consulted the work of Glattfelder et al. (2011) in the search for a new
omnipresent property of high-frequency markets which can be ultimately used to verify any agent-
based model. The main criteria: it should be possible to employ the property as a benchmark
independently on the relation of physical time and the sequence of agents actions. Fortunately,
there is one, out of 12, newly described scaling laws which perfectly fits the mentioned criteria: the
‘overshoot scaling law’. It is fully based on the directional-change concept where only relative price
moves dictate sequences of events dismaying time intervals between them. According to Glattfelder
et al. (2011) the average length of the overshoot section is approximately equal to the size of the
directional-change threshold. Golub et al. (2014) analytically showed that in the continuous process
with zero trend the probability of overshoot ω(δ) reaching the length l equals to exp (−l/δ), i.e.

l
P(ω(δ) = l) = exp − . (7)
δ
This dependence reveals the exponential relation between the length of a directional-change thresh-
old δ and the expected length of the overshoot ω(δ). From equation (7) it follows that the expected
overshoot E[ω(δ)] is equal to the size of the directional change threshold δ:
E[ω(δ)] = δ. (8)
Glattfelder et al. (2011) have empirically shown that the average coefficients EOS and COS of
EOS
δ
the overshoot scaling law hω(δ)i = COS across all 13 currency pairs analysed in the work
are EOS = 1.04 and COS = 1.06. Lowest and the highest registered values specific for the pairs
are: COS,low = 0.98, COS,high = 1.17; EOS,low = 0.98, EOS,high = 1.08. This statistical property
is fully agnostic to the volatility and does not relate on any values defined in physical time. Its
application as a benchmark to an agent-based model does not require any additional assumption
on the connection of agents activity and the flow of real physical time. Therefore, we call that
scaling law as the main benchmark of our model.
We used the same notation proposed in Glattfelder et al. (2011) to validate scaling laws mentioned
above. We build log-log plots where on X- and Y -axis the base and dependent values of scaling
laws are placed. We assume a linear relationship between the response variable Y (for example, the
average size of price returns) and the fixed variables X (for example, a period of time): Y = A+BX,
where A and B are unknown parameters to be estimated. Thus, B defines the slope of the line on
13
Electroniccopy
available at:
1.20
Average
1.15
1.10
Price
1.05
1.00
0.95
0.90
0 2000 4000 6000 8000 10000
Step
Figure 5. 10 time series generated by the agents from the entire grid of 50 by 50 equally distributed points (see
figure 3). The presented price curves were obtained by applying the exponential function to the set of logarithmic
returns generated by the model: P rice = exp(Slog /1000). Here 1000 is just a scaling factor.
the log-log plot and A is the intersection of Y -axis. In this way, a scaling law takes the following
form:
x E
y= , (9)
C
where y = exp Y , x = exp X, E = B and C = exp(−A/B).
5. Results
In this chapter, we highlight the main findings of the research work and review how components
of the agent-based model affect the properties of the generated time series.
All experiments have been performed in two steps. First, we analysed time series generated by
the agents from the entire grid operating simultaneously; second, we examined the effects induced
by I or III part of the grid being performing solely. The following two sub-chapters outline details
of each experiment.
5.1. The entire grid

