Hindawi

Complexity
Volume 2023, Article ID 3996948, 12 pages
https://doi.org/10.1155/2023/3996948

Research Article
Impact of High-Frequency Trading with an Order Book Imbalance
Strategy on Agent-Based Stock Markets

Isao Yagi ,1 Mahiro Hoshino,2 and Takanobu Mizuta3

1
Department of Information Systems and Applied Mathematics, Faculty of Informatics, Kogakuin University,
Tokyo 163-8677, Japan
2
NTT East-Minamikanto Corporation, Yokohama 231-0023, Japan
3
SPARX Asset Management Co. Ltd., Tokyo 108-0075, Japan

Correspondence should be addressed to Isao Yagi; iyagi2005@gmail.com

Received 1 September 2022; Revised 7 January 2023; Accepted 30 January 2023; Published 17 February 2023

Academic Editor: Ning Cai

Copyright © 2023 Isao Yagi et al. Tis is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

It is known that there is a positive correlation between order book imbalance and future returns. Although some previous studies
using actual trading data have suggested that high-frequency trading (HFT) may take this characteristic into account, HFT frms
have not disclosed their specifc strategies. Furthermore, there has been a long-standing debate in the empirical research feld as to
whether HFT is the cause of fash crashes, but no fnal conclusion has been reached. In the present study, we analysed the impacts
of HFT taking into account the correlation between order book imbalance and future returns on a stable market and on a market
with a fash crash, using agent-based simulations, which are said to be capable of analysing events in their essence. We also
analysed how HFT investment performance difers between those two market conditions. Te results showed that HFT has the
efect of further stabilizing the market when the market is stable but does not take place during fash crashes and so is unable to
afect the market either for the good or the bad. Te results also suggest that the proposed HFT’s performance is more sensitive to
market price fuctuations than conventional HFT (i.e., HFT following a position market-making strategy) and tends to have high
risk and high returns.

1. Introduction have increased in depth, which means the number of orders

whose prices are around the best price in the order book.
In recent years, high-frequency trading (HFT), which uses Using NASDAQ trading data, Carrion [5] found that
powerful computers to trade at high speed, has had a sig- spreads widened in trades where HFT provided liquidity and
nifcant impact on the fnancial markets. For example, narrowed in trades where HFT took away liquidity. From
Ohyama and Suzuki [1] reported that the proportion of HFT these results, the authors conclude that HFT provides li-
orders among all orders was about 70%, whereas the pro- quidity when liquidity is scarce and consumes liquidity when
portion of the trading value of HFT was about 40% in the liquidity is plentiful.
Tokyo Securities Exchange (TSE). Accordingly, there have On the other hand, researchers are divided on
been numerous studies on the impact of HFT on a market. whether HFT provides liquidity during a fnancial crisis.
Some studies have reported that HFT provides market li- For example, Ohyama and Suzuki [1] claimed HFT
quidity [2–5]. Jarnecic et al. [3] compared HFT orders with provided liquidity in terms of spread, i.e., it narrowed
the remainder of trading orders in the limit order book and bid/ask spreads, during the COVID-19 crisis. However,
found that the former were submitted at multiple prices to Weill [6] studied the optimal liquidity provision of HFT
the limit order book, concentrated around the quote. Tis during fnancial crashes and revealed that HFT provided
fnding indicated HFT improves market liquidity. Hosaka the optimal amount of liquidity if HFT had sufcient
[4] has put forward the suggestion that the HFT transactions capital at the time of the disruption. Tis result showed
2 Complexity

