Physica A 525 (2019) 171–191

Contents lists available at ScienceDirect

Physica A
journal homepage: www.elsevier.com/locate/physa

Order book model with herd behavior exhibiting long-range

highlights

• The model is based on empirical order book model and financial herd behavior model.
• The model is compared against high frequency data from Bitcoin exchanges and NYSE.
• Fracture in PSD can be explained by the convergence towards the Walras equilibrium.

article info a b s t r a c t

Article history: In this work, we propose an order book model with herd behavior. The proposed model
Received 8 September 2018 is built upon two distinct approaches: a recent empirical study of the detailed order
Received in revised form 15 January 2019 book records by Kanazawa et al. (2018) and financial herd behavior model. Combining
Available online 28 March 2019
these approaches allows us to propose a model that replicates the long-range memory of
Keywords: absolute return and trading activity. We compare the statistical properties of the model
Financial markets against the empirical statistical properties of the Bitcoin exchange rates and New York
Order book stock exchange tickers. We also show that the fracture in the spectral density of the high-
Long-range memory frequency absolute return time series might be related to the mechanism of convergence
Bitcoin towards the equilibrium price.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction

In the recent decades an increasing effort by social scientists, physicists and broader interdisciplinary community has
been applied to create agent-based models (ABMs) of different social phenomena [1–8]. Notable part of these ABMs were
created to explain various recurrent anomalous statistical patterns, collectively referred to as stylized facts, observed in
the financial markets [2–4,7]. Financial ABMs vary in complexity usually trading plausibility for analytical tractability [4].
Some of the more complex financial ABMs seem to be reasonably plausible, but they are not analytically tractable. A well
known example of such model is the Lux–Marchesi model [9]. Hence it significantly harder to understand their dynamics
and to build upon them to further improve their agreement with the empirical data.
In the recent few years we have developed a reasonably plausible yet analytically tractable financial ABM [10,11].
This financial ABM was derived from a widely recognized behavioral model [12], which emphasizes imitation (herd)
behavior among socially interacting individuals. This financial ABM is able to rather precisely fit the probability density
functions (PDFs) and power spectral densities (PSDs) of the empirical absolute return time series. While the model has
other desirable features, e.g., it scales well with change in the time scale, it also has some drawbacks. First of all it does
not implement realistic trading strategies, though most of the financial ABMs lack this feature [4]. But it is not the main

∗ Corresponding author.
E-mail address: [email protected] (A. Kononovicius).

https://doi.org/10.1016/j.physa.2019.03.059
0378-4371/© 2019 Elsevier B.V. All rights reserved.
172 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191

Table 1
Comparison of the approaches reviewed in Section 2. Differences between our approach and the reviewed
approaches are highlighted using italics.
Approach Agent types Fixed/flexible Strategies Volume Tractable
[27] Chartists, fundamentalists Fixed Sound Strategic Partly analytically
[28] Chartists, fundamentalists Fixed Simple Unit Only numerically
[29–31] Chartists, fundamentalists Fixed Sound Random Only numerically
[32] Chartists, fundamentalists Fixed Realistic Unit Only numerically
[33,34] Chartists, random traders Fixed Simple Random only numerically
[35] Chartists, fundamentalists Fixed Realistic Strategic Only numerically
Our Chartists, fundamentalists Flexible Simple Unit Possibly analytically

issue we currently see. Our financial ABM relies on a two assumptions, which are not fully and transparently justified. We
have assumed the presence of the omnipotent market maker, who is able to clear the market instantaneously, (a rather
common assumption in many financial ABMs [2–4]) and the presence of the exogenous noise. In our previous papers we
have speculated that exogenous noise is likely to originate from the order book dynamics. Consequently it seems that
including order book dynamics in our financial ABM could potentially resolve these two issues at once. Such approach
has an additional perk, it would allow us to consider trading activity time series as well.
Though the empirical order book data has become available at the same time as the empirical high-frequency financial
time series data, in the 1980’s, it took a longer time for detailed empirical studies of the order book dynamics to be
undertaken [13–16]. Interestingly some of the observations were not as universal as the stylized facts discovered in the
empirical time series. There are few recent empirical order book studies, which confirm some of the earlier findings,
but fail to confirm the other findings [17]. Similarly there is a variety of simple limit order book models (e.g., [18–20]),
which are not mutually compatible, but are able to reproduce some of the empirical observations in the order book
data. Some of the more recent order book modeling approaches, such as [21–24], are more sophisticated and able to
reproduce a variety of empirical observations. The most recent approach by Kanazawa et al. [25,26] combines empirical
and theoretical modeling approaches. Namely, Kanazawa et al. have observed the behavior of the high-frequency traders
in a highly detailed order book level data set. Based on the observations a microscopic model inspired by the kinetic
theory was proposed. Nevertheless most of these models have not considered a detailed reproduction of the stylized
facts established for the time series data.
In this paper our goal is to introduce the order book mechanics, similar to the ones introduced in [25,26], into our
financial ABM [10,11] thus producing a novel order book model with herd behavior, which would be able to reproduce
power law statistical properties of the absolute return and trading activity time series. In Section 2 we will compare
the order book modeling approach we will take here against a few similar recently published approaches [27–35]. In
Section 3 we will briefly introduce herd behavior model proposed by Kirman [12], which is the basis of our agent-based
approach. In Section 4 we will discuss two different financial market models: the previously proposed herd behavior
model with instantaneous clearing [10,11] and the order book model with herd behavior. We will show that under certain
parameter values both of the models produce mostly identical statistical properties. Further in Section 5 we will compare
the order book model with herd behavior against the empirical high-frequency Bitcoin and NYSE data. We will examine
the sensitivity of the model to parameter value changes in Section 6. While we will provide concluding remarks and future
outlooks in Section 7.

2. Review of the other similar order book modeling approaches

Recently a few approaches similar to the one we will take here were published [27–35]. While these approaches
are different in many aspects, we will keep our review brief by highlighting only the most important similarities and
differences between these approaches themselves and our approach to be taken in the next sections of this paper. The
review is summarized in Table 1.
First of all it is important to note that all of the models under our review are able to reproduce the main stylized facts
of absolute return time series. Some of the models are also able to reproduce the features of price time series [27] or
specific peculiarities observed in the modeled markets [33,34]. Neither of the considered approaches reports results on
the reproducibility of the statistical properties of trading activity. Our goal will be to reproduce PDFs and PSDs of both
absolute return and trading activity time series. Note that some of the approaches reviewed here report auto-correlation
functions of the return time series, while we prefer to use PSDs. This makes no significant difference as PSD is directly
related to the auto-correlation function via the Wiener–Khinchin theorem.
All of the considered models assume that there are two types of traders. One type of traders relies on the internal
market information. These traders are usually called chartists. Another type of traders use exogenous information. In
most of the works exogenous information is some kind of parametrization of the fundamental price, hence such traders
are called chartists. Among the considered works only Cocco et al. [33,34] make a different assumption. Namely, Cocco
et al. assume that exogenous information is related to the objective needs of traders, which are unknown and random,
 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191 173