As the first step, we tested the performance of the whole grid of directional-change agents. Param-
eters of the experiment are specified in Section 3. An example of 10 price curves generated by the
model with the help of the squared root impact function (2) is presented in figure 5. The curves
consist of various intervals with plateaus and sudden jumps thus mimicking features of the real
FX market. At the same time, there is no any prevailing trend which would define the evolution
of each given price curve. The red line represents the average aggregated return computed as the
result of 1000 independent simulations. This line remains horizontal throughout all 10 000 steps.
The latter confirms that the model where all agents participate in the trading does not possess any
deterministic impact on the trend.
Benchmarks introduced in Section 4 were applied to validate the synthetically generated data
sets. Autocorrelation function (ACF) of a time series consisting of 10 000 000 simulated returns is
14
Electroniccopy
available at:
1.2 1.0
1.0
1 step return
0.8 5 steps return
0.8
10 steps return
0.6 20 steps return
0.6
ACF
ACF
0.4
0.2 0.4
0.0
0.2
−0.2
−0.4 0.0
0 5 10 15 20 0 5 10 15 20
Lag Lag
(a) (b)
Figure 6. (a) Autocorrelation function of the generated time series. Lags are measured in the numbers of ele-
mentary price moves. (b) Autocorrelation function of absolute returns computed on the same time series using
four different lag steps.
shown in figure 6(a). The maximum negative correlation (-0.32) is observed for the lag size equal
to one step. The rest of the autocorrelation values is significantly smaller. The autocorrelation
function rapidly decays and becomes indistinguishable from zero after about ten steps.
Figure 6(b) shows slowly decaying autocorrelation functions of absolute returns computed over
four different lag steps. The bigger the lag step, the lower the decline of the ACF. This effect
confirms the ability of the agent-based model of replicating the slow decay of autocorrelation in
absolute returns. That is, the aggregational normality has been successfully validated.
The absolute price change scaling law is shown in figure 7(a). C and E are the scaling parameters
from equation (9) described in Section 4.2. R is the Pearson product-moment correlation coefficient.
One can clearly observe the linear dependence of the absolute price change and the elapsed time
interval expressed in logarithmic values. It is important to highlight that this scaling law employs
physical time to measure the intervals between given prices of the time series. There is no academic
agreement on the periodicity and intensity of trades assumed to be accomplished by artificial agents
in physical time. Both these values are determined by the assumption made in every particular
model. High trades frequency induced big size of price changes over short periods of time. Low
frequency results in the opposite. That is equivalent to the compression of time for the same agents
activity level. It leads to the increased or decreased volatility of the process and the bigger or lower
slope of the log-log plot Y = A + BX (see Section 4.2 for more details). The scaling parameters
E and C are directly connected to the coefficients of the plot through equation (9). Therefore, the
choice of the trade frequency in physical time directly affects the obtained scaling coefficients of the
absolute price change scaling law. As results, it is not possible to compare the parameters observed
in the FX markets with the ones generated by the model. Nevertheless, the fact that the absolute
price change of generated returns depicted in figure 7(a) is represented by a straight line on the
log-log plot (described by a power law function) is enough to confirm the successful replication of
the absolute price change scaling law.
Figure 8 shows distributions of returns generated by the model. The returns are computed over
four gradually increasing lags: 10, 50, 100, and 1000 steps. As it can be seen from the figure, the
distributions with lags up to the hundreds of steps (figures 8(a), 8(b), and 8(c)) are characterised
by noticeable fat tails. The fat tails disappear around the lag level of 1000 steps (figure 8(d)).
Assuming that the agents generate a new return every minute, the observed effect is in line with
15
Electroniccopy
available at:
103
Up
Down
101 Average
102
hω(δ)i (%)
∆S
100
101
C = 4.4 C = 1.04
E = 0.53 E = 1.04
R = 1.0 R = 1.0
10−1
100 1
10 102 103 104 105 10−1 100 101
Steps δ (%)
(a) (b)
Figure 7. (a) Average absolute price move as the function of the number of steps (the absolute price move scaling
law). (b) Overshoot scaling law build from a time series generated by the agents from the entire grid. Parameters
on the plot correspond to the average line. Approximation was done for δ > 0.3%. The same equation provided in
the description of figure 5 was used to transform thresholds from absolute values to relative ones (into percentage
terms). Coefficients of the Up line: C = 1.04, E = 1.05, R = 1.0; of the Down: C = 1.03, E = 1.03, R = 1.0. The
plot is based on 20 000 000 steps.
empirical results found in real markets (for example, values from Kullmann et al. (1999)). Measured
excess kurtosis k is equal to 2.73 at 10-steps lag and only 0.06 when the lag rises to 1000 steps.
The excess kurtosis rapidly decreases together with the growing lag size confirming the empirically
observed dependence. We present two probability plots for lags of 10 and 1000 steps in figure 9. The
probability plots confirm the aggregated normality once again. Nevertheless, the empirical analysis
of Kullmann et al. (1999) shows that the excess kurtosis at such small lags is usually much higher
(k = 10 and more). The intrinsic time agent-based model can also be used to generate returns
characterised by the size of the kurtosis typical for the real price data. The size of the grid and
thus the quantity and the diversity of trading agents directly contribute to the excess kurtosis. The
grid 50 by 50 points produces returns with k = 2.73 at 10-steps lag; the grid 100 by 100 points
result in k = 3.46; the grid 200 by 200 points generates time series with k = 6.01.
Finally, we validated the agent-based model using the properties of the overshoot scaling law. The
log-log plot of the average overshoot length versus the size of the directional-change threshold was
constructed (figure 7(b)). The revealed dependence appears to be linear. The scaling parameters are
E = 1.04 and C = 1.04. These values are exceedingly close to the ones observed in the real foreign
exchange market (see Section 4.2). We also investigated two supplementary versions of the overshoot
scaling law in addition to the one described in Glattfelder et al. (2011): the dependence computed
using only overshoots following upward directional changes (red dashed line in the figure) and
only overshoots following downward directional changes (green dashed line). There is no noticeable
difference between all three lines. This observation confirms that the directional-change agent-based
model is able to reproduce the overshoot scaling law. Results of the experiment are also in line
with the theoretical equations (A6) and (A6) provided in Appendix A.
Markets react to exogenous shocks such as news in multiple ways (Frank and Sanati 2018).
Positive and negative news could as permanently change the preceding price trend as well as
make a short-term disturbance. The latter can become rapidly absorbed by the system or become
pronounced in the corresponding price change. This sort of price change might be also observed
as instantly after the respective event as well as over significantly longer period. We checked how
16
Electroniccopy
available at:
0.25 0.09
10−20 10−11 10−3
10−1110−7 10−3
0.08
0.20 µ = 0.0 0.07 µ = 0.0
σ =2.73 σ =5.69
Frequency
Frequency
0.06
0.15 −20 −10 0 10 20
0.05 −30−20−10 0 10 20 30
0.04
0.10
Lag = 10 0.03
Lag = 50
0.05 0.02
0.01
0.00 0.00
−30 −20 −10 0 10 20 −40 −30 −20 −10 0 10 20 30
Return Return
(a) (b)
0.06
0.018
10−11 10−7 10−3
0.05
10−6 10−4 10−2