that HFT does not always provide buy orders at the time occur, it is difcult to extract only its direct impact on the
of a crash. market because various other external factors exist simul-
Tere have been many studies about the impact of HFT on taneously in the market.
the market during a fash crash, such as the fash crash of May 6, Tis study attempts to circumvent these barriers and
2010, in particular [7–11]. For example, the CFTC-SEC reported analyze the intrinsic impact of a fash crash on a market
that HFT initially provided liquidity to the market, but between by using an artifcial market (AM). An AM is a fnancial
2:41 and 2:44 p.m., HFT traders aggressively sold about 2,000 E- market multiagent system constructed virtually on
Mini contracts in order to reduce their temporary long positions a computer [21–25]. In an AM, each agent, which is
[7, 8]. Sornette and Von der Becke [9] stated that it was an assumed to be an investor, is given a unique trading
almost certain that HFT would lead to a higher frequency of strategy and is allowed to trade fnancial assets based on
crashes if there was much aggressive high-frequency selling that its own strategy. We can then see how the market is
would frequently clear out all 10 levels of depth before the ofer afected by the behaviors of many agents and how the
price could adjust downward during the fash crash on May 6, agents in turn are afected by changes in the market. Tere
2010, as reported by Nanex (Nanex report on May 6th fash have been previous studies on market liquidity using AMs
crash) https://www.nanex.net/20100506/FlashCrashAnalysis_ [26–29], and the relationship between HFT and fash
CompleteText.html). Golub et al. [10] concluded that it ap- crashes has also been investigated with AMs [30, 31].
pears likely that mini fash crashes are caused by HFT activity, Karvik et al. [30] used AMs to fnd that the frequency of
given the speed and the magnitude of the crashes. fash crashes may be proportional to the frequency of
Various HFT strategies have also been reported HFT transactions. Tis fnding suggests that HFT is
[4, 12, 13]. ASIC [12] categorized strategies into three types: a cause of fash crashes, as also suggested by the empirical
electronic liquidity provision strategies, statistical arbitrage studies mentioned above. Vuorenmaa and Wang [31]
strategies, and liquidity detection strategies. A market- found using an AM that the probability of a fash crash
making strategy is a type of electronic liquidity provision increases as the number of HFT agents increases or their
strategy in which the majority of HFT trading volume and positions decrease.
more than 80% of HFT limit order submissions are asso- However, there has been no discussion of the impacts of
ciated with market-making [13]. HFT with an OBI strategy on stable markets and markets
Tere have also been a number of studies on the market with fash crashes, based on the results of analyses with AMs;
impact of order imbalance, which means the imbalance between furthermore, as mentioned above, empirical studies have not
sell and buy order submissions. One type of such studies looks clarifed whether HFT behaviour is the cause of fash crashes.
at order book imbalance (OBI). OBI is the diference between Terefore, the contributions in our paper are as follows.
the number of buy orders and that of sell orders in the order
book around the best quote. It is known that OBI is correlated (i) A stable market and an unstable market with a fash
with future returns, i.e., positive returns are obtained when there crash as AMs were built to investigate how three
are more buy orders than sell orders around the best quote, and types of HFT agents afect the markets. Te three
negative returns are obtained when there are more sell orders types of HFT agents were as follows. One was an
than buy orders [14, 15]. Hereinafter, we call this property the HFT agent with a market-making strategy, which is
OBI property. Empirical analysis reveals that HFTmay be based common among HFT strategies. Another was an
on the correlation between OBI and future returns [16–18]. HFT agent with an OBI strategy that combines
Cartea et al. [15] investigated the correlation between high- a market-making strategy with one looking at the
frequency price changes and OBI (they referred to OBI as the positive correlation between OBI and future
order fow imbalance) and proposed a linear model of the returns. Te third was an HFTagent with the reverse
relationship between them. Goldstein et al. [18] claimed that the OBI strategy that combines a market-making
relative level of stock prices in the future increases as the strategy with one looking at the negative correla-
number of buy orders relative to sell orders in the limit order tion between OBI and future returns.
book increases. Although few reports have been published on (ii) Te fnding that HFT agents’ transactions contrib-
specifc strategies, Stoikov [19] proposed a micro-price that is ute to more stable prices in the stable market,
a mid-price adjustment that incorporates OBIs and bid-ask whereas their transactions are not expected to be
spreads. He found that the micro-price was a better predictor efective in stabilizing price in the market with the
for short-term movements of mid-prices than mid-prices and fash crash, was obtained.
volume-weighted mid-prices.
(iii) It was investigated how the performance of each
In recent years, although many complex forecasting
HFTagent is afected by each of the two markets and
methods based on deep learning have been developed for it was found that the HFT strategy taking into ac-
stock price prediction [20], HFT is often conducted using count the OBI tends to be high-risk and high-
relatively simple algorithms for speed. However, such al- return.
gorithms can cause large losses if unexpected events such as
fash crashes occur. Nevertheless, when trying to analyze the Te rest of this paper is organized as follows. Section 2
impact of an unexpected event such as a fash crash on HFT, explains our proposed AM. Section 3 reports on simulations
it is impractical and difcult to cause such an event to occur that we conducted using our AM. Section 4 presents the
in an actual market. Even if a fash crash can be made to simulation results. Section 5 gives our conclusions.
Complexity 3

2. AM Model We now explain the order process and the learning

process of NA. In the order process, NA j decides the order
We built a new AM with three types of HFT agents based on prices as follows. Te rate of the expected price of NA j at
the model constructed by Yagi et al. [29], as their model is time t, i.e., the expected return, is given by the following
able to reproduce the statistical characteristics of the long- equation.
term price fuctuations found in previous empirical studies.
In our market model, normal agents (Section 2.1) and three 1
types of HFT agents (Section 2.2) are available. Te position- re tj � t t t t t
􏼐w1 j r1 j + w2 j r2 j + uj ϵj 􏼑. (1)
w1 tj + w2 tj + uj
based market maker (PMM) adopts a position market-
making strategy, the position-based OBI market maker
(POMM) combines the position market-making strategy Te initial term in equation (1) normalizes the impact of the
with an OBI strategy, and a position-based reverse OBI three trading strategies. ri tj is the i-th expected return of NA j at
market maker (PrOMM) combines the position market- time t. r1 tj denotes the expected return of the fundamental
making strategy with the reverse OBI strategy. Note that strategy and is calculated as ln (Pf /Pt− 1 ), which means that NA
PrOMM may not be following a realistic strategy, but it is j expects a positive (negative) return when the fundamental
provided for comparison with PMM and POMM. Te price is higher (lower) than the previous market price. Note that
pricing mechanism in the model is a continuous double Pt is the market price at time t, and Pf is the fundamental price,
auction. Tis means that if there are buy (sell) order prices in which is constant over time. Te initial market price P0 is set to
the order book that are higher (lower) than the sell (buy) Pf , and the market price is set to the most recent price when no
order price of the agent, then an agent’s order is immediately trades have been made. r2 tj denotes the expected return of the
matched to the highest buy order (lowest sell order) in the technical strategy and is calculated as ln (Pt− 1 /Pt− 1− τt ), which
order book. On the other hand, if there are no orders in the means that NA j expects a positive (negative) return when the
order book, then the order does not match any other order historical return is positive (negative). Note that τ t is taken from
and remains in the order book. We call the former type of the uniform distribution between 1 and τ max at the start of the
order placed by an agent a market order and the latter type simulation for NA j. ϵtj is set as a normally distributed random
a limit order. Te orders in the order book are canceled at error with a mean of 0 and standard deviation σ ε .
time tc (i.e., the order efective period) after the order was wi tj is the strategy weight of NA j at time t, w1 tj denotes
placed. Te tick size ∆P is the minimum unit for the price. the fundamental strategy weight, and w2 tj denotes the
When orders are sell (buy) orders, fractional values smaller technical strategy weight. Both strategy weights are decided
than ∆P are rounded up (down). according to the uniform distribution between 0 and wi, max
Normal agent (NA) j � 0 to NA j � n − 1 submit orders at the beginning of the simulation. Each weight is changed
for one share at a time in sequence, where after the order by by using the learning process described later. uj is the noise
NA n − 1 is submitted, NA 0 submits the next order. Te strategy weight and is decided according to the uniform
time t is incremented by 1 each time an NA places an order. distribution between 0 and umax at the beginning of the
Terefore, the process goes forward one step even if an order simulation. We model the strategy weights as random
placed by an NA becomes a limit order, that is, if the order variables chosen independently.
does not match to any other orders in the order book. An Te expected price of NA j at time t, Pe tj , is calculated
HFTagent places both a buy order and a sell order just before according to the following equation.
the NA submission. Te HFT agent cancels their previous Pe tj � Pt− 1 exp􏼐re tj 􏼑. (2)
buy and sells in the order book if any remains there and
places new buy and sell limit orders. Te process does not Te order price of NA j at time t, Po tj , is determined by
move forward one step after an HFT order submission. All the uniform distribution between Pe tj − Pd and Pe tj + Pd ,
agents trade only one type of risk asset. All agents can trade where Pd is constant. If Po tj is less than Pe tj , then, NA j
assets indefnitely because the quantity of cash of each agent submits a buy order whose price is Po tj for one share. If Po tj is
is set indefnitely, and they can short sell. greater than Pe tj , then NA j submits a sell order whose price
is Po tj for one share.
2.1. Normal Agents (NAs). NAs are assumed to be general Te learning process is performed before the order
investors in the real world and designed to be able to process at each time. Comparing the sign of ri tj (i � 1, 2) with
replicate the characteristic of the real markets. An NA de- that of rtl � ln (Pt− n /Pt− 1− tl ), if both signs are the same, wti,j
termines the order price by combining the following three is updated as follows:
trading strategies: the fundamental strategy, the technical 􏼌􏼌 􏼌􏼌 t
strategy, and the noise strategy. Te fundamental strategy wti,j ←wti,j + kl 􏼌􏼌rtl 􏼌􏼌qj 􏼐wi, max − wti,j 􏼑, (3)
refers to the fundamental price to make investment de-
cisions. Te technical strategy uses past price movements to where kl is a constant, and qtj is set according to the uniform
make investment decisions. Te noise strategy represents distribution between 0 and 1. If ri tj and rtl have opposite
trial-and-error investment decisions. Te weights of the signs, wti,j is updated as follows:
fundamental and technical strategies are changed by 􏼌􏼌 􏼌􏼌 t
wti,j ←wti,j − kl 􏼌􏼌rtl 􏼌􏼌qj wti,j . (4)
learning as market conditions change.
4 Complexity