therefore it causes random trading. Our approach is similar to the most of these works, we will assume that the market
is populated by the two types of agents: chartists and fundamentalists.
Unlike in these models in our approach agents are flexible, namely they are able to switch their trading strategies. It is
quite a puzzle to explain why the models with fixed and flexible agent types are able to reproduce the main stylized facts.
Most likely it is because these models rely on the heterogeneity of the agents within each type. Whenever this intra-type
heterogeneity breaks down, the agents act similarly on the received information, and thus large fluctuations emerge.
In our approach large fluctuations also emerge when most agents become chartists, i.e., when inter-type heterogeneity
breaks down. In all cases fluctuations will be the largest under heterogeneity breaking.
The approaches also differ on how the trading strategies are specified. Most realistic specification can be found
in [32,35] as the model proposed in these works utilize technical trading strategies which are actually used for trading
in the real markets. A bit less practical, yet sound from an economic perspective, strategies can be found in [27,29–31].
These strategies are economically sound because they allow the agents to maximize their expected wealth, though these
strategies are not necessarily common among the traders in the real markets. While our approach, and also [28,33,34],
is the simplest in this context. Namely, trading strategies in this case are assumed to be highly stylized and do not
necessarily guarantee wealth gain for the agents. In some sense it could be said that the agents under this approach
have zero intelligence.
We could categorize the approaches to the trading strategies from another point of view. The strategies also
significantly differ on how the order volumes are treated. We find three different possible treatments: order volume
is selected strategically in [27,35] (it is important to note that neither of these models uses order books), order volume is
drawn from random distribution in [29–31,33,34] or it is set to be of unit size [28,32]. In [25,26], on which we will based
our order book approach, order volumes are also assumed to be of unit size, this assumption is backed up by empirical
analysis, which shows that around 80% of completed transactions fill a unit of volume.
Although the differences in respect to trading strategies can be extremely important when trying to build a model
for economic policy making [29–32,35], it does not help to make model more tractable, which is also a rather desirable
feature. All of these models, with the exception of [27], are not analytically tractable and can be studied only via numerical
simulation. Thus only limited knowledge about their dynamics can be obtained. Though our financial ABM seems to be
comparatively lacking in economic plausibility it makes up by being analytically tractable, though one has to rely on
Stochastic Calculus for that. As the model in [27] our financial ABM lacks order book clearing mechanism, but in this
work we will introduce that in to our financial ABM. As the order book modeling approach we will take in this work is
analytically tractable [25,26], we can expect that the combination of the both approaches would also become analytically
tractable at some point.

3. Kirman’s herd behavior model

Let us start with discussion about Kirman’s herd behavior model [12], which is the base upon we build financial market
models in the following sections. In the seminal paper Kirman shared an observation that social scientists and behavioral
biologists observe remarkably similar patterns in rather distinct systems. In the experiments involving ants described by
Kirman, entomologists observed the emergence of asymmetry in a symmetric experimental setup: despite having two
identical food sources available, majority of the ants in the ant colony preferred to forage from a single food source at
a time. Numerous references in Kirman’s paper suggest that humans also seem to prefer the more popular product over
the less popular despite both being of a similar quality.
To account for these empirical observations Kirman proposed a simple probabilistic herd behavior model in which the
probability for an agent to switch to another state is proportional to the fraction of agents in that state:
( )
X
p (X → X + 1) = (N − X ) σ1 + h ∆t , (1)
N
[ ]
N −X
p (X → X − 1) = X σ2 + h ∆t , (2)
N
where X is a number of agents in the first state, N – a total number of agents, σi – an idiosyncratic behavior parameter
(encodes preferences for the states), h herd behavior parameter, ∆t – arbitrarily small time step. This formulation of the
herd behavior model is often referred to as ‘‘local’’ or ‘‘extensive’’, because the fluctuations of X quickly disappear as N
becomes larger [36–39]. In other words X rapidly converges to a certain value and remains almost constant afterwards.
What we described above in the literature is often referred to as the N-dependence problem [36–38]. This problem
can be circumvented by assuming that the probability to switch is proportional to a total number of agents X :

p (X → X + 1) = (N − X ) (σ1 + hX ) ∆t , (3)
p (X → X − 1) = X [σ2 + h (N − X )] ∆t , (4)
To contrast the previous formulation of the model, this formulation of the herd behavior model is often referred to as
the ‘‘non-extensive’’ or ‘‘global’’ formulation. In this formulation X no longer converges to a fixed value even in the limit
of infinite N. This is desirable feature to have in the financial market models as it is well-known that the stylized facts
174 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191

hold for small and large markets alike [40]. Hence there is a variety of the financial ABMs, which were inspired by the
non-extensive formulation of the Kirman’s model [9–11,41–44]. There are also papers in the opinion dynamics, which
claim that the fluctuating nature of opinion change can be explained by assuming the presence of collective peer-pressure
instead of inter-personal communication [37–39,45,46].
Let us take the infinite N limit and introduce an almost continuous state variable x = NX . This allows us to rewrite the
model driven by Eqs. (3) and (4) as a stochastic differential equation [10,41,42]:
√
dx = h [ε1 (1 − x) − ε2 x] dt + 2hx (1 − x)dW , (5)
σi
where εi = h . From the Eq. (5) it is straightforward to conclude that x is Beta distributed, x ∼ Beta (ε1 , ε2 ). If we would
set ε1 = ε2 and ε1 < 1 then we would observe the same pattern entomologists did as the Beta distribution is multi modal
in this case.
One can derive a similar SDE, by taking finitely large N limit under extensive formulation of the model. Yet in this
case SDE becomes ordinary differential equation as with larger N the diffusion term becomes negligible and only the drift
term remains:
√
2hx (1 − x)
dx = h [ε1 (1 − x) − ε2 x] dt + dW
N
≈ h [ε1 (1 − x) − ε2 x] dt . (6)

In a special case, p (X → X − 1) = 0, one can also derive the well-known Bass diffusion equation [47]:

dx = (1 − x) (σ1 + hx) dt . (7)

4. Herd behavior model in the context of the financial markets

Kirman’s herd behavior model is in a sense a generic model. In the financial market context it would be natural for these
states to represent different trading strategies. In the agent-based modeling literature one can find the most common is
to consider interaction between agents using fundamentalist and chartist trading strategies [4,6,35]. It is worth to note
a couple approaches which consider modeling market sentiments instead [42,48,49]. In our previous works [10,11] we
have used both of these approaches to build a highly sophisticated ABM which was able to fit the empirical absolute
return PDF and PSD rather well. Here we will start with the fundamentalist-chartist model under instantaneous clearing
and later use it to build order book model.

4.1. Herd behavior model with instantaneous clearing

Let us define the trading strategies in a rather stylized manner. One could in general define more sophisticated
strategies (as discussed in [35]), but we would like to keep the model compatible with our earlier works. This will allow
us to use some of the earlier analytically obtained results which would be impossible while using a more realistic trading
strategies.
In [10] we have assumed that excess demand by chartist traders is conditioned on their mood ξ (t ):

Dc (t ) = r0 Nc (t ) ξ (t ) , (8)

where r0 is the relative impact of chartists’ trading activity and Nc (t ) is the number of chartists. In contrast fundamental-
ists’ demand is conditioned on their knowledge about the market fundamentals, which is quantified as the fundamental
price Pf (which has a physical dimension of generic price units, p.u.):
Pf Pf
Df (t ) = Nf (t ) ln = [N − Nc (t )] ln , (9)
P (t ) P (t )
where Nf (t ) is the number of fundamentalists and P (t ) is the current price. Note that here we assume that the
fundamental price is fixed, which is not true for the real markets. Nevertheless adding variability to the fundamental
price, e.g., by assuming that it follows Brownian motion, would not have a significant impact on the statistical properties
of the model.
We have also assumed that a market maker instantaneously clears the market by setting the price to the Walras
equilibrium price, which is obtained in the following manner:
Nc (t )
( )
Df (t ) + Dc (t ) = 0 ⇒ Peq (t) = Pf exp r0 · · ξ (t ) . (10)
N − Nc (t )
If ξ (t ) fluctuates significantly faster than Nc (t ), then the absolute return
P (t )
⏐ ∝ Nc (t ) = y.
⏐ ⏐
|r (t )| = ⏐⏐ln
⏐ ⏐
(11)
P (t − T ) ⏐ N − Nc (t )
 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191 175