0.016
µ = 0.0 µ = 0.02
0.014
σ =7.91
Frequency
0.04
σ =23.75
Frequency
−40 −20 0 20 40
0.012
0.03 0.010 −100 −50 0 50 100
0.008
0.02 Lag = 1000
Lag = 100 0.006
0.01 0.004
0.002
0.00 0.000
−40 −20 0 20 40 −100 −50 0 50 100
Return Return
(c) (d)
Figure 8. Distribution of returns generated by the agents from the entire grid and computed at four different
lags. Their approximation by the Gaussian function is marked by red lines. Included sub-plots are the same
distributions but presented in the logarithmic scale.
(a) (b)
Figure 9. (a) Probability plots of returns computed with the lag equal to 10 steps. (b) Probability plots of returns
computed with the lag equal to 1000 steps. The total number of prices in the generated time series is 10 000 000.
the agent-based model operating in intrinsic time reacts to instant shocks of various amplitudes.
Each shock was simulated by a small portion of extra volume on the sell or the buy side added to
the net volume. The Monte Carlo simulation demonstrated that the extra volume added to the net
volume at the arbitrary step has instant and permanent impact on the average return. The size of
the impact is defined by equation (2). Results of the simulation can be requested from the authors.
Volume impact function connects the net volume generated by the set of agents at time t with
17
Electroniccopy
available at:
the induced price change observed at t + 1. The square root dependence (2) was selected for the
agent-based model although there is no scientific agreement on the universal form of the functions.
Therefore, linear and logarithmic impacts have also been tested in our work. Time series generated
with the help of these dependencies do not excel the same quality of stylised facts reproduced by
the square root function. The linear impact induces rapid price fluctuations around the initial level
and none of the statistical properties become observed. In turn, time series generated using the
logarithmic dependence expectantly have similar properties to the ones of time series produced
with the help of the squared root. The fat-tailed distribution of returns is of the same scale for
both functions. Nevertheless, scaling coefficients of the overshoot scaling law become noticeably
worse in case of the logarithmic impact. The coefficients are: C = 0.86, E = 0.95, R = 0.16 for
the linear impact; C = 1.38, E = 0.99, R = 1.0 for the logarithmic impact; C = 1.04, E = 1.04,
R = 1.0 for the squared root function (2).
The linear distribution of thresholds and the squared shape of the grid are aimed to replicate
the diversity of real groups of traders in FX markets. The selected parameters only approximately
replicate the composition of the trading strategies existent in the real world. Nevertheless, we
experimented with different sets of thresholds, such as their logarithmic distribution as well as the
radial fit and the circular shape of the grid. The obtained performance of the agent-based model
was close to the one offered by the simple squared grid whenever the number of agents remains
substantial.
5.2. Asymmetric regions