Tese equations mean that the weights of strategies this, Ptfv,pm is calculated nonlinearly from Pt,mid in
whose predicted direction of price change coincides with the equation (5). Te power exponent of the weighted portion
actual direction of price change are raised, while the weights of the position is set to an odd number (�3) to take into
of strategies that are out of line are lowered. Furthermore, account the direction of the position. As a result, the
wti,j is set according to the uniform distribution between larger the HFT agent’s buy (sell) position is, the easier its
0 and wi, max with probability δl . Tis is an objective model of sell (buy) order fnds a match.
the search for a better strategy by trial and error to fun- OBI at time t, otpm , is defned by the following equation:
damentally reevaluate the previous strategy.
Dpt,buy − Dpt,sell
2.2. HFT Agents. Tere are three types of HFT agents, i.e., otpm � , (6)
PMM, POMM, and PrOMM, in our market model. PMM Bpt,buy + Bpt,sell
adopts a position market-making strategy. Tis strategy is
as follows. First, PMM calculates the basic order price as where Dpt,buy , Dpt,sell , Bpt,buy , and Bpt,sell are the buy depth
the mid-price. Te mid-price is the average of the best-bid at time t, the sell depth at time t, the amount of buy orders,
price and the best-ask price plus a value depending on its and the amount of sell orders in the order book, respectively.
own position, which means the amount of assets held by When the coefcient of OBI is wom , the basic order price
PMM. Next, PMM simultaneously submits a buy order of POMM Ptfv,pom and that of PrOMM Ptfv,prom , the buy
t,buy t,buy
whose price is the value of the basic order price minus the order price of POMM Po,pom and that of PrOMM Po,prom ,
t,sell
spread, which is equal to the amount of its own expected and the sell order price of POMM Po,pom and that of PrOMM
return per transaction, and a sell order whose price is the Pt,sell
o,prom are given by equations (7)–(12), respectively.
value of the basic order price plus the spread [29].
Generally, PMM tries to keep its position neutral to avoid 3 2
the price fuctuation risk of its own asset. Tus, the more Ptfv,pom � 􏼒1 − wpm 􏼐stpm 􏼑 + wom 􏼐otpm 􏼑 􏼓Pt,mid , (7)
PMM holds an asset, the lower PMM tends to set its buy
and sell order prices so that its sell order has a better 3 2
Ptfv,prom � 􏼒1 − wpm 􏼐stpm 􏼑 − wom 􏼐otom 􏼑 􏼓Pt,mid , (8)
chance of matching an NA’s buy market order. Other-
wise, the more PMM short-sells the asset, the higher it
tends to set its buy and sell order prices so that its buy 1
Pt,buy t
o,pom � Pfv,pom − Pf θpm , (9)
order has a better chance of matching an NA’s sell market 2
order [32, 33].
1
POMM and PrOMM determine their basic order prices Pt,sell t
o,pom � Pfv,pom + Pf θpm , (10)
using not only their own positions but also the diference 2
between the market buy depth and the market sell depth.
1
Note that depth means the amount of orders around the Pt,buy t
o,prom � Pfv,prom − Pf θpm , (11)
highest range of buy orders and the lowest range of sell 2
orders. In this paper, the market buy depth is the amount of 1
buy orders from the best-bid price to Dp (�50) ticks lower Pt,sell t
o,prom � Pfv,prom + Pf θpm . (12)
2
than the best-bid price, and the market sell depth is the
amount of sell orders from the best-ask price to Dp ticks
higher than the best-ask price. Note that the larger the buy (sell) depth is than the sell
Let the best-ask and best-bid prices, the mid-price, the (buy) depth, the more likely POMM’s buy (sell) limit
base spread of the HFT agent, the HFT agent’s position, and orders are to be executed with sell (buy) market orders
the coefcient of its position be Pt,sell , Pt,buy , Pt,mid � (Pt,sell + relative to NAs’ buy (sell) limit orders. On the other hand,
Pt,buy )/2, θpm , stpm , and wpm , respectively. Ten, the HFT the larger the buy (sell) depth is than the sell (buy) depth,
t,buy the more likely PrOMM’s sell (buy) limit orders are to be
agent’s basic order price Ptfv,pm , buy order price Po,pm , and
t,sell
sell order price Po,pm are as determined by the following executed with buy (sell) market orders of NAs. Fur-
equations: thermore, when OBI is small, POMM (PrOMM) does not
want to have much infuence on its own order strategy
3 because OBI may be coincidental, but when it is large,
Ptfv,pm � 􏼒1 − wpm 􏼐stpm 􏼑 􏼓Pt,mid ,
POMM (PrOMM) wants to make a large correction
because of the possibility of some large buying (selling)
1 pressure. To model these characteristics, the weighted
Pt,buy t
o,pm � Pfv,pm − Pf θpm , (5)
2 portions of OBI in equations (7) and (8) are made
nonlinear.
1
Pt,sell t
o,pm � Pfv,pm + Pf θpm .
Here, the order prices of HFT agents are reset according
2 to equations (13) and (14) and do not become market orders
t,buy
As mentioned above, PMM trades without taking of the HFT agents if either or both of Po,pm ≥ Pt,sell and
large positions to avoid price fuctuation risk. To model Pt,sell
o,pm ≤ P
t,buy
are satisfed.
Complexity 5