Through out our papers we have referred to y as modulating return as it describes longer-term fluctuations of the absolute
return, while ξ (t ) dictates the rapid fluctuations and changes in the sign of the return. In some of the earlier works ξ (t )
was even modeled as a noise [41].
Previously [10] we have also extended the original herd behavior model by introducing the feedback of the modulating
return y on the switching dynamics:
] ∆t
p (Nc → Nc + 1) = (N − Nc ) σfc + hNc ,
[
(12)
τ (Nc )
] ∆t
p (Nc → Nc − 1) = Nc σcf + h (N − Nc ) ,
[
(13)
τ (Nc )
where
( )−α
Nc
τ (Nc ) = ≡ y−α . (14)
N − Nc
That is, τ (Nc ) adjusts the characteristic time scale of microscopic switching events according to the current global value
of the modulating return. Such feedback scenario implements the coupling between returns and trading activity, which is
well established empirical fact [50]. Note that in [10] switching dynamics were assumed to correlate with trading activity.
Introduction of the feedback scenario enables us to obtain a more general form of the SDE for y, which has tunable
noise multiplicativity exponent:
√
) ] 1+y 2hy
dy = h εfc + 2 − εcf y (1 + y) dW
[ (
dt +
τ (y) τ (y)
= h εfc + 2 − εcf y (1 + y) yα dt + 2hy1+α (1 + y) dW
[ ( ) ] √
(15)
√
≈ h 2 − εcf y2+α dt + 2hy3+α dW .
( )

This SDE, assuming y ≫ 1, belongs to a class of SDEs exhibiting power law statistics described in [51]. Thus it the y
time series should exhibit power law statistics [10]:
εcf +α−2
P (y) ∼ y−εcf −α−1 , S (f ) ∼ f .
−1− 1+α (16)

This simple model already reproduces two main stylized facts. In the later papers, e.g., [11], we have extended this
model by describing the mood dynamics using the same herd behavior model. Though in order to fit the empirical absolute
return PDF and PSD we had to introduce exogenous noise, which we assumed to represent additional randomness arising
from the order book dynamics and possibly an exogenous information inflow. In the next section we build the order book
model to address this assumption.

4.2. Order book model with herd behavior

Most of the ABMs, which consider statistical properties of the various financial time series, directly or indirectly assume
presence of the market maker [4]. While the real financial markets are not cleared by an idealized market maker, most of
the contemporary financial markets implement trading by using the order books. Similarly to the market makers order
books record and execute orders that the traders submit. The difference is that the orders in the order book are executed
only if there is an overlap between the buy (bid) and the ask (sell) sides of the order book or if a market order is submitted.
While there is a significant body of literature considering order book modeling [19–26,28], most of these models consider
reproducing patterns observed at the order book level and usually neglect stylized facts related to the financial time series.
It is worth to note that there are numerous papers in economics, which suggest that the prices in various auctions converge
towards Walras equilibrium and that this convergence might be comparatively fast [52–57]. Yet this convergence is not
instantaneous and we might observe some interesting effects in the high-frequency financial time series. Here, while
building our order book model, we will partly rely on an empirically motivated order book model proposed by Kanazawa
et al. [25,26].
As in [25,26] we assume that chartists as high-frequency traders submit unit volume limit orders to the both sides of
the order book. The submitted quotes, Qiask and Qibid , are placed by the ith agent the same distance, Si , away from the
current valuation of the stock, Vi :

Qiask (t ) = Vi (t ) + Si , (17)
Qibid (t ) = Vi (t ) − Si , (18)
Si ∼ Gamma (k, θ) , (19)

where k is the shape parameter of the Gamma distribution and θ is the scale parameter of the Gamma distribution. Here
physical dimension of θ is p.u. (generic price units) as it is also the dimension of Qi , Vi and Si . Further in this paper we
176 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191

Fig. 1. Full order book profile (a) and a zoom in to 5 best quotes present in the order book (b). Blue curve (a) and circles (b) denote bid (buy) limit
orders, red curve (a) and circles (b) denote ask (sell) limit orders, while black circles (b) show Vi for each of the best quoters. In sub figure (a) ask
profile is cumulative, giving number of limit orders willing to sell for a given price or larger, while bid profile is inversely cumulative, giving the
number of limit orders willing to buy at a given price or smaller. In sub figure (b) the arrow shows where the valuation would move if market ask
order would be submitted.

will use the empirically determined values k = 4 and θ = 15.5 p.u. (see [25,26]) unless specified otherwise. Note that
these values were obtained specifically for USD/JPY exchange rate in Forex and they might take different values for the
other markets or exchange rates. Yet we will rely on these values as best available estimate at this point.
Next we will replace the sophisticated trend following mechanism originally present in [25,26] with a simpler market
order submission mechanism, which is similar to the one used in the herd behavior model with instantaneous clearing.
This simplification leads to another simplifying assumption that valuations are homogeneous Vi (t ) = V (t ). Hence the
valuation changes only after a market order is executed. After a market order is executed, the valuation is set to the value
of realized quote. As in the original order book approach, after the valuation is reset chartists update their submitted
quotes.
An example of the full order book profile is shown in Fig. 1(a). Sub figure (b) of Fig. 1 provides zoomed in picture of
the 5 best quotes, and shows the movement of valuation if market ask order would be submitted. Blue and red circles
represent bid and ask quotes respectively, these circles are connected to show that the both quotes are submitted by
the same agent, whose valuation is represented by the black circle. Note that under our simplifications the profile is
symmetric around quote equal to the current valuation, which is identical for all agents. This would not be the case for
the original approach in [25,26].
Now let discuss the replacement of the trend following mechanism of [25,26] with a simpler one used in the herd
behavior model with instantaneous clearing. Let the chartists submit unit volume market orders at rate
λe
λtC (t ) = λtc Nc (t ) , (20)
τ (Nc (t ))
where λe is the reference event rate (which has a physical dimension of s−1 ) and λtc – the relative market order submission
rate by a single chartist agent. The submitted market order is bid order with probability:
1 + ξ (t )
pbid (t ) = , (21)
2
and ask market order is submitted otherwise, pask (t ) = 1 − pbid (t ). Let us assume that the mood simply flips its sign at
rate
λe
λM (t ) = λm , (22)
τ (Nc (t ))
where λm is relative mood flipping rate. As the sign flip does not change the modulus the mood will take only two possible
values, ξ (t ) ∈ {−ξ0 , +ξ0 } (here ξ0 is initial value of the mood).
Fundamentalists are not present in [25,26], so we can keep our earlier assumptions about their behavior. Namely, we
assume that fundamentalists are willing to buy stock if there is such ask quote for which Qjask (t ) < Pf , and are willing
to sell stock if there is such bid quote for which Qjbid (t ) > Pf . They will submit unit volume market orders, if there are
suitable quotes in the order book, at rate:
λe P (t ) ⏐
⏐ ( )⏐
λtF (t ) = λtf [N − Nc (t )] ⏐ln ⏐,
⏐
⏐ (23)
τ (Nc (t )) Pf ⏐
where λtf is the relative market order submission rate by a single fundamentalist agent.
Note that in [25,26] power law return distribution is obtained only after taking into account the fact that the number of
high-frequency traders (chartists) changes over time. If the number of high-frequency traders (chartists) would be fixed,
 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191 177