The entire grid of directional-change agents generate time series with zero average trend (Section
5.1). We claim that the reason of the observed phenomenon is the selected symmetry of the grid:
to each agent from the region I with not equal thresholds δup I and δ I
down there is an agent in the
III III I III
region III with thresholds δup and δdown such that δup = δdown and δdown I III . To confirm
= δup
the statement, we performed a series of tests where the imbalance of trading agents was created.
The imbalance is achieved by removing a part of agents from the initial grid. The effect observed
in the test is the clearly pronounced non-zero average trend of the generated time series. We
demonstrate results of two trivial experiments where only agents from the region I or region III
have been selected to trade (figure 10). The induced deviations of the average trend are upward
and downward correspondingly. The permanent trend observed in both experiments is present due
to the created imbalance between the number of agents supporting trends of opposite directions.
Agents from the region I tend to trade more often when the price is going down while agents from
the region III prefer the rising trend.
Several factors affect the average slope of all generated price curves: the total number of agents
forming the initial grid, the section of the grid used to generate a time series, the selected time
interval between two consecutive steps. We performed a set of experiments to visualise the impact
of the created grid asymmetry on generated prices. An extra directional-change agent with the
specific pair of thresholds was added to the initial set of agents forming the 50 by 50 points grid.
This resulted in 2500 + 1 agents trading simultaneously. We show in figure C3 the distribution of
final aggregated returns after 10 000 steps averaged over 1000 independent simulations. Additional
agents with symmetric thresholds δup = δdown have no pronounced impact on the average aggre-
gated return (parameter µ of the Gaussian distribution used for approximations) as well as on the
standard deviation (parameter σ). Agents with asymmetric thresholds tend to deviate the average
trend in the direction which coincides with the direction of their bigger thresholds. The deviation
is bigger in case of smaller δup and δdown . Agents with such thresholds are similar to high-frequency
traders and eager to react to all trend changes. As result, they make more trades than the traders
with substantially bigger directional-change thresholds (see figure 1(a)). These facts explain the
observed phenomenon. A set of overshoot scaling laws computed using the data generated in each
experimental setup is presented in figure C4. It is worth noting that lines related to the upward
18
Electroniccopy
available at:
1.30
Average
1.25
1.20
1.15
Price
1.10
1.05
1.00
0.95
0.90
0 2000 4000 6000 8000 10000
N
(a)
1.15
Average
1.10
1.05
1.00
Price
0.95
0.90
0.85
0.80
0.75
0 2000 4000 6000 8000 10000
N
(b)
Figure 10. (a) Time series generated by agents from the region I of the initial grid and (b) by agents from the
region III. Red lines represent average prices based on 1000 independent simulations. The initial grid size is 50
by 50 points.
and downward overshoots are symmetrically shifted from the diagonal line.
The number of agents participating in the trading effects the excess kurtosis of generated returns
(see Section 5.1). Grids of different sizes were tested to investigate whether there is any other
impact. Only agents from the I region were involved in the experiment. The average aggregated
returns of 1000 independent simulations are shown in figure 1(b). As it can be seen from the figure,
the bigger the size of the grid, the higher the deviation of the aggregated trend from the trend-less
case. The net volume which is the only parameter of the impact function (2) is the reason of the
observed deviation. Bigger number of agents creates higher standard deviation of the net volume.
The latter induces the revealed in figure 1(b) dependence of the grid size and the average trend.
The dependence has been also confirmed in another experiment results of which are presented in
figure C5. The average positive and negative volumes generated by the model were computed as
the function of the grid size. The experiment proves the statement that the bigger the size of the
19
Electroniccopy
available at:
grid, the higher the generated absolute net volume.
We would like to note that the stylised facts selected for this project are still reproduced even
when a part of the agents from the grid is deactivated. In figure C2 we include overshoot scaling laws
computed using the data generated by two separate parts of the grid: by region I (figure 2(a)) and
region III (figure 2(b)). The scaling coefficients C and E of studied dependencies insignificantly
differ from the ones observed empirically in Glattfelder et al. (2011). Therefore, we claim that the
agent-based model operating in directional-change intrinsic time can be used to simulate financial
time series with statistical properties closely related to high-frequency markets. Moreover, the trend
of the collection of generated returns can be specifically predefined. The set of the directional-
change agents is the parameter which can be used to setup the trend. A comprehensive analysis of
its precise direction is a topic for an independent work which is beyond the scope of this research
paper.
6. Conclusion
An agent-based model where trades happen in directional-change intrinsic time was tested in
this work. A set of artificial agents mimics the behaviour of real market participants by buying/
selling one lot of the traded asset. All trades happen at the moments when the directional-change
intrinsic time ticks. Two types of intrinsic events are considered: directional changes and overshoots.
Directional changes are moments when the price curve makes reversals equal to the size of the
selected threshold. Overshoot intrinsic events happen every time when the overshoot length is a
multiple of the of the corresponding directional change threshold. The agents flip their opened
positions from long to short and vice versa at times when they observe intrinsic events of given
magnitude. They flip the positions with the empirically and theoretically found probability to
register a new overshoot event before the directional-change one. The probability is aimed to
mimic behaviour of real market participants who do not exploit every trading opportunity. Traders
decisions are not conditional on time intervals between registered intrinsic events. The agents are
ignorant to the flow of physical time and consider prices as the only source of information about
the markets activity needed to perform.
The agent-based model replicates a wide range of trading patterns observed in the real world.
A unique pair of directional-change thresholds was assigned to each simulated trader. The size of
the thresholds as well as their ratio define the scale of price trends analysed by each agent. The
difference between the number of all agents decided to flip a long or short position at each new
price represented the net volume or imbalance between the supply and demand. The net volume
was used to calculate the impact of the imbalanced decisions on the further price change. The
empirically observed squared root function connects the net volume and the corresponding market
return.
A set of benchmarks was chosen to test the performance of the constructed agent-based model.
The model is assumed to be successful if it manages to reproduce ‘stylised facts’ discovered in the
real financial world. The utilised benchmarks: low auto-correlation of returns, fat-tailed distribution
of returns, aggregated normality, price jump and the overshoot scaling laws. The last benchmark
is fully agnostic to the connection of the simulated agents actions and the flow of physical time.
We propose that the scaling law is the universal test for any agent-based model.
The presented agent-based model operating in the directional-change intrinsic time has success-
fully passed all chosen tests. The latter lets us make an educated guess that real market participants
intentionally or unintentionally make trades using their own intrinsic time. In other words, they
have preferences on the scale of price returns after which the traders reverse opened positions. The
obtained knowledge can be used to substantially improve the quality of the inferences we make
about the connection of the aggregated traders behaviours and observed price changes.
20
Electroniccopy
available at:
Appendix A: Overshoot as the function of trend
The average length of an overshoot section is approximately equal to the length of the correspond-
ing directional change threshold (Glattfelder et al. 2011):
hω(δ)i ≈ δ. (A1)
This dependence was found not only in historical tick data but also in arithmetic Brownian motion.
Nevertheless, the average length of overshoots does not resemble the length of the selected threshold
if the time series has a constant trend. In turn, the average overshoot length depends on the size of
the trend (Golub et al. 2017). To find the exact form of this dependence, we select an arithmetic
Brownian motion with increments dSt , trend µ, and volatility σ:
dSt = St − St−1 = µdt + σdWt . (A2)
Golub et al. (2014) derived the probability of the upward and downward overshoot to reach some
fixed value x. The probabilities are the following:
 n o
 x (|µ| − µ) + (|µ| + µ) exp − 2|µ|δ
σ 2
up

P(ω(δup ) ≥ x) = exp − 2 · n o , (A3)
 σ 1 − exp − 2|µ|δup 
σ 2
 n o
 x (|µ| + µ) + (|µ| − µ) exp − 2|µ|δσdown
2