t,buy
If Po,pm ≥ Pt,sell , then Table 1: Parameters.

Pt,buy
o,pm � P
t,sell
− ∆P, Parameter Initial value
(13) te 800,000
Pt,sell
o,pm � 􏼐P
t,sell
− ∆P􏼑 + Pf θpm . N 1,000
w1,max 1
If Pt,sell
o,pm ≤ P
t,buy
, then w2,max 10
umax 1
Pt,buy
o,pm � 􏼐P
t,buy
+ ∆P􏼑 − Pf θpm , τmax 10,000
(14) σε 0.06
Pt,sell
o,pm � P
t,buy
+ ∆P. Pd 1,000
tc 10,000
∆P 1.0
3. Simulation Pf 10,000
tl 10,000
3.1. Overview. In this study, we built two market simulation kl 4
environments. In one, the market price transition is stable; δl 0.01
that is, market volatility is low. Hereinafter, we call this the θpm 0.003
stable market. In the other, a fash crash occurs, the market wpm 5.0 × 10− 8
price transition becomes unstable, and market volatility is wom 10.0
high. We call this the unstable market. Letting each of the
three types of HFT agents participate in each market, we
Table 2: Stylized facts.
investigate how the HFT agent afects market price, vola-
tility, and the performance of the HFT agent. Kurtosis 4.089
Te stable market is achieved by keeping a fundamental Lag
price Pf constant throughout the simulation. Te unstable 1 0.158
market is realized by having a fash crash occur, i.e., having 2 0.123
Autocorrelation coefcients for squared returns
NAs submit sell orders with order price 1 with a probability 3 0.090
20% during the 30,000 time units from time 100,001 to time 4 0.074
130,000. HFT agents place their orders according to the 5 0.064
proposed strategies in the simulations (including during the Proportion of the OBI property 2.396%
plunge).
Each simulation ends at the time te � 800,000. Te values
of the other parameters are shown in Table 1.
recently executed time t, we calculate the value obtained by
′
subtracting the number of times the return (ln (Pt /Pt )) is
3.2. Validation of Proposed AM. Real fnancial markets have negative (positive) from the number of times the return is
particular statistical properties such as volatility clustering positive (negative) at the next execution time t′, and then
and a fat tail, as empirical studies have indicated [34–36]. We divide the result by the number of transactions (volume).
confrmed whether our proposed AM reproduces the sta- Tis value is positive if the OBI property is satisfed and is
tistical features of the price transitions [37]. Although early called here the proportion of OBI property.
AMs were not able to reproduce the ubiquitous scaling laws Te proportion of OBI property in Table 2, which is the
of returns, recent AMs have been able to replicate volatility average of 20 simulation runs, is indeed positive in this
clustering and a fat tail [29, 38–42]. Tus, we set the model model and is statistically signifcant at the 0.05 level (since
parameters as listed in Table 1 so as to reproduce these previous studies [14, 15] focused on the correlation between
features (these are also the settings used in Yagi et al. [29]). the magnitude of OBI and the magnitude of returns, the
Table 2 shows that both the autocorrelation coefcients for correlation was analysed quantitatively using, for example,
squared returns with several lags and kurtosis are positive, regression lines. On the other hand, since we focused on the
which means that all runs replicated volatility clustering and qualitative impact of the OBI property on HFT agents’
a fat tail. Tus, the model reproduces long-term statistical behaviours, it is sufcient to confrm that the OBI property
characteristics observed in actual fnancial markets. Te holds in our AM). Terefore, it can be confrmed that the
statistics for stylized facts in Table 2 are averages over 20 proposed model is valid.
simulation runs, for which we calculated price returns at
intervals of 100 time units. 4. Results and Discussion
As mentioned in Section 1, it is known that OBI is
correlated with future returns (OBI property) in actual f- 4.1. Price Transitions. In this section, we compare market
nancial markets, i.e., positive returns are obtained when price transitions in a stable market with those in an unstable
there are more buy orders than sell orders around the best market with fash crashes when each of the three types of
quote, and negative returns when there are more sell orders HFT agents (PMM, POMM, and PrOMM) participates. In
than buy orders [14, 15]. When the buy depth Dpt,buy is the stable markets, price changes in the market where an
greater than (less than) the sell depth Dpt,sell at the most HFT agent participates are smaller than those in the market
6 Complexity

