SSRN Id3870658

Michael J.
Brennan Irish Finance

Working Paper Series
This paper can be downloaded without charge from

Financial Economic Network Electronic Paper Collection:
http://ssrn.com/abstract=3870658

Electronic copy available at: https://ssrn.com/abstract=3870658

Rogue Traders∗
Huayuan Dong† Paolo Guasoni‡ Eberhard Mayerhofer§
June 30, 2021
Abstract
Investing on behalf of a firm, a trader can feign personal skill by committing fraud
that with high probability remains undetected and generates small gains, but that with low
probability bankrupts the firm, offsetting ostensible gains. Honesty requires enough skin
in the game: if two traders with isoelastic preferences operate in continuous-time and one
of them is honest, the other is honest as long as the respective fraction of capital is above
an endogenous fraud threshold that depends on the trader’s preferences and skill. If both
traders can cheat, they reach a Nash equilibrium in which the fraud threshold of each of
them is lower than if the other one were honest. More skill, higher risk aversion, longer
horizons, and greater volatility all lead to honesty on a wider range of capital allocations
between the traders.
JEL: G11
MSC (2010): 91G15, 91A23
Keywords: rogue trading, internal fraud, operational risk, stochastic differential games.
∗
We thank for helpful comments Jodi Dinetti, Tiziano De Angelis, Giorgio Ferrari, Martin Herdegen, Sergey
Nadtochiy, Frank Riedel, Kwok Chuen Wong, and seminar participants at Bielefeld University and AMaMeF
2021 conference.
†
Dublin City University, School of Mathematical Sciences, D09 W6Y4 Dublin, Ireland.
huayuan.dong2@mail.dcu.ie
‡
Dublin City University, School of Mathematical Sciences, D09 W6Y4 Dublin, Ireland.
paolo.guasoni@dcu.ie. Supported by the SFI (16/IA/4443,16/SPP/3347).
§
University of Limerick, Department of Mathematics and Statistics, V94 T9PX Limerick, Ireland.
eberhard.mayerhofer@ul.ie

Introduction
The expression “rogue trader” entered popular culture in 1995, when Nicholas W. Leeson, a
trader of an overseas office of Barings Bank in Singapore, made unauthorized bullish bets on
the Japanese stock market, concealing his losses in an error account. At first, losses were
recovered with a profit, but in the aftermath of the Kobe earthquake they reached $1.4 billion
(Brown and Steenbeek, 2001), forcing the 233-year old bank into bankruptcy. Earlier episodes
of rogue trading ante litteram include the losses of Robert Citron in 1994 for Orange County
($1.7 billion, Kenyon (1997)) and of Toshihide Iguchi in 1983-1995 for Daiwa Bank ($1.1 billion,
Iguchi (2014)). The earliest case is possibly the one involving the law firm of Grant & Ward in
1884, which embarrassed former president Ulysses S. Grant, one of the firm’s partners (Krawiec,
2000).
Since the demise of Barings Bank, rogue trading episodes have increased in frequency and
magnitude. In 2008, Jerome Kerviel, a junior trader at Société Générale who had been exceeding
positions limits through fictitious trades to avoid detection, eventually lost $7.6 billion, the
largest rogue trading loss in history. In his defense, he claimed that colleagues also engaged in
unauthorized trading.1
The rise in rogue trading and its threat to both financial institutions and financial stabil-
ity has been recognized by the Basel Committee as operational risk, defined as “the risk of
loss resulting from inadequate or failed internal processes, people and systems or from external
events” (Committee et al. (2011)). The Capital Accord of Basel II (and Basel III, to be en-
acted in 2023) include provisions for protection from operational losses: while insurance can
cover high-frequency, low-impact events, rogue trading falls squarely in the low-frequency, high-
impact category of uninsurable risks, which incur capital charges. Such charges are in turn
based on standardized approaches or statistical models, due in part to the absence of consensus
on the origin of rogue trading, which is the focus of this paper.
Our starting point is that “The continued existence of rogue trading [...] presents a mystery
for many scholars and industry observers.” (Krawiec (2000)). “Operational-risk is unlike mar-
ket and credit risk; by assuming more of it, a financial firm cannot expect to generate higher
returns.” (Crouhy et al. (2004)). In other words, prima facie it is hard to reconcile rogue
traders’ actions with the optimizing behavior of sophisticated rational agents.
We propose a model in which rational, self-interested, risk-averse traders deliberately engage
in fraudulent activity that has zero risk premium. While undetected, fraud allows a trader to
feign superior returns, ostensibly without additional risk. In reality, higher returns are exactly
offset by a higher probability of bankruptcy, thereby creating no value for the firm. Yet,
under some circumstances, fraud may be optimal for a trader because, while its benefits are
personal, potential bankruptcy costs are shared with other traders. Furthermore, a trader who
understands the circumstances leading to others’ fraud, can anticipate them and act accordingly,
leading to a dynamic Nash equilibrium.
In equilibrium, each trader abstains from fraud as long as the respective share of wealth
under management exceeds an endogenous fraud threshold that depends on both traders’ pref-
erences (risk aversions and average horizon) and investment characteristics (expected returns
and volatilities). Thus, a trader must have enough skin in the game to remain honest: when the
share of managed assets drops below the fraud threshold, then the marginal utility of fraudulent
trades becomes positive, and a trader cheats as little and as quickly as possible to restore the
wealth share to the honesty region. Importantly, such fraudulent activity does not generate
extra volatility, so that it cannot be detected by monitoring wealth before bankruptcy occurs.
These results bring several insights. First, our model suggests that rogue trading has an
1
See “Bank Outlines How Trader Hid His Activities” https://www.nytimes.com/2008/01/28/business/
worldbusiness/28bank.html and “Timeline of events in SocGen rogue trader case” https://www.reuters.com/
article/uk-socgen-kerviel-events-timeline-idUKL1885652420080318.