P(ω(δdown ) ≥ x) = exp − 2 · n o , (A4)
 σ 1 − exp − 2|µ|δdown 
σ 2
where ω(δup ) and ω(δdown ) stand for upward and downward overshoots correspondingly.
The expected value of the shown probability equations (A3) and (A4) F(x) = P(X ≥ x) is
Z ∞
E[X] = F(x)dx. (A5)
0
Using (A5) one can find that:

n o
σ 2 1 − exp − 2|µ|δ
σ 2
up
E[ω(δup , µ, σ)] = n o, (A6)

(|µ| − µ) + (|µ| + µ) exp − 2|µ|δ σ2
up
n o
σ 2 1 − exp − 2|µ|δσdown
2
E[ω(δdown , µ, σ)] = n o. (A7)
(|µ| + µ) + (|µ| − µ) exp − 2|µ|δσdown
2
The expected length of the all overshoots is the average of upward and downward expected
overshoots (equations (A6) and (A7)):
 n o n o 
σ2 1 − exp − 2|µ|δ
σ2
up
1 − exp − 2|µ|δdown
σ2
E[ω(δup , δdown , µ, σ)] =  n o+ n o .
2 (|µ| − µ) + (|µ| + µ) exp − 2|µ|δ2
up
(|µ| + µ) + (|µ| − µ) exp − 2|µ|δdown
2
σ σ
(A8)
21
Electroniccopy
available at:
3.0
Overshoot type
Upward
2.5 Downward
Total
2.0
1.5
ω
1.0
0.5
0.0
−1.0 −0.5 0.0 0.5 1.0
µ
Figure A1. Expected length of the overshoot section ω as the function of trend µ computed using equations (A6)
and (A7). The variance σ and threshold δ are fixed and equal to 1.
In figure A1 we demonstrate the dependence of the overshoot length ω on various trends µ when
volatility σ and the threshold size δ are fixed. The last two variables are assumed to be equal to one
(σ = 1; δ = 1). It is easy to notice that the lengths of upward and downward overshoots coincide
with each other only in case of zero trend. One can observe significant divergence of the curves for
any other value not equal to zero. The obtained result is quite intuitive: price tends to overshoot
more after an upward directional-change event in case of the overall ascending trend. At the same
time, it tends to make a directional-change reversal right after a downward directional-change
event. This observation suggests that for each non-zero trend µ there are such thresholds δup and
δdown that the total number of overshoot intrinsic events registered in the given time series is equal
to the number of overshoots observed within a time series of the same length and zero trend. This
property was directly used in the proposed agent-based model to replicate the diversity of trading
patterns (Section 3).
Appendix B: Dissection Algorithm
By Stick we mark the latest observed price, by Sext the local extreme, mode is the current mode
of the alternating trend which can be equal either up or down, δup and δdown are upward and
downward thresholds respectively, SIE is the price at which the latest intrinsic event was observed.
The algorithm returns 1 and -1 when the price curve hits levels of upward and downward directional-
change events correspondingly. 2 and -2 will be returned in case of overshoot intrinsic events
registered on ascending or descending trends.
22
Electroniccopy
available at:
Algorithm 1 Intrinsic Event
1: if first tick then
2: Sext ← Stick
3: SIE ← Stick
4: return 0
5: else if mode is up then
6: if Stick − Sext ≥ δup then
7: mode ← down
8: Sext ← Stick
9: SIE ← Stick
10: return 1
11: else if Stick < Sext then
12: Sext ← Stick
13: if SIE − Sext ≥ δdown then
14: SIE ← Stick
15: return −2
16: else
17: return 0
18: else if mode is down then
19: if Sext − Stick ≥ δdown then
20: mode ← up
21: Sext ← Stick
22: SIE ← Stick
23: return −1
24: else if Stick > Sext then
25: Sext ← Stick
26: if Sext − SIE ≥ δup then
27: SIE ← Stick
28: return 2
29: else
30: return 0
Appendix C: Additional experiments
23
Electroniccopy
available at:
40
103
30
δup
102
20
10 101
0
0 10 20 30 40
δdown
(a) (b)
Figure C1. (a) The average number of trades performed by directional-change agents with parameters δup and
δdown within 10 000-steps simulation. Parameters of the agents used for the simulation are provided in Section
3. (b) Average aggregated returns recorded within 10 000 steps (N ) of 1000 independent simulation when only
agents from the region I are active. 13 tests with grids of various sizes were performed. The smallest grid is 10
by 10 points. The biggest is 130 by 130. The applied increment is 10.
(a) (b)
Figure C2. Overshoot scaling laws computed using time series generated by the agents from (a) I and (b) III
regions of the grid only. Red dashed line (Up) corresponds to the average overshoot computed after directional-
changes upward. Green dashed line (Down) corresponds to the average overshoot measured after directional-
changes downward. The scaling parameters C, E, and R are shown for the lines representing the the average
overshoot lengths (blue dashed line).
24
Electroniccopy
available at:
δup = 1, δdown = 1 δup = 1, δdown = 2 δup = 1, δdown = 3 δup = 1, δdown = 4
µ =-1.2 µ =-167.6 µ =-82.8 µ =-48.7
σ =59.0 σ =73.1 σ =74.7 σ =73.0
0.006
0.004
0.002
0.000
−200 −100 0 100 200 −400 −300 −200 −100 0 −200 0 −200 −100 0 100
0.006
µ =171.0 µ =1.2 µ =-36.4 µ =-35.1
0.005 σ =71.9 σ =69.5 σ =70.7 σ =74.8
0.004
0.003
0.002
0.001
0.000
0 200 400 −200 0 200 −200 0 200 −200 −100 0 100
0.006
µ =81.5 µ =33.6 µ =-1.1 µ =-21.9
σ =74.9 σ =75.0 σ =70.7 σ =76.1
0.005
0.004
0.003
0.002
0.001
0.000
0 200 −200 0 200 −200 −100 0 100 200 −200 0 200
0.006
µ =52.6 µ =31.5 µ =18.2 µ =-1.0
0.005 σ =76.5 σ =72.9 σ =73.3 σ =68.2
0.004
0.003
0.002
0.001
0.000
−200 0 200 −200 0 200 −200 0 200 −200 −100 0 100 200
Figure C3. Impact of an additional intrinsic time agent on the average trend generated by the entire grid.
2500 + 1 traders participate in each simulation. Parameters of the supplementary agents are put on top of each
subplot. Experiment consists of 1000 independent simulations and 10000 steps. Normal distributions (red lines)
approximate the final aggregated returns. Blue rectangular mark zero trend level. Red rectangular stand for the
centre of the obtained distributions. The subplots are centred on the average aggregated return.
25
Electroniccopy
available at:
102 102 102 102
C = 0.88 C = 5.66 C = 2.56 C = 1.52