where HFT agents do not participate (refer to Figure 1). Tis price rebound either. Terefore, in Phase 3, many buy
result is consistent with the result of Yagi et al. [29] that orders are submitted, most of them become market or-
when HFT agents trade in the stable market, prices converge ders, and the price rebounds. In Phase 4, the proportions
approximately between the buy and sell order prices of HFT of both buy and sell limit orders of HFT agents that
agents (the HFT agents’ spread), resulting in price stability. matched market orders are higher than those of NAs,
On the other hand, in the unstable market, price changes are indicating that the orders of HFT agents match many of
similar between whether or not HFT agents participate (refer the market orders.
to Figure 2). If the orders of HFT agents played the same role As a result, although there are some diferences between
as in the stable market, the price declines would not be as HFT agents’ transactions, the above explanation roughly
large as those in the market without HFT agents because the accounts for why HFT agents do not participate much in
buy orders of HFT agents would support the price when the trading when prices change widely, and why HFT is not
fash crash occurred. Tus, we checked the extent to which expected to be efective in stabilizing volatility.
buy orders of HFT agents were executed during each phase
in the unstable market.
Tables 3–5 summarize the numbers of orders in the order 4.2. HFT Agents’ Performances. For both the stable and un-
book for each agent type when a new order is placed and stable markets, the performances of the three types of HFT
a trade is executed during each of the four phases. Tat is, agents at the end of the simulation are shown in Table 6. Tese
Tables 3–5 show the number of limit orders, which market performances are the averages over 20 simulation runs. Figures 3
orders match for each agent type by phase. Tese values were and 4 show the performance transitions of each HFTagent in the
calculated as averages of 20 simulation runs. Phase 1 is the stable and unstable markets, respectively.
period before the fash crash begins. Phase 2 is from the time Te performance of POMM is the best among the three
NAs’ market sell orders at price 1 begin to be placed until the types of HFT agents in the stable market; however, it is the
market price reaches its lowest point. Phase 3 is during the worst among them in the unstable market.
rebound of the price from the lowest point. In this exper- In the stable market, the position market-making
iment, the rebound is considered to have ended when prices strategies of all HFT agents appear to work properly, al-
recover to Pf − dpf (dpf � 50). Phase 4 is the period from the though there are some diferences between their positions.
end of Phase 3 to the end of the simulation. Te reasons for the diferences in performance between the
In Phase 1, most of the buy (sell) orders in the order HFT agents may be as follows.
book, which matched market sell (buy) orders were placed First, the reason that the performance of POMM is better
by HFT agents. For example, the proportion of buy (sell) than that of PMM seems to be that the proft from the OBI
orders of PMM to the total limit orders executed is 83.5% strategy is added to the proft from the position market-
(82.6%) in Table 3. Terefore, we can see that most of the making strategy, as there is a positive correlation between
market orders matched the limit orders of HFT agents. OBI and future returns. On the other hand, the performance
In Phase 2, the proportion of buy limit orders of HFT of PrOMM is worse than that of PMM, as PrOMM’s strategy
agents, which matched market orders is signifcantly lower expected a negative correlation between OBI and future
than that in Phase 1 (e.g., that of PMM is 2.0%), while the returns.
proportion of buy limit orders of NAs is much higher (e.g., Next, we attempt to explain the mechanisms of HFT
that of NA is 98.0%). On the other hand, the amount of sell agents’ performances in the unstable market.
limit orders of HFT agents (e.g., that of PMM is 136.0) is When the market price declines sharply, PMM’s buy
larger than that of NAs (that of NAs is 6.8); however, the position becomes large (refer to Figure 5) because PMM’s
total amount of sell limit orders, which matched market buy buy limit orders match the sell orders of NAs at price 1
orders (e.g., 7,033.4 (�6,893.3 + 140.1) in Table 3), is much during the fash crash. Terefore, PMM’s performance de-
more than that of buy limit orders, which matched market teriorates rapidly as the price falls and the buy limit orders of
sell orders (e.g., 142.8). PMM are executed one after another (see Figure 4).
From the abovementioned statements, it can be seen that When the market price reaches the bottom, the
HFT agents rarely trade, while NAs trade with each other in downward trend of PMM’s performance moderates. Tis
Phase 2. Tis result suggests that HFT agents’ trades have is because PMM changed from a buy position to a sell
little efect on market price declines. position around the bottom price. When the price falls
Finally, we also check the behaviours of HFT agents signifcantly, the divergence between it and the funda-
and NAs. In Phase 3, the proportions of buy limit orders mental price increases, so the infuence of the funda-
of HFT agents are higher than those of sell limit orders of mental strategy of NAs becomes stronger. As a result,
HFT agents (e.g., those of PMM are 80.6% and 21.3%, many NAs begin to place buy orders, and the market price
respectively). However, as the total amount of sell limit stops falling and rebounds. PMM’s performance de-
orders that matched market buy orders (those of PMM teriorates because the price rises while PMM maintains
and NAs are 5437.8.) is more than that of buy limit orders a sell position, but the deterioration of PMM’s perfor-
that matched market sell orders (those of PMM and NAs mance is slow because the price rises slowly.
are 1420.5), HFT agents’ trades have not had as great of an Later, as the price rebound progresses and the market
impact on price formation as they did during the plunge; price begins to converge with the fundamental price, PMM’s
i.e., HFT does not seem to have contributed much to the performance begins to improve. Te reason for this is as
Complexity 7

10100
10080
10060
10040
10020
Price

10000
9980
9960
9940
9920
9900
91000
127000
163000
199000
235000
271000
307000
343000
379000
415000
451000
487000
523000
559000
595000
631000
667000
703000
739000
775000
Time

PMM PrOMM

POMM NONE
Figure 1: Price transitions of the stable market.