then returns would be exponentially distributed. This intuition was confirmed empirically by splitting the time series into
two hour periods. For each of these periods returns were found to be approximately exponentially distributed, though the
different values of rates were found for the different periods. The values of rates were found to be positively correlated
with the average number of high-frequency traders (chartists) present during the same periods. Next let us include this
variation to the current model by including switching behavior present in the herd behavior model with instantaneous
clearing.
It is straightforward to determine event rates for the trading strategy switching: the fundamentalist will switch to the
chartist trading strategy at rate,
λe
λfc (t ) = [N − Nc (t )] εfc + Nc (t ) ,
[ ]
(24)
τ (Nc (t ))
while the chartist will switch to the fundamentalist trading strategy at rate,
λe
λcf (t ) = Nc (t ) εcf + {N − Nc (t )} .
[ ]
(25)
τ (Nc (t ))
Note that these transition rates do not include parameter h, which is because it is equivalent to λe . As soon as chartist
becomes fundamentalist his limit orders are canceled, also if fundamentalist becomes chartist, then he immediately
submits his limit orders.
( )number of agents in this model will always be finite, the probability of τ (Nc ) = 0 (or alternatively Nc (t ) = N)
As the
and τ Nf = ∞ (or alternatively Nc (t ) = 0) will be non-zero, which would lead to ‘‘over-heating’’ or ‘‘freezing’’ of the
strategy switching dynamics. To avoid these edge cases let us redefine the feedback scenario as:
[2Nc (t )]α
⎨if Nc (t ) = N [
⎧
]α
τ (Nc (t )) = λ0 +
−1
Nc (t ) (26)
⎩else ,
N − Nc (t )
where λ0 is the relative minimum switching rate. In the above we have multiplied Nc by 2 when taking the Nc = N edge
case into account, because the previous increase in y, number of chartists increasing from Nc = N − 2 to Nc = N − 1, is
approximately double given large N.
We use the Gillespie algorithm [58,59] to implement the order book model. The main idea behind the Gillespie
algorithm is that we can sum all of the event rates to obtain the total event rate:
λT = λcf + λfc + λM + λtF + λtC , (27)
which enables us to generate random inter-event times, which are distributed exponentially:
∆ti ∼ Exp λT .
( )
(28)
After each ∆ti one of the possible events happens. The probability for any of the possible events to happen is proportional
to its rate:
λcf λfc λM λtF λtC
pcf = , pfc = , pM = , ptF = , ptC = . (29)
λT λT λT λT λT
As these probabilities sum to 1, one of the five possible events is bound to happen: either randomly selected chartist
switches to fundamentalist trading strategy (with probability pcf ) or randomly selected fundamentalist switches to
chartist trading strategy (with probability pfc ) or the mood flips its sign (with probability pM ) or the randomly selected
fundamentalist submits market order (with probability ptF ) and randomly selected chartist submits market order (with
probability ptC ).
The exact algorithm behind this model is summarized as a flowchart in Fig. 2. The code implementing this model is
publicly available on https://github.com/akononovicius/herding-OB-model. All of the parameters used in this model are
summarized in Table 2.

4.3. Comparison between the models

It is possible to approximately estimate equilibrium prices for the order book model. From the discussion in the
previous section we know that chartists submit unit volume market orders at rate λtC , with probability pbid they buy
the stock and with probability 1 − pbid they sell the stock. Hence their excess demand (on average) is given by:
λe
D̄c = λtC pbid − λtC (1 − pbid ) = λtc Nc (t ) ξ (t ) . (30)
τ (Nc (t ))
Note that the final result is similar to Eq. (8). Fundamentalists on the other hand submit unit volume market orders at rate
λtF , they submit ask orders if there is such j for which Qjbid (t ) > Pf and bid orders if there is such j for which Qjask (t ) < Pf .
Assuming that Pf is not in the spread, the excess demand of fundamentalists (on average) will be given by:
λe P (t )
( )
D̄f = λtf [N − Nc (t )] ln . (31)
τ (Nc (t )) Pf
178 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191

Fig. 2. Flowchart illustrating the order book model with herd behavior.

Table 2
List of parameters used in the order book model with herd behavior.
Symbol Meaning
N Number of agents
λe Reference event rate (has physical dimension of 1/s)
εcf Relative idiosyncratic switching rate from chartists to fundamentalists
εfc Relative idiosyncratic switching rate from fundamentalists to chartists
ξ0 Absolute value of chartists’ mood
λm Relative rate at which mood flips its sign
λ0 Relative minimum switching rate between chartists and fundamentalists
α Exponent of the feedback scenario
k Order book shape parameter
θ Order book scale parameter (has physical dimension of generic price units, p.u.)
Pf Fundamental price (has physical dimension of generic price units, p.u.)
λtc Relative market order submission rate for chartists
λtf Relative market order submission rate for fundamentalists

Note that the final result is similar to Eq. (9). Assuming that order book is non-empty and almost uniformly filled we can

obtain the equilibrium price:

λtc Nc (t )
( )
D̄c + D̄f = 0 ⇒ Peq (t) = Pf exp · · ξ (t ) . (32)
λtf N − Nc (t )
 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191 179

Fig. 3. Comparison between the statistical properties, (a) PDFs and (b) PSDs, of y time series (gray curves) and absolute return time series (red,
blue and green curves), black curves show the expected slopes of the statistical properties, as per Eq. (16). For the best comparison all of the time
series were normalized to unit standard deviation. The following parameter values were used in numerical simulations: N = 500, λe = 10−7 s−1 ,
εcf = εfc = 1, ξ0 = 0.2, λm = 103 , λ0 = 0.1, α = 1, k = 4, θ = 15.5 p.u., Pf = 3 · 104 p.u. (all cases), λtc = λtf = 3 · 104 (red curves), 300 (blue and
gray curves), 3 (green curves).

Fig. 4. Sample fragments of the absolute return (red curves) and y (gray curves) time series. Parameter values are identical to the ones used in
Fig. 3 except: (a) λtc = λtf = 3 · 104 (red curve in Fig. 3), (b) λtc = λtf = 3 (green curve in Fig. 3). The correlation coefficients between the samples
are ρ = 0.67 (a) and 0.03 (b).

As the expression for the equilibrium price has the same form as in Eq. (10), we can expect that y and return will have
similar statistical properties. The main condition we have to ensure for the similarity to be observable is that enough
trades happen between Nc changes, so that the equilibrium price could be reached. This means that λtc and λtf have to be
rather large. As you can see in Fig. 3 the agreement between the statistical properties of y and absolute return improves as
λtc and λtf grow larger. In Fig. 4 we have shown sample y and absolute return time series, which can be seen to correlate
in sub figure (a) and be almost uncorrelated in sub figure (b).
Correlation between y and absolute return time series, see sample series in Fig. 4, also is stronger with larger λtc and
λtf . For the series shown in Fig. 4(a) correlation is strong ρ ≈ 0.7 (with λtc = λtf = 3 · 104 ), while in Fig. 4(a) correlation
is negligible ρ ≈ 0 (with λtc = λtf = 3).
Similar intuition can be obtained from Fig. 5. As you can see in (a) and (c), for large λtc and λtf (parameters the same as
for the red curve in Fig. 3) the price tends to catch up with the changes in the equilibrium price. Though the following is far
from being perfect as can be seen by zooming in on the series, (c). The correlation between the price and the equilibrium
price time series is mild. While for small λtc and λtf , (b) and (d), (parameters the same as for the green curve in Fig. 3)
it is evident that the price does not manage to catch up with the changes in the equilibrium price. As expected, there is
almost no correlation between the time series.
Based on these results we would like to argue that the fracture in the PSD of the high-frequency absolute return time
series happens due to order book dynamics. Namely, the absolute return PSD in the high frequency range is less sloped,
because the markets are unable to discover the equilibrium price that fast. It seems that it could take a day or two (as
the fracture is usually between 10−5 and 10−4 Hz) for the markets to discover the new equilibrium price.