101
E = 0.96 101
E = 2.6 101
E = 1.49 101
E = 1.16
R = 1.0 R = 0.95 R = 0.98 R = 0.99
101 102 101 102 101 102 101 102
102 102 102 102
C = 5.42 C = 0.94 C = 1.25 C = 1.22

101
E = 2.53 101
E = 0.97 101
E = 1.08 101
E = 1.07
R = 0.95 R = 1.0 R = 1.0 R = 1.0
101 102 101 102 101 102 101 102
ω (δ)

102 102 102 102
C = 2.65 C = 1.2 C = 0.97 C = 1.05

101
E = 1.51 101
E = 1.06 101
E = 0.99 101
E = 1.01
R = 0.98 R = 1.0 R = 1.0 R = 1.0
101 102 101 102 101 102 101 102
102 102 102 102
C = 1.6 C = 1.2 C = 1.07 C = 1.02

101
E = 1.19 101
E = 1.06 101
E = 1.02 101
E = 1.0
R = 0.99 R = 1.0 R = 1.0 R = 1.0
101 102 101 102 101 102 101 102
δ
Figure C4. Overshoot scaling laws computed using the time series generated by the entire grid of agents plus
one extra trader (see description of figure C3). Green lines correspond to average overshoots computed on the
downward trend and red lines to the average overshoots captured on the upward trend. The black line is the
average of both lines. Scaling law coefficients C, E and the Pearson product-moment correlation correspond to
the average (black) line.
26
Electroniccopy
available at:
2.5 −1.9
Buy
2.4 −2.0
2.3 −2.1
Exceed Buy
Exceed Sell
2.2 −2.2
2.1 −2.3
2.0 −2.4
Sell
1.9 −2.5
0 20 40 60 80 100 120 140
Grid size, L
Figure C5. Positive (Exceed Buy) and negative (Exceed Sell) average net volumes as the function of the grid
size. Average total net volume is equal to zero.
27
Electroniccopy
available at:
References
Aloud, M., Fasli, M., Tsang, E., Dupuis, A. and Olsen, R., Modeling the High-Frequency FX Market: An
Agent-Based Approach. Computational Intelligence, 2017, 33, 771–825.
Andriani, P. and McKelvey, B., Beyond Gaussian averages: redirecting international business and manage-
ment research toward extreme events and power laws. Journal of International Business Studies, 2007,
38, 1212–1230.
Antypas, A., Koundouri, P. and Kourogenis, N., Aggregational Gaussianity and barely infinite variance in
financial returns. Journal of Empirical Finance, 2013, 20, 102–108.
Arneodo, A., Bouchaud, J.P., Cont, R., Muzy, J.F., Potters, M. and Sornette, D., Comment on“Turbulent
cascades in foreign exchange markets”. arXiv preprint cond-mat/9607120, 1996.
Basu, S., Investment performance of common stocks in relation to their price-earnings ratios: A test of the
efficient market hypothesis. The journal of Finance, 1977, 32, 663–682.
Bauwens, L., Omrane, W.B. and Giot, P., News announcements, market activity and volatility in the
euro/dollar foreign exchange market. Journal of International Money and Finance, 2005, 24, 1108–
1125.
Black, F. and Scholes, M., The pricing of options and corporate liabilities. Journal of political economy,
1973, 81, 637–654.
Blattberg, R.C. and Gonedes, N.J., A comparison of the stable and student distributions as statistical models
for stock prices. The journal of business, 1974, 47, 244–280.
Bollerslev, T. and Melvin, M., Bidask spreads and volatility in the foreign exchange market: An empirical
analysis. Journal of International Economics, 1994, 36, 355–372.
Bouchaud, J.P., Power laws in economics and finance: some ideas from physics. , 2001.
Bouchaud, J.P., Price impact. Encyclopedia of quantitative finance, 2010.
Brown, S.J. and Warner, J.B., Using daily stock returns: The case of event studies. Journal of financial
economics, 1985, 14, 3–31.
Chakraborti, A., Toke, I.M., Patriarca, M. and Abergel, F., Econophysics: Empirical facts and agent-based
models. arXiv preprint arXiv:0909.1974, 2009.
Cheung, Y.W. and Chinn, M.D., Currency traders and exchange rate dynamics: a survey of the US market.
Journal of international money and finance, 2001, 20, 439–471.
Chordia, T., Roll, R. and Subrahmanyam, A., Market liquidity and trading activity. The journal of finance,
2001, 56, 501–530.
Christie, A.A., The stochastic behavior of common stock variances: Value, leverage and interest rate effects.
Journal of financial Economics, 1982, 10, 407–432.
Cohen, L., Diether, K.B. and Malloy, C.J., Supply and demand shifts in the shorting market. The Journal
of Finance, 2007, 62, 2061–2096.
Cont, R., Empirical properties of asset returns: stylized facts and statistical issues. , 2001.
Cont, R., Potters, M. and Bouchaud, J.P., Scaling in stock market data: stable laws and beyond. Scale
invariance and beyond, 1997, 7, 75–85.
Dacorogna, M.M., Müller, U.A., Pictet, O.V. and De Vries, C.G., Extremal forex returns in extremely large
data sets. Extremes, 2001, 4, 105–127.
Danielsson, J. and de Vries, C.G., Tail index and quantile estimation with very high frequency data. Journal
of empirical Finance, 1997, 4, 241–257.
Debreu, G., Theory of value: An axiomatic analysis of economic equilibrium, 17, 1987, Yale University Press.
Delage, V., Brandlhuber, C., Tuyls, K. and Weiss, G., Multi-Agent based simulation of FOREX exchange
market. Maastricht University, Department of Knowledge Engineering, the Netherlands Teramark Tech-
nologies GmbH, Munich, Germany, 2010.
Di Matteo, T., Multi-scaling in finance. Quantitative finance, 2007, 7, 21–36.
Ederington, L.H. and Lee, J.H., How markets process information: News releases and volatility. The Journal
of Finance, 1993, 48, 1161–1191.
Ehrentreich, N., Agent-based modeling: The Santa Fe Institute artificial stock market model revisited, Vol.
602, , 2007, Springer Science & Business Media.
Elder, A., The new trading for a living: psychology, discipline, trading tools and systems, risk control, trade