10400

10200

10000

9800
Price

9600

9400

9200

9000
91000
127000
163000
199000
235000
271000
307000
343000
379000
415000
451000
487000
523000
559000
595000
631000
667000
703000
739000
775000

Time

PMM PrOMM

POMM NONE
Figure 2: Price transitions of the unstable market.

follows. When the price becomes as high as the fundamental Since POMM takes the position of market-making strategy,
price, NAs place buy orders at a price equal to the funda- its performance transition is basically the same as that of PMM,
mental price, but PMM, wishing to unwind its sell position, but the infuence of the OBI strategy is further refected in its
places buy orders at higher prices, which leads to the un- strategy. Tus, during the fash crash when the buy depth is
winding of PMM’s buy position. Tis is because the position greater than the sell depth in the order book (refer to Figures 5
market-making strategy begins to function when the excess and 6), as POMM takes a larger buy position than PMM,
sell position is eliminated. POMM’s performance is more afected by the fash crash than
8 Complexity

Table 3: Numbers and proportions of NA’s and PMM’s orders in the order book that matched a new order.
NA PMM
Buy orders 190.9 (16.5%) 965.6 (83.5%)
Phase 1
Sell orders 202.5 (17.4%) 963.9 (82.6%)
Buy orders 6,893.3 (98.0%) 140.1 (2.0%)
Phase 2
Sell orders 6.8 (4.8%) 136.0 (95.2%)
Buy orders 275.5 (19.4%) 1145.0 (80.6%)
Phase 3
Sell orders 4277.7 (78.7%) 1160.1 (21.3%)
Buy orders 592.5 (15.0%) 3,357.8 (85.0%)
Phase 4
Sell orders 986.2 (22.7%) 3,350.6 (77.3%)

Table 4: Numbers and proportions of NA’s orders and POMM’s orders in the order book that matched a new order.
NA PMM
Buy orders 178.3 (15.2%) 993.0 (84.8%)
Phase 1
Sell orders 192.7 (16.3%) 990.3 (83.7%)
Buy orders 6905.5 (97.7%) 163.3 (2.3%)
Phase 2
Sell orders 9.4 (6.5%) 134.2 (93.5%)
Buy orders 329.3 (24.2%) 1,031.3 (75.8%)
Phase 3
Sell orders 4,327.8 (80.1%) 1,077.5 (19.9%)
Buy orders 567.8 (14.2%) 3,439.8 (85.8%)
Phase 4
Sell orders 982.0 (22.3%) 3,424.1 (77.7%)

Table 5: Numbers and proportions of NA’s orders and PrOMM’s orders in the order book that matched a new order.
NA PMM
Buy orders 365.5 (32.7%) 753.4 (67.3%)
Phase 1
Sell orders 369.5 (32.9%) 755.4 (67.1%)
Buy orders 6,987.4 (98.9%) 77.3 (1.1%)
Phase 2
Sell orders 7.9 (6.7%) 109.2 (93.3%)
Buy orders 297.2 (21.2%) 1,105.7 (78.8%)
Phase 3
Sell orders 4,308.9 (80.2%) 1,063.8 (19.8%)
Buy orders 1,177.8 (32.0%) 2,498.7 (68.0%)
Phase 4
Sell orders 1,575.9 (38.6%) 2,510.2 (61.4%)

Table 6: Performances of HFT agents.

PMM POMM PrOMM
Stable market 29,241.9 30,950.1 6,485.9
Unstable market −647.2 −23,927.0 32,834.8

that of PMM. Terefore, its performance is worse than that of PrOMM’s performance is more strongly afected by
PMM. Note that the reason why the sell depth is extremely small the OBI strategy than by the position market-making
during the price sharp decline is that the price falls so sharply strategy, because the diference between the sell depth
that NAs’ sell orders do not have time to accumulate in the price and buy depth during the price rebound period after
range between the market price and Dp ticks above it. During a sharp decline is so large. Note that the transitions of the
the rebound period, because the sell depth is larger than the buy sell and buy depths in the market with PrOMM are al-
depth, POMM’s sell position becomes larger than PMM’s sell most the same as those in the market with POMM (refer
position. Ten, POMM’s performance deteriorates further to Figure 6). Since PrOMM takes a strategy that expects
because POMM tries to decrease its sell position even if the OBI and future returns to be negatively correlated, after
market price rises. After that, when the diference between the reaching a buy position due to the position market-
sell depth and the buy depth converges to some extent, the making strategy at the time of the sharp decline, it
impact of the OBI strategy becomes smaller, and that of the quickly changes to a sell position (refer to Figure 5).
position market-making strategy becomes larger, POMM’s Tus, PrOMM’s performance improves during the
performance recovers. subsequent sharp decline. Moreover, when the price
Complexity 9

35000

30000

25000
Performance
20000

15000

10000

5000

0
91000
127000
163000
199000
235000
271000
307000
343000
379000
415000
451000
487000
523000
559000
595000
631000
667000
703000
739000
775000
Time

PMM

POMM

PrOMM
Figure 3: HFT agents’ performances in the stable market.

50000 10200

40000
10000
30000
9800
20000
Performance

10000 9600

Price
0 9400
-10000
9200
-20000
9000
-30000

-40000 8800
91000

136000

181000

226000

271000

316000

361000

406000

451000

496000

541000

586000

631000

676000

721000

766000

Time

PMM PrOMM
POMM Price (PMM)
Figure 4: HFT agents’ performances in the unstable market.
10 Complexity

80 10200

60 10000

40 9800

20 9600
Position

Price
0 9400

-20 9200

-40 9000

-60 8800
91000

133000

175000

217000

259000

301000

343000

385000

427000

469000

511000

553000

595000

637000

679000

721000

763000
Time

PMM PrOMM
POMM Price (PMM)
Figure 5: Positions of HFT agents and price transitions in the unstable market.

800 10200

700 10000

600
9800
500
9600
Depth

Price
400
9400
300
9200
200

100 9000

0 8800
91000

133000

175000

217000

259000

301000

343000

385000

427000

469000

511000

553000

595000

637000

679000

721000

763000

Time

BuyDepth

SellDepth

Price
Figure 6: Price and depth in the unstable market.