5. Comparison against the empirical data

Before making a comparison against the empirical data let us state that neither order book model presented in [25,26]
(which forms a basis of our order book approach) nor our financial ABM [10,11] (which forms a basis of our agent-based
approach) is able to reproduce all of the empirical statistical properties considered further in this Section. Our earlier
180 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191

Fig. 5. Sample fragments of the price (red curves) and the equilibrium price (gray curves) time series. Parameter values are identical to the ones
used in Fig. 3 except: (a) and (c) λtc = λtf = 3 · 104 (red curve in Fig. 3), (b) and (d) λtc = λtf = 3 (green curve in Fig. 3). The correlation coefficients
between the samples are ρ = 0.54 (a) and 0.02 (b).

Table 3
List of the considered Bitcoin time series.
Exchange Exchange pair Period available Period used Trades
bitfinex BTC/USD 2013-03-31/2016-12-22 From 2013-05-01 107 (99.4%)
bitflyer BTC/JPY 2017-07-04/2018-07-04 Whole 3 · 107
bitstamp BTC/USD 2011-09-13/2018-07-04 Except 2016-06-23 2.6 · 107 (99.9%)
btcbox BTC/JPY 2014-04-09/2018-07-04 Whole 8.9 · 106
btce BTC/USD 2011-08-14/2017-07-25 Whole 3.3 · 107
btcn BTC/CNY 2011-06-13/2017-09-30 From 2013-04-01 1.1 · 108 (99.9%)
btcoid BTC/IDR 2014-02-09/2018-07-04 Whole 8 · 106
btctrade BTC/CNY 2013-05-19/2017-09-30 Whole 2 · 107
coinbase BTC/EUR 2015-04-23/2018-07-04 Whole 1.5 · 107
coinbase BTC/USD 2014-12-01/2018-07-04 From 2017-05-01 3 · 107 (67.5%)
coincheck BTC/JPY 2014-10-31/2018-07-04 Except 2017-08-07 108 (99.9%)
kraken BTC/EUR 2014-01-08/2018-07-04 Whole 2.1 · 107
kraken BTC/USD 2014-01-07/2018-07-04 Whole 107
okcoin BTC/CNY 2013-06-12/2015-04-05 Whole 108

financial ABM [10,11] was shown to reproduce statistical properties of absolute return reasonably well, yet there was no
way to consider trading activity in that model. While [25,26] has focused on reproduction of order book dynamics only
briefly touching upon statistical properties of absolute return.
In this paper we use publicly available tick by tick trading data from 12 different Bitcoin exchanges. We have
downloaded the data from bitcoincharts.com website on July 5, 2018. List of the considered Bitcoin time series is given
in Table 3. Note that Coinbase and Kraken exchanges appear twice in the table, because they contribute more than one
exchange pair. These time series were selected, because their data files were among top 5% of the largest. Fisco’s BTC/JPY,
Zaif’s BTC/JPY and Zyado’s BTC/EUR were also among top 5% of the largest, but these time series were excluded, because
their statistical properties were too different from the rest of the time series. Note that for the same reason we have
truncated some of the Bitcoin time series which remained under our consideration.
For each of the considered Bitcoin time series we have produced one minute absolute return (normalized to the
standard deviation) and trading activity (defined as trades per time interval and normalized to the mean) time series.
For each of the produced one minute time series we have calculated PDF and PSD. The obtained statistical properties
were averaged to produce average profile for each of the statistical properties. To select the model parameters we have
 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191 181

Fig. 6. Comparison between the empirical Bitcoin statistical properties (gray curves) and statistical properties generated by the model (red curves):
(a) one minute absolute return PDF, (b) one minute absolute return PSD, (c) trading activity per one minute PDF, (d) trading activity per one minute
PSD. The following parameter set was used in numerical simulations: N = 500, λe = 10−7 s−1 , εfc = 5, εcf = 2, ξ0 = 0.2, λm = 10, λtc = 25, λtf = 75,
λ0 = 0.4, α = 2, k = 4, θ = 15.5 p.u., Pf = 3 · 104 p.u.

used simulated annealing technique with a goal to reproduce these averaged statistical properties. As you can see in Fig. 6
the obtained agreement is rather good.
Earlier works, such as the ones found in [60,61], has already carried out detailed analysis of the Bitcoin time series
and established that the stylized facts for the Bitcoin are somewhat different from the stylized facts established for the
stocks. Namely, [61] has reported that the Bitcoin returns exhibit heavier tails than ordinary stock returns. Similar finding
was also reported in [60]. Yet these works disagree on whether the tail index changes over time. In [61] it was reported
that while the tail index has slightly increased, but not enough statistical evidence was found to formally claim that the
tail index is increasing. In [60], on the other hand, a significant change in the tail index is reported. Furthermore other
statistical properties of absolute return, such as slope of the auto-correlation function, Hurst exponent and multi-scaling
properties, seem to approach values observed for the ordinary stocks. This is interpreted as a sign that Bitcoin market
is becoming ‘‘mature’’ market. In this paper we did not carry out a rigorous empirical analysis, at least as rigorous as
in [60,61]. We have just checked whether PDFs and PSDs of absolute return and trading activity change over time and
found that the change, if present, is negligible for the exchange rates that have remained under our consideration. While
the return tail index seems to be similar to the one obtained for NYSE stocks under our consideration (λ ≈ 3.7; analysis
of NYSE stocks are discussed in the next paragraph). A thorough analysis of the Bitcoin time series or a meta analysis of
the methods used in [60,61] would be due, but this topic is out of the scope of this paper.
We have also considered the statistical properties of 26 tickers from NYSE. The considered tickers include: ABT, ADM,
BA, BMY, C, CVX, DOW, FON, FNM, GE, GM, HD, IBM, JNJ, JPM, KO, LLY, MMM, MO, MOT, MRK, SLE, PFE, T, WMT and
XOM. All their time frames are from January, 2005 to March, 2007. As with the Bitcoin time series, we have obtained
averaged statistical properties for NYSE data set. Using simulated annealing technique we have obtained another best fit
parameter set for our model. To obtain a better fit we had to divide absolute return time series generated by the model
by factor of 3. This indicates that the model still lacks something, though it seems to reproduce correct behavior for
the tail of the distribution. As we can see in Fig. 7 after this correction the agreement between the model and the data
appears to be rather good. The obtained parameter set is similar to the one obtained for the Bitcoin case. Though there
are some differences. The mood swings seem to be larger in NYSE case (the respective ξ0 is larger), while Bitcoin trading
would seem to be more random (which is somewhat consistent with assumptions made in [33,34]). On the other hand
chartists seem to submit less order in NYSE case (λtc is smaller), which would indicate prices in NYSE a more impacted
by the market fundamentals. Base trading activity seems to be higher, λ0 is larger, for NYSE. The last two differences
could potentially indicate that NYSE is a more mature market than Bitcoin exchanges, as is also noted based on empirical
observations in [60].
182 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191

Fig. 7. Comparison between the empirical NYSE stocks’ statistical properties (gray curves) and statistical properties generated by the model (red
curves). Parameter values are identical to the ones used in Fig. 6 except: ξ0 = 1, λtc = 2, λ0 = 1.5.