management, 2014, John Wiley & Sons.
Fama, E.F., Efficient capital markets: A review of theory and empirical work. The journal of Finance, 1970,
25, 383–417.
28
Electroniccopy
available at:
Farmer, J.D., Lillo, F. et al., On the origin of power-law tails in price fluctuations. Quantitative Finance,
2004, 4, 7–11.
Frank, M.Z. and Sanati, A., How does the stock market absorb shocks?. Journal of Financial Economics,
2018, 129, 136 – 153.
Gabaix, X., Gopikrishnan, P., Plerou, V. and Stanley, H.E., A theory of power-law distributions in financial
market fluctuations. Nature, 2003, 423, 267–270.
Gençay, R., Dacorogna, M., Muller, U.A., Pictet, O. and Olsen, R., An introduction to high-frequency finance,
2001, Academic press.
Ghashghaie, S., Breymann, W., Peinke, J., Talkner, P., Dodge, Y. et al., Turbulent cascades in foreign
exchange markets. Nature, 1996, 381, 767–770.
Glattfelder, J., Dupuis, A. and Olsen, R., Patterns in high-frequency FX data: discovery of 12 empirical
scaling laws. Quantitative Finance, 2011, 11, 599–614.
Golub, A., Chliamovitch, G., Dupuis, A. and Chopard, B., Multi-scale representation of high frequency
market liquidity. Available at SSRN 2393428, 2014.
Golub, A., Glattfelder, J. and Olsen, R.B., The Alpha Engine: Designing an Automated Trading Algorithm.
, 2017.
Guillaume, D.M., Dacorogna, M.M., Davé, R.R., Müller, U.A., Olsen, R.B. and Pictet, O.V., From the
bird’s eye to the microscope: A survey of new stylized facts of the intra-daily foreign exchange markets.
Finance and stochastics, 1997, 1, 95–129.
Hasbrouck, J., Measuring the information content of stock trades. The Journal of Finance, 1991, 46, 179–
207.
Hausman, J.A., Lo, A.W. and MacKinlay, A.C., An ordered probit analysis of transaction stock prices.
Journal of financial economics, 1992, 31, 319–379.
Huberman, G. and Stanzl, W., Price Manipulation and Quasi-Arbitrage. Econometrica, 2004, 72, 1247–1275.
Joulin, A., Lefevre, A., Grunberg, D. and Bouchaud, J.P., Stock price jumps: news and volume play a minor
role. arXiv preprint arXiv:0803.1769, 2008.
Kaldor, N., Capital accumulation and economic growth. In The theory of capital, pp. 177–222, 1961, Springer.
Kempf, A. and Korn, O., Market depth and order size. Journal of Financial Markets, 1999, 2, 29–48.
Kozhan, R. and Salmon, M., The information content of a limit order book: The case of an FX market.
Journal of Financial Markets, 2012, 15, 1–28.
Kullmann, L., Töyli, J., Kertesz, J., Kanto, A. and Kaski, K., Characteristic times in stock market indices.
Physica A: Statistical Mechanics and its Applications, 1999, 269, 98–110.
Kyle, A.S., Continuous auctions and insider trading. Econometrica: Journal of the Econometric Society,
1985, pp. 1315–1335.
LeBaron, B., Agent-based computational finance. Handbook of computational economics, 2006, 2, 1187–1233.
Lillo, F., Farmer, J.D. and Mantegna, R.N., Econophysics: Master curve for price-impact function. Nature,
2003, 421, 129.
Mandelbrot, B. and Taylor, H.M., On the distribution of stock price differences. Operations research, 1967,
15, 1057–1062.
Mandelbrot, B.B. and Hudson, R.L., The (mis) behaviour of markets: a fractal view of risk, ruin and reward,
2010, Profile books.
Mantegna, R.N. and Stanley, H.E., Scaling behaviour in the dynamics of an economic index. Nature, 1995,
376, 46–49.
Müller, U.A., Dacorogna, M.M., Davé, R.D., Olsen, R.B., Pictet, O.V. and von Weizsäcker, J.E., Volatil-
ities of different time resolutionsanalyzing the dynamics of market components. Journal of Empirical
Finance, 1997, 4, 213–239.
Müller, U.A., Dacorogna, M.M., Olsen, R.B., Pictet, O.V., Schwarz, M. and Morgenegg, C., Statistical
study of foreign exchange rates, empirical evidence of a price change scaling law, and intraday analysis.
Journal of Banking & Finance, 1990, 14, 1189–1208.
Muth, J.F., Rational expectations and the theory of price movements. Econometrica: Journal of the Econo-
metric Society, 1961, pp. 315–335.
Naciri, N. and Tkiouat, M., Economic Agent Based Models. International Journal of Applied Engineering
Research, 2016, 11, 5492–5502.
Pagan, A., The econometrics of financial markets. Journal of empirical finance, 1996, 3, 15–102.
Petrov, V., Golub, A. and Olsen, R., Instantaneous Volatility Estimator Based on Directional-Change In-
trinsic Time [online]. , 2018. (accessed Nov. 2018).
29
Electroniccopy
available at:
Samanidou, E., Zschischang, E., Stauffer, D. and Lux, T., Agent-based models of financial markets. Reports
on Progress in Physics, 2007, 70, 409.
Schwert, G.W., Why does stock market volatility change over time?. The journal of finance, 1989, 44,
1115–1153.
Scott, L.O., Option pricing when the variance changes randomly: Theory, estimation, and an application.
Journal of Financial and Quantitative analysis, 1987, 22, 419–438.
Walras, L., Elements of pure economics, 2013, Routledge.
Zhang, L., Mykland, P.A. and Aı̈t-Sahalia, Y., A tale of two time scales: Determining integrated volatility
with noisy high-frequency data. Journal of the American Statistical Association, 2005, 100, 1394–1411.
Zhou, B., High-frequency data and volatility in foreign-exchange rates. Journal of Business & Economic
Statistics, 1996, 14, 45–52.
Funding details
This project has received funding from the European Union’s Horizon 2020 research and innovation
programme under the Marie Skłodowska-Curie grant agreement No 675044
30
Electroniccopy
available at:

Agent-Based Model in Directional-Change Intrinsic Time

Uploaded by

Copyright:

Available Formats

Agent-Based Model in Directional-Change Intrinsic Time

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Agent-Based Model in Directional-Change Intrinsic Time

Uploaded by

Copyright:

Available Formats

Agent-Based Modelling in Directional-Change

V. Petrov∗ †, A. Golub‡ and R. Olsen§

†University of Zurich, 15 Andreasstrasse, 8050 Zurich, Switzerland

∗ Corresponding author. Email: [email protected]

Pf lip = e−1 = P(ω(δ) ≥ δ). (3)

Random Return b) c) Generated Return

4.1. Traditional benchmarks

E[(Mt − µ)(Mt+τ − µ)]

h(r(t, τ ) − hr(t, τ )i)4 i

4.2. The ultimate benchmark

where y = exp Y , x = exp X, E = B and C = exp(−A/B).

5.1. The entire grid

10−20 10−11 10−3

10−6 10−4 10−2

5.2. Asymmetric regions

dSt = St − St−1 = µdt + σdWt . (A2)

Using (A5) one can find that:

E[ω(δup , µ, σ)] = n o, (A6)

Appendix B: Dissection Algorithm

Appendix C: Additional experiments

C = 0.88 C = 5.66 C = 2.56 C = 1.52

C = 5.42 C = 0.94 C = 1.25 C = 1.22

δup = 3, δdown = 1 δup = 3, δdown = 2 δup = 3, δdown = 3 δup = 3, δdown = 4

C = 2.65 C = 1.2 C = 0.97 C = 1.05

C = 1.6 C = 1.2 C = 1.07 C = 1.02

You might also like