begins to rebound, it achieves a buy position immedi- 5. Conclusion

ately, and the performance is further improved as the
price rebounds. In this study, we prepared a stable market and an unstable
Tese results suggest that when fash crashes occur, market with a fash crash as AMs and investigated how
PrOMM may have the best performance, while the per- three types of HFT agents afect the markets and how the
formances of PMM and POMM may deteriorate. performance of each HFT agent is afected by each of
Complexity 11

these markets. Te three types of HFT agents in our AM behaviours among PMM, POMM, and PrOMM may depend
were as follows. One was an HFT agent with a position only on the OBI strategy if the basic order prices of these
market-making strategy (PMM), which is common agents are calculated by the same method, regardless of what
among HFT strategies. Another was a PMM with the OBI the base price is calculated by. Terefore, our second future
strategy (POMM), which it combines with the positive work is to investigate HFT behaviour and its impact on the
correlation between OBI and future returns. Te third market when the basic order price is diferent from the mid-
was a PMM with the reverse OBI strategy (PrOMM), price, as in the case of pure price. Finally, as we did not
which it combines with the negative correlation between compare our results with realistic data, we intend to research
OBI and future returns. the relationship between our results and realistic data as our
As a result, we found that price changes in the market third future work.
where HFT agents participate were smaller than those in the
market where HFT agents did not participate, whereas price Data Availability
changes were similar regardless of whether HFT agents
participated in the market where a fash crash occurred. Te data used to support the fndings of this study are
Tese results suggest that HFT agents’ transactions con- available at https://onl.bz/rHUULVT.
tribute to more stable prices in the stable market, whereas
their transactions are not expected to be efective in stabi- Disclosure
lizing price in the market with the fash crash.
Regarding the performances of HFT agents, it was Te opinions expressed herein are solely those of the authors
found that PMM accumulated solid profts in stable and do not necessarily refect those of SPARX Asset Man-
markets and recovered relatively quickly in the fash crash agement Co., Ltd. Te work presented here in is original and
market, even though its performance temporarily de- has not been previously published, but the initial idea was
teriorated. On the other hand, POMM achieved high presented at the 28th JSAI Special Interest Group on Fi-
performance in the stable market. Tis result is consistent nancial Informatics Workshop (https://sigfn.org/028/, in
with a fnding of Goldstein et al. [18], which claimed that Japanese) in part.
the relative level of stock prices in the future increases as
the number of buy orders relative to sell orders in the Conflicts of Interest
limit order book increases. Stoikov [19] proposed a HFT
Te authors declare that they have no conficts of interest.
strategy, which took into account a micro-price that is
a mid-price adjustment that incorporates OBIs and bid-
ask spreads. He reported that the micro-price was a better Acknowledgments
predictor for short-term movements of mid-prices than Tis work was supported by JSPS KAKENHI, grant no.
mid-prices and volume-weighted mid-prices. Tis fnd- 20K04977.
ing of Stoikov [19] also seems to be consistent with our
result. However, POMM performed worse than PMM in
References
the fash crash market. It seems difcult to compare the
result with the fndings of Goldstein et al. [18] and [1] Y. F. Ohyama and K. Suzuki, “Characterization of high-speed
Stoikov [19], as their fndings were based on market data trading,” FSA Staf Reports, Financial Services Agency,
from a relatively stable period. Finally, PrOMM’s per- Chiyoda, Japan, 2021.
formances were the opposite of those of POMM. Tese [2] A. J. Menkveld, “High frequency trading and the new market
results may suggest that a trading strategy that takes into makers,” Journal of Financial Markets, vol. 16, no. 4,
account the OBI property tends to be high-risk and high- pp. 712–740, 2013.
[3] E. Jarnecic, M. Snape, and M. Snape, “Te provision of li-
return.
quidity by high-frequency participants,” Financial Review,
As mentioned above, there have been previous studies vol. 49, no. 2, pp. 371–394, 2014.
that empirically analysed HFT investment strategies that [4] G. Hosaka, “Analysis of high-frequency trading at tokyo stock
take into account OBI, but the periods examined in those exchange,” Japan Exchange Group, Chuo, Japan, JPX
studies were limited to when the market was relatively stable. Working Papers, 2014.
Terefore, our frst future work is to analyse how an HFT [5] A. Carrion, “Very fast money: high-frequency trading on the
agent with the OBI strategy behaves during periods of NASDAQ,” Journal of Financial Markets, vol. 16, no. 4,
market instability, based on empirical data. Although all pp. 680–711, 2013.
HFT agents in this study set their order prices using the mid- [6] P.-O. Weill, “Leaning against the wind,” Te Review of Eco-
price as the basic order price, HFT investment behaviours nomic Studies, vol. 74, no. 4, pp. 1329–1354, 2007.
may change if the basic order price difers. For example, [7] S. E. C. Cftc, “Recommendations regarding regulatory re-
sponses to the market events of May 6, 2010,” Technical
Peña [43] proposed pure price as a diferent reference price
Report, CFTC-SEC Joint Advisory Committee, Washington,
from the mid-price and used pure price when proposing DC, USA, 2011.
algorithmic trading strategies to construct portfolios that [8] D. Easley, M. M. López de Prado, and M. Ohara, “Te mi-
balance proftability and default risk. Tus, the pure price crostructure of the fash crash: fow toxicity, liquidity crashes,
deserves consideration as one of the candidates for the basic and the probability of informed trading,” Journal of Portfolio
order price. On the other hand, the diferences in investment Management, vol. 37, no. 2, pp. 118–128, 2011.
12 Complexity