6. Impact of the model parameters

In this section we check to see how changing the model’s parameter values impact the statistical properties of absolute
return and trading activity generated by the model. In all figures in this section we will show three curves. Usually one
will be generated with a larger parameter value than used to produce Fig. 6 (blue curve), one smaller (green curve) and
one identical (red curve).
Models built on the non-extensive formulation of the herd behavior model are known to avoid the N–dependence
problem [10,11,36–39,41,42], but as we can see in Fig. 8 this model has some kind of N–dependence. Yet the fluctuations
do not disappear with larger N, it seems that the model starts to exhibit even fatter tails. This is most likely occurs due
to the implemented mood mechanism: the more agents and the more chartist agents, the more mood is reflected in the
time series.
We can partly eliminate this dependence be recalling that while y dynamics are not influenced by N, but the number
of trades per time window does depend on N. By requiring that N λtc = const and N λtf = const we eliminate this
dependence. And as we can see in Fig. 9 then changing N does not influence the statistical properties of absolute return.
Though N retains the impact on the trading activity.
Changing λe parameter value seems to have a similar impact as changing h in the original herd behavior model (see
Fig. 10). Namely the PSD of the absolute return shifts to the right as we increase λe . Though due to the absolute return
formula comparing log-prices at two different points in time, the PDF of the absolute return might also be impacted: it
seem that the PDF might obtain heavier tails as λe becomes larger. Interestingly changing λe does not seem to have any
qualitative effect on the statistical properties of the trading activity. Most likely increasing λe simply increase the mean
of trades per time window without changing anything else.
Parameter εfc does not have significant impact on the statistical properties of the model with instantaneous clearing,
Eq. (16), but it seems that it is able to impact the statistical properties of the absolute return in the order book model
(see Fig. 11). As εfc increases the tails of the PDF becomes heavier and the PSD becomes flatter. The statistical properties
of the trading activity do not seem to change qualitatively, the tail of the PDF remains the same nor does the steepness
of the PSD change. Most likely larger εfc simply increases mean trading activity.
Parameter εcf seems to have the opposite effect on the statistical properties of absolute return (see Fig. 12). As εcf
increases the tail of the PDF become lighter, while the slope of the PSD becomes steeper. The impact on the statistical
properties of the trading activity seems to be both quantitative, the mean number of trades per time window decreases
as εcf increases, and qualitative, the tail of the PDF becomes lighter and the slope of the PSD becomes steeper. These
effects are most likely caused by the dynamics reflected by the model with instantaneous clearing as such dependence is
predicted by Eq. (16).
 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191 183

Fig. 8. Influence of N parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except: N = 5000
(blue curve), N = 50 (green curve).

Fig. 9. Influence of N parameter on the model’s statistical properties, when N λtc = const and N λtf = const. Parameter values are identical to the
ones used in Fig. 6 except: N = 5000, λtc = 2.5, λtf = 7.5 (blue curve), N = 50, λtc = 250, λtf = 750 (green curve).

The mood dynamics, ξ0 and λmc , does not seem to have a significant impact on the statistical properties of both absolute
return and trading activity (see Figs. 13 and 14). Though small effect of ξ0 on the absolute return PDF and PSD are visible.
Larger ξ0 makes the tail of the PDF less fat and the PSD flatter. Changing λmc would have a larger if the base value of ξ0
184 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191

Fig. 10. Influence of λe parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except:
λe = 3 · 10−7 s−1 (blue curve), λe = 3 · 10−8 s−1 (green curve).

Fig. 11. Influence of εfc parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except: εfc = 15
(blue curve), εfc = 0.7 (green curve).

was larger. Increasing λmc would have similar effect as decreasing ξ0 as with larger λm the effective mood (the average
trend) would be smaller than the true value of ξ0 .
 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191 185

Fig. 12. Influence of εcf parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except: εcf = 6
(blue curve), εcf = 0.7 (green curve).

Fig. 13. Influence of ξ0 parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except: ξ0 = 0.6
(blue curve), ξ0 = 0.1 (green curve).

As we have seen in the previous section larger λtc and λtf values force the realized prices to more closely follow
the equilibrium prices. While looking at the statistical properties of the model we see that λtf does not seem to have a
noticeable effect (see Fig. 16). This is because the fundamentalists activate only if the current price deviates far from the
186 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191

Fig. 14. Influence of λm parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except: λm = 300
(blue curve), λm = 0.3 (green curve).

Fig. 15. Influence of λtc parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except: λtc = 250
(blue curve), λtc = 2.5 (green curve).

fundamental price and their trades rapidly push the price back to the fundamental price. On the other hand λtc seems to
have a significant effect (see Fig. 15): larger values lead to the fatter tails of the PDFs, while the PSDs flatten.
 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191 187

Fig. 16. Influence of λtf parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except: λtf = 750
(blue curve), λtc = 7.5 (green curve).

Fig. 17. Influence of λ0 parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except: λ0 = 4
(blue curve), λ0 = 0.04 (green curve).

As we can see in Fig. 17 changing λ0 values does not have a significant effect besides increasing the overall level of
trading activity.
188 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191

Fig. 18. Influence of α parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except: α = 1 (blue
curve), α = 0 (green curve).

Changing the power of the feedback scenario α seems to have an adverse effect (see Fig. 18): with smaller values the
tails of the PDFs become lighter and the slopes of the PSDs become steeper.
Interestingly, as can be seen in Figs. 19 and 20, the parameters influencing the overall shape of the order book itself, k
and θ , do not seem to have any effect on the statistical properties obtained from the normalized time series. This result
could be attributed to the simplifying assumption about the homogeneity of the valuation. If valuation would be allowed
to be heterogeneous as in [25,26], the order book shape parameters would likely have a more profound effect on the
observed statistical properties.

7. Conclusions

Here we have proposed an order book model with herd behavior, which is able to reproduce the main stylized facts
of the financial markets. The order book model with herd behavior was built upon empirical insights by Kanazawa
et al. [25,26], who have studied a very detailed records of the order book level events, and our previously proposed
theoretical ABM [10,11], which is known to successfully reproduce the statistical properties of the high-frequency absolute
return. Incorporating order book dynamics improves upon our previous work in numerous ways. First of all, we were able
to scrap two not very realistic, but still common in the literature, assumptions: we no longer need to introduce an efficient
market maker to define the market price (introduced in [10]), we also no longer need to introduce the exogenous noise
as was done in [11]. Another key improvement is that now we are able to consider statistical properties of the trading
activity alongside the statistical properties of absolute return.
Using simulated annealing, optimizing the root mean squared error of the worst match, we were able to calibrate the
model parameters to match the Bitcoin’s statistical properties on one minute timescale. Calibrating the model to match
the statistical properties observed in NYSE (one minute timescale) was not as successful, which indicates that the model
still lacks something. We believe that introducing heterogeneity into chartist and fundamentalist valuation of the stock
might be the key, but this would further complicate the model introducing additional parameters. Another possibility
would be to complicate the modeling of the chartist mood swings. Finally, the model structure itself suggests that some
of the parameter values could be gleaned from the order book level data. Currently we are gathering publicly available
Bitcoin order book level data in hopes to use the collected data to improve calibration of the model.
To limit the complexity of the model and to allow for analytical tractability we have made few highly simplifying
assumptions. Most restrictive of them are homogeneous valuation, unit order volume and mood dynamics assumptions.
As relaxing these assumptions could enrich the dynamics of the proposed order book model with herding behavior, we
consider doing so in the future works.
 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191 189

Fig. 19. Influence of k parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except: k = 16 (blue
curve), k = 1 (green curve).

Fig. 20. Influence of θ parameter on the model’s statistical properties. Parameter values are identical to the ones used in Fig. 6 except: θ = 100 p.u.
(blue curve), θ = 2.5 p.u. (green curve).

A relevant future goal could be improvement economic soundness of the trading strategies and switching behavior
of the agents. This would make model more easily comparable with the other recent approaches [27,29–32,35] as well
190 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191

as open up the possibility to provide a deeper insight for the economic policy makers. Similar transition was already
undertaken by Biondo (starting from [28] and arriving to [29–31]).
We could also extend our approach further by considering its dynamics analytically. This should be possible, because
one of the underlying models can be alternatively described using stochastic differential equations, while the order book
part of the model can also be approached analytically using Bogoliubov–Born–Green–Kirkwood–Yvon hierarchy. Though
fully integrating the both parts could prove to be a challenging task. There could be a couple of possible approaches: a
superstatistical or a coupled SDE approach as used in [62,63] or a coarse-grained approximation of the model as discussed
in [64].