[9] D. Sornette and S. Von der Becke, Crashes and High Frequency [29] I. Yagi, Y. Masuda, and T. Mizuta, “Analysis of the impact of
Trading, Swiss Finance Institute Research Paper, Zürich, high frequency trading on artifcial market liquidity,” IEEE
Switzerland, 2011. Transactions on Computational Social Systems, vol. 7, no. 6,
[10] A. Golub, J. Keane, and S.-H. Poon, High Frequency Trading pp. 1324–1334, 2020.
and Mini Flash Crashes, SSRN, Pamplemousses, Mauritius, [30] G.-A. Karvik, J. Noss, J. Worlidge, and D. Beale, “Te deeds of
2012. speed: an agent-based model of market liquidity and fash
[11] A. Kirilenko, A. S. Kyle, M. Samadi, and T. Tuzun, “Te fash episodes,” Technical Report, Bank of England, London, UK,
crash: high-frequency trading in an electronic market,” Te 2018.
Journal of Finance, vol. 72, no. 3, pp. 967–998, 2017. [31] T. A. Vuorenmaa and L. Wang, “An agent-based model of the
[12] A. Securities and I. Commission, “Report 215: Australian fash crash of May 6, 2010, with policy implications,” Social
equity market structure,” Regulatory Report, ASIC, Kobe, Science Research Network, Pamplemousses, Mauritius, 2014.
Port island, 2010. [32] Y. Nakajima and Y. Shiozawa, “Usefulness and feasibility of
[13] B. Hagströmer and L. Nordén, “Te diversity of high- market maker in a thin market,” in Proceedings of the In-
frequency traders,” Journal of Financial Markets, vol. 16, ternational Conference Experiments in Economic Sciences: New
no. 4, pp. 741–770, 2013. Approaches to Solving Real world Problems, pp. 1000–1003,
[14] Á. Cartea, S. Jaimungal, and J. Penalva, Algorithmic and Kyoto, Japan, December 2004.
Highfrequency Trading, Cambridge University Press, Cam- [33] Y. Kusada, T. Mizuta, S. Hayakawa, and K. Izumi, “Impacts of
bridge, UK, 2015. position-based market makers on markets’ shares of trading
[15] Á. Cartea, R. Donnelly, and S. Jaimungal, “Enhancing trading volumes -an artifcial market approach,” Social Modeling and
strategies with order book signals,” Applied Mathematical Simulations+ Econophysics Colloquium 2014, vol. 30, 2014.
Finance, vol. 25, no. 1, pp. 1–35, 2018. [34] M. Sewell, “Characterization of fnancial time series,” 2011,
[16] C. Cao, O. Hansch, and X. Wang, “Te information content of http://www.cs.ucl.ac.uk/fleadmin/UCL-CS/research/Research_
an open limit-order book,” Journal of Futures Markets, vol. 29, Notes/RN_11_01.pdf.
no. 1, pp. 16–41, 2009. [35] R. Cont, “Empirical properties of asset returns: stylized facts
[17] R. Cont, A. Kukanov, and S. Stoikov, “Te price impact of and statistical issues,” Quantitative Finance, vol. 1, no. 2,
order book events,” Journal of Financial Econometrics, vol. 12, pp. 223–236, 2001.
no. 1, pp. 47–88, 2014. [36] R. Tsay, “Analysis of fnancial time series,” in Wiley Series in
[18] M. A. Goldstein, A. Kwan, and R. Philip, “High-frequency Probability and Statistics, Wiley-Interscience, Hoboken, NJ,
trading strategies,” Management Science, 2022. USA, 2nd edition, 2005.
[19] S. Stoikov, Te micro-price: A High Frequency Estimator of [37] H. Levy, M. Levy, and S. Solomon, Microscopic Simulation of
Future Prices, SSRN, Pamplemousses, Mauritius. Financial Markets: From Investor Behavior to Market Phe-
[20] A. Takkar and K. Chaudhari, “A comprehensive survey on nomena, Elsevier, Amsterdam, Netherland, 2000.
deep neural networks for stock market: the need, challenges, [38] B. LeBaron, “Agent-based computational fnance: suggested
and future directions,” Expert Systems with Applications, readings and early research,” Journal of Economic Dynamics
vol. 177, Article ID 114800, 2021. and Control, vol. 24, no. 5-7, pp. 679–702, 2000.
[21] R. G. Palmer, W. Brian Arthur, J. H. Holland, B. LeBaron, and [39] M. Raberto, S. Cincotti, S. M. Focardi, and M. Marchesi,
P. Tayler, “Artifcial economic life: a simple model of “Agent based simulation of a fnancial market,” Physica A:
a stockmarket,” Physica D: Nonlinear Phenomena, vol. 75, Statistical Mechanics and Its Applications, vol. 299, no. 1-2,
no. 1-3, pp. 264–274, 1994. pp. 319–327, 2001.
[22] W. Arthur, J. Holland, B. Lebaron, R. Palmer, and P. Tayler, [40] C. Chiarella and G. Iori, “A simulation analysis of the mi-
“Asset pricing under endogenous expectations in an artifcial crostructure of double auction markets,” Quantitative Fi-
stock market,” in Te Economy as an Evolving Complex System nance, vol. 2, no. 5, pp. 346–353, 2002.
II, pp. 15–44, Addison-Wesley, Boston, 1997. [41] C. H. Hommes, “Heterogeneous agent models in economics
[23] C. Chiarella, G. Iori, and J. Perelló, “Te impact of hetero- and fnance,” Handbook of Computational Economics, vol. 2,
geneous trading rules on the limit order book and order pp. 1109–1186, 2006.
fows,” Journal of Economic Dynamics and Control, vol. 33, [42] I. Yagi, S. Maruyama, and T. Mizuta, “Trading strategies of
no. 3, pp. 525–537, 2009. a leveraged ETF in a continuous double auction market using
[24] S.-H. Chen, C.-L. Chang, and Y.-R. Du, “Agent-based eco- an agent-based simulation,” Complexity, vol. 2020, Article ID
3497689, 7 pages, 2020.
nomic models and econometrics,” Te Knowledge Engineering
[43] G. Peña, An Algorithmic Trading Strategy to Balance Proft-
Review, vol. 27, no. 2, pp. 187–219, 2012.
ability and Risk,” Big Data in Finance, Palgrave Macmillan,
[25] C.-H. Yeh and C.-Y. Yang, “Do price limits hurt the market?”
London, UK, 2022.
Journal of Economic Interaction and Coordination, vol. 8,
no. 1, pp. 125–153, 2013.
[26] G. Iori, “A microsimulation of traders activity in the stock
market: the role of heterogeneity, agents interactions and
trade frictions,” Journal of Economic Behavior and Organi-
zation, vol. 49, no. 2, pp. 269–285, 2002.
[27] C. Chiarella and X.-Z. He, “Heterogeneous beliefs, risk, and
learning in a simple asset-pricing model with a market
maker,” Macroeconomic Dynamics, vol. 7, no. 04, pp. 503–536,
2003.
[28] S. Ganesh, N. Vadori, M. Xu, H. Zheng, P. Reddy, and
M. Veloso, “Reinforcement learning for market making in
a multi-agent dealer market,” 2019.