Acknowledgments

This research was funded by the European Social Fund under the No 09.3.3–LMT–K–712 ‘‘Development of Competences
of Scientists, other Researchers and Students through Practical Research Activities’’ measure.

References

[1] D. Helbing, Pluralistic modeling of complex systems, Sci. Culture 76 (2) (2010) 315, arXiv:1007.2818.
[2] A. Chakraborti, I.M. Toke, M. Patriarca, F. Abergel, Econophysics review: I. Empirical facts, Quant. Finance 7 (2011) 991–1012, http://dx.doi.org/
10.1080/14697688.2010.539248.
[3] A. Chakraborti, I.M. Toke, M. Patriarca, F. Abergel, Econophysics review: II. Agent-based models, Quant. Finance 7 (2011) 1013–1041,
http://dx.doi.org/10.1080/14697688.2010.539249.
[4] M. Cristelli, L. Pietronero, A. Zaccaria, Critical overview of agent-based models for economics, in: F. Mallnace, H.E. Stanley (Eds.), Proceedings
of the School of Physics ‘‘E. Fermi", Course CLXXVI, SIF-IOS, Bologna-Amsterdam, 2012, pp. 235–282, http://dx.doi.org/10.3254/978-1-61499-
071-0-235.
[5] D. Sornette, Physics and financial economics (1776–2014): puzzles, ising and agent-based models, Rep. Progress Phys. 77 (2014) 062001,
arXiv:1404.0243.
[6] F. Abergel, H. Aoyama, B. Chakrabarti, A. Chakraborti, N. Deo, D. Raina, I. Vodenska (Eds.), Econophysics and Sociophysics: Recent Progress and
Future Directions, Springer, 2017.
[7] T.-T. Chen, B. Zheng, Y. Li, X.-F. Jiang, New approaches in agent-based modeling of complex financial systems, Front. Phys. 12 (6) (2017) 128905,
http://dx.doi.org/10.1007/s11467-017-0661-2.
[8] A. Flache, M. Mas, T. Feliciani, E. Chattoe-Brown, G. Deffuant, S. Huet, J. Lorenz, Models of social influence: Towards the next frontiers, J. Artif.
Soc. Soc. Simul. 20 (4) (2017) 2, http://dx.doi.org/10.18564/jasss.3521, URL http://jasss.soc.surrey.ac.uk/20/4/2.html.
[9] T. Lux, M. Marchesi, Scaling and criticality in a stochastic multi-agent model of a financial market, Nature 397 (1999) 498–500.
[10] A. Kononovicius, V. Gontis, Agent based reasoning for the non-linear stochastic models of long-range memory, Physica A 391 (4) (2012)
1309–1314, http://dx.doi.org/10.1016/j.physa.2011.08.061, arXiv:1106.2685.
[11] V. Gontis, A. Kononovicius, Consentaneous agent-based and stochastic model of the financial markets, PLoS ONE 9 (7) (2014) e102201,
http://dx.doi.org/10.1371/journal.pone.0102201, arXiv:1403.1574.
[12] A.P. Kirman, Ants, rationality, recruitment, Q. J. Econ. 108 (1993) 137–156.
[13] B. Biais, P. Hillion, C. Spatt, An empirical analysis of the limit order book and the order flow in the Paris Bourse, J. Finance 1995 (1995)
1655–1689.
[14] P. Gopikrishnan, V. Plerou, X. Gabaix, H.E. Stanley, Statistical properties of share volume traded in financial markets, Phys. Rev. E 62 (2000)
4493–4496.
[15] J.P. Bouchaud, M. Mezard, M. Potters, Statistical properties of stock order books: empirical results and models, Quant. Finance 2 (2002) 251.
[16] E. Smith, J.D. Farmer, L. Gillemot, S. Krishnamurthy, Statistical theory of the continuous double auction, Quant. Finance 3 (6) (2003) 481–514,
http://dx.doi.org/10.1088/1469-7688/3/6/307.
[17] M. Alexandre, F.F. Ferreira, C.R. Neto, Order. book, Order book order flow and returns: evidence from the Brazilian stock exchange, Redeca 5
(1) (2018) 24–38.
[18] P. Bak, M. Paczuski, M. Shubik, Price variations in a stock market with many agents, Physica A 246 (1997) 430–453.
[19] S. Maslov, Simple model of a limit order-driven market, Physica A 278 (2000) 571–578, http://dx.doi.org/10.1016/S0378-4371(00)00067-4.
[20] D. Challet, R. Stinchcombe, Analyzing and modeling 1+1d markets, Physica A 300 (2001) 285–299.
[21] T. Preis, S. Golke, W. Paul, J.J. Schneider, Multi-agent-based order book model of financial markets, Europhys. Lett. 75 (3) (2006) 510,
http://dx.doi.org/10.1209/epl/i2006-10139-0.
[22] M. Cristelli, V. Alfi, L. Pietronero, A. Zaccaria, Liquidity. crisis, Liquidity crisis granularity of the order book and price fluctuations, Eur. Phys. J.
B 73 (1) (2010) 41–49, http://dx.doi.org/10.1140/epjb/e2009-00353-6.
[23] F. Abergel, A. Jedidi, A mathematical approach to order book modeling, Int. J. Theor. Appl. Finance 16 (5) (2013) 1350025, http://dx.doi.org/10.
1142/S0219024913500258.
[24] J. Donier, J.P. Bouchaud, From Walras’ auctioneer to continuous time double auctions, J. Statist. Mech. 2016 (2016) 123406, http://dx.doi.org/
10.1088/1742-5468/aa4e8e.
[25] K. Kanazawa, T. Sueshige, H. Takayasu, M. Takayasu, Derivation of the Boltzmann equation for the financial Brownian motion, Phys. Rev. Lett.
120 (2018) 138301, http://dx.doi.org/10.1103/PhysRevLett.120.138301, arXiv:1703.06739.
[26] K. Kanazawa, T. Sueshige, H. Takayasu, M. Takayasu, Kinetic theory for finance Brownian motion from microscopic dynamics, Phys. Rev. E 98
(2018) 052317, http://dx.doi.org/10.1103/PhysRevE.98.052317, arXiv:1802.05993.
[27] T. Kaizoji, M. Leiss, A. Saichev, D. Sornette, Super-exponential endogenous bubbles in an equilibrium model of fundamentalist and chartist
traders, J. Econom. Behav. Organiz. 112 (2015) 289–310, http://dx.doi.org/10.1016/j.jebo.2015.02.001.
[28] A.E. Biondo, A. Pluchino, A. Rapisarda, A multilayer model of order book dynamics, J. Netw. Theory Finance 2 (3) (2016) 37–52, http:
//dx.doi.org/10.21314/JNTF.2016.021, URL http://www.siecon.org/online/wp-content/uploads/2016/09/BIONDO.pdf.
[29] A.E. Biondo, Learning.to. forecast, risk. aversion, Learning to forecast risk aversion and microstructural aspects of financial stability, Econom.:
Open-Access, Open-Assessment E-J. 12 (2018) 1–21, http://dx.doi.org/10.5018/economics-ejournal.ja.2018-20.
[30] A.E. Biondo, Order book modeling and financial stability, Journal of Economic Interaction and Coordination https://doi.org/10.1007/s11403-018-
0227-6.
 A. Kononovicius and J. Ruseckas / Physica A 525 (2019) 171–191 191

[31] A.E. Biondo, Order book microstructure and policies for financial stability, Stud. Econom. Finance 35 (2018) 196–218, http://dx.doi.org/10.1108/
SEF-04-2017-0087.
[32] R.M. Navarro, H. Larralde, A detailed heterogeneous agent model for a single asset financial market with trading via an order book, PLOS ONE
12 (2) (2017) 1–27, http://dx.doi.org/10.1371/journal.pone.0170766.
[33] L. Cocco, G. Concas, M. Marchesi, Using an artificial financial market for studying a cryptocurrency market, J. Econ. Interact. Coord. 12 (2)
(2017) 345–365, http://dx.doi.org/10.1007/s11403-015-0168-2.
[34] L. Cocco, R. Tonelli, M. Marchesi, An agent based model to analyze the Bitcoin mining activity and a comparison with the gold mining industry,
Future Internet 11 (1) (2019) 8, http://dx.doi.org/10.3390/fi11010008.
[35] B. Llacay, G. Peffer, Using realistic trading strategies in an agent-based stock market model, Comput. Math. Organ. Theory 24 (3) (2018) 308–350,
http://dx.doi.org/10.1007/s10588-017-9258-0.
[36] S. Alfarano, M. Milakovic, Network structure and N-dependence in agent-based herding models, J. Econom. Dynam. Control 33 (1) (2009) 78–92,
http://dx.doi.org/10.1016/j.jedc.2008.05.003.
[37] A. Flache, M.W. Macy, Local convergence and global diversity: From interpersonal to social influence, J. Conflict Resol. 55 (6) (2011) 970–995,
http://dx.doi.org/10.1177/0022002711414371, arXiv:0808.2710.
[38] A. Kononovicius, J. Ruseckas, Continuous transition from the extensive to the non-extensive statistics in an agent-based herding model, Eur.
Phys. J. B 87 (8) (2014) 169, http://dx.doi.org/10.1140/epjb/e2014-50349-0, arXiv:1404.0856.
[39] A.F. Peralta, A. Carro, M. San Miguel, R. Toral, Analytical and numerical study of the non-linear noisy voter model on complex networks.
arXiv:1803.06861.
[40] R. Cont, Empirical properties of asset returns: Stylized facts and statistical issues, Quant. Finance 1 (2001) 1–14.
[41] S. Alfarano, T. Lux, F. Wagner, Estimation of agent-based models: The case of an asymmetric herding model, Comput. Econ. 26 (1) (2005)
19–49.
[42] S. Alfarano, T. Lux, F. Wagner, Time variation of higher moments in a financial market with heterogeneous agents: An analytical approach, J.
Econom. Dynam. Control 32 (2008) 101–136.
[43] V. Alfi, M. Cristelli, L. Pietronero, A. Zaccaria, Minimal agent based model for financial markets I: Origin and self-organization of stylized facts,
Eur. Phys. J. B 67 (3) (2009) 385–397, arXiv:0808.3562.
[44] V. Alfi, M. Cristelli, L. Pietronero, A. Zaccaria, Minimal agent based model for financial markets II: Statistical properties of the linear and
multiplicative dynamics, Eur. Phys. J. B 67 (3) (2009) 399–417, arXiv:0808.3565.
[45] F. Sano, M. Hisakado, S. Mori, Mean field voter model of election to the house of representatives in Japan, Phys. Soc. Jpn (2016) 011016,
http://dx.doi.org/10.7566/JPSCP.16.011016, arXiv:1702.03603.
[46] A. Kononovicius, Empirical analysis and agent-based modeling of Lithuanian parliamentary elections, Complexity 2017 (2017) 7354642,
arXiv:1704.02101.
[47] A. Kononovicius, V. Gontis, V. Daniunas, Agent-based versus macroscopic modeling of competition and business processes in economics and
finance, Int. J. Adv. Intell. Syst. 5 (1–2) (2012) 111–126, arXiv:1202.3533. URL http://www.thinkmind.org/index.php?view=article&articleid=
intsys_v5_n12_2012_9.
[48] Z. Chen, T. Lux, Estimation of sentiment effects in financial markets: A simulated method of moments approach, Comput. Econ. 2016 (2016)
1–34, http://dx.doi.org/10.1007/s10614-016-9638-4.
[49] T. Lux, Estimation of agent-based models using sequential Monte Carlo methods, J. Econom. Dynam. Control 91 (2018) 391–408, http:
//dx.doi.org/10.1016/j.jedc.2018.01.021.
[50] R. Rak, S. Drozdz, J. Kwapien, P. Oswiecimka, Stock returns versus trading volume: is the correspondence more general?, Acta Phys. Polon. B
44 (2013) 2035–2050, http://dx.doi.org/10.5506/APhysPolB.44.2035, arXiv:1310.7018.
[51] J. Ruseckas, B. Kaulakys, 1/f noise from nonlinear stochastic differential equations, Phys. Rev. E 81 (2010) 031105, arXiv:1002.4316.
[52] V.L. Smith, An experimental study of competitive market behavior, J. Polit. Economy 70 (1962) 111–137, http://dx.doi.org/10.1086/258609.
[53] M. Bayati, C. Borgs, J. Chayes, Y. Kanoria, A. Montanari, Bargaining dynamics in exchange networks, J. Econom. Theory 156 (2015) 417–454,
http://dx.doi.org/10.1016/j.jet.2014.02.007.
[54] H.H. Nax, B.S.R. Pradelski, Evoliutionary dynamics and equitable core selection in assignment games, Internat. J. Game Theory 44 (2015) 903–932,
http://dx.doi.org/10.1007/s00182-014-0459-1.
[55] J. Leshno, B. Pradelski, Efficient price discovery and information in the decentralized assignment game (2017). http://dx.doi.org/10.2139/ssrn.
3057554.
[56] E.O. Kimbrough, A. Smyth, Testing the boundaries of the double auction: The effects of complete information and market power, J. Econom.
Behav. Organiz. 150 (2018) 372–396, http://dx.doi.org/10.1016/j.jebo.2018.01.017.
[57] H.H. Nax, D.N. Duran, B. Pradelski, Price discovery in online markets: Convergence, asymmetries and information (2018). http://dx.doi.org/10.
2139/ssrn.3131004.
[58] D.T. Gillespie, Exact stochastic simulation of coupled chemical reactions, J. Phys. Chem. 81 (1977) 2340–2361.
[59] D.T. Gillespie, Markov Processes: An Introduction for Physical Scientists, Academic Press, San Diego CA, 1992.
[60] S. Drozdz, R. Gebarowski, L. Minati, P. Oswiecimka, M. Watorek, Bitcoin market route to maturity? Evidence from return fluctuations, temporal
correlations and multiscaling effects, Chaos 28 (2018) 071101, http://dx.doi.org/10.1063/1.5036517, URL https://aip.scitation.org/doi/10.1063/1.
5036517.
[61] S. Begusic, Z. Kostanjcar, H.E. Stanley, B. Podobnik, Scaling properties of extreme price fluctuations in Bitcoin markets, Physica A
510 (2018) 400–406, http://dx.doi.org/10.1016/j.physa.2018.06.131, arXiv:1803.08405, URL https://www.sciencedirect.com/science/article/pii/
S0378437118308550?via.
[62] J. Ruseckas, V. Gontis, B. Kaulakys, Nonextensive statistical mechanics distributions and dynamics of financial observables from the nonlinear
stochastic differential equations, Adv. Complex Syst. 15 (supp01) (2012) 1250073, http://dx.doi.org/10.1142/S0219525912500737.
[63] J. Ruseckas, R. Kazakevicius, B. Kaulakys, Coupled nonlinear stochastic differential equations generating arbitrary distributed observable with
1/f noise, J. Statist. Mech. 2016 (4) (2016) 043209.
[64] Y. Zou, V.A. Fonoberov, M. Fonoberova, I. Mezic, I.G. Kevrekidis, Model reduction for agent-based social simulation: Coarse-graining a civil
violence model, Phys. Rev. E 85 (2012) 066106, http://dx.doi.org/10.1103/PhysRevE.85.066106.