important social component: A sole trader investing all the firm’s capital would not engage in
fraud because such a trader would bear in full both the costs and the benefits of fraudulent
activity (Proposition 1.1). Furthermore, the fraud threshold is higher if a trader knows that
nobody else is cheating (Lemma 2.7 and Theorem 2.9).
Second, the model emphasizes the danger that traders with relatively small amounts of cap-
ital can pose to a financial institution, due to their insufficient stakes in the firm. This concern
is confirmed by the cases of the junior trader Jerome Kerviel and Nick Leeson. By reviewing
Mr. Leeson’s trading record and the investigation reports from Singaporean authorities, Brown
and Steenbeek (2001) suggest that he had excluded the error account (meant for traders to
settle minor trading mismatches) from the market reports to the headquarter and built up
unauthorized speculative position in the early days after assuming the duty at Baring’s office
in Singapore in 1992.
Third, our comparative statics offer some clues for assessing and mitigating rogue-trading
risk. The incidence of fraud is higher in less skilled traders, which means that emphasis on
performance evaluation has the indirect benefit of fraud reduction. Fraud also declines signifi-
cantly as risk aversion increases, suggesting that, ceteris paribus, the most fearless traders are
also the ones most tempted by fraud, and that the most dangerous combination is found in a
trader with high risk tolerance and low share of managed assets. Somewhat counter-intuitively,
a longer horizon does not necessarily imply lower fraud, though fraud eventually declines when
the horizon is long enough.
Fourth, our model hints at a subtle trade-off between investment performance and opera-
tional risk. Classical portfolio theory implies that diversification can only increase performance,
hence the addition of a trader with expertise in a new asset class always improves the risk-return
tradeoff. Yet, our results caution that a higher number of traders, each with a lower share of
assets under management, may also increase the appeal of fraud for each of them, potentially
worsening the firm’s risk profile. (The quantitative analysis of the trade-off between diversifi-
cation and fraud requires very different technical tools, hence is deferred to future research.)
This paper offers the first structural model of rogue trading, in which fraud arises from
agency issues between traders and their firms. A priori, it is traders’ hidden action that enables
fraudulent activity. A posteriori, the traders’ optimal strategies imply that fraud is both con-
tinuous and of finite-variation, which makes it hard to detect even for a hypothetical observer
who could continuously monitor traders’ wealth.
In the interest of both simplicity and relevance, the model assumes that each trader is
compensated with a fraction of trading profits, i.e., contracts are linear. As a result, the
fraudulent activity that arises in the model does not stem from nonlinear incentives that may
encourage risk-taking (cf. Carpenter (2000)), but merely from the asymmetric opportunity of
taking personal credit from fraudulent gains while sharing bankruptcy costs. In this sense, each
trader’s fraud represents an externality for other traders and the firm, whence overall demand
for fraud is socially suboptimal (i.e., nonzero).
The literature on rogue trading is relatively sparse. Most existing work explores the legal
(Krawiec (2000, 2009)), social psychological (Wexler (2010)), and regulatory (Moodie (2009))
aspects of rogue trading, and offer a number of hypotheses for mechanisms that may foster
malfeasance in trading. Armstrong and Brigo (2019) find that common risk measures are
ineffective in preventing excessive risk-taking by traders with tail risk-seeking preferences. In a
similar vein, Gwilym and Ebrahim (2013) argue that position limits are inadequate in restraining
rogue trading. Taking the perspective of a firms’ management, Xu et al. (2020) use stochastic
control to minimize operational risk through preventative and corrective policies. Xu et al.
(2017) review the recent literature on operational risk.
At the technical level, our analysis is based on the theory of nonzero-sum stochastic dif-
ferential games with singular controls. In contrast to single-agent singular stochastic control
problems, which date back to the finite fuel problem of Bather and Chernoff (1967), research

on singular stochastic differential games is relatively recent: Guo and Xu (2018) generalize the
finite fuel problem to an n-player stochastic game and a mean-field game, in which each player
minimizes the distance of an object to the center of N objects, while keeping her total amount
of controls at minimal. Guo et al. (2018) extend this analysis to a larger class of games with pos-
sible moving reflecting boundaries in Nash equilibria. De Angelis and Ferrari (2018) establish
the connection between a class of stochastic games with singular controls and a certain optimal
stopping game, where the underlying state process differ but the reflective and exit boundaries
coincide.
Kwon and Zhang (2015) and Ekström et al. (2020) study optimal stopping games in which
all or one of the players control an exit time that terminates the game. Note that the fraud in
Ekström et al. (2020) differs from the one considered here, in that their model entails an agent
stealing from another one, who seeks to detect fraud and can terminate the game. In these
papers, players are forbidden in part to execute discontinuous actions simultaneously, whereas
the present model does not impose such a restriction. Karatzas and Li (2012) and Hamadène
and Mu (2015) use BSDE techniques to investigate existence and uniqueness of Nash equilibrium
in singular (and exit) games, while Dianetti and Ferrari (2018) employ fixed point methods.
The results in the paper bear also a curious analogy with portfolio choice with proportional
transaction costs in that, similar to Davis and Norman (1990), the solution to the present
model leads to an inaction region, surrounded by two regions in which actions are performed
as little as necessary to return to the inaction region. Although the mechanisms underlying the
two models are very different, it is worth pointing out the common feature that leads to the
common structure. In both cases (and in many other singular control problems), an action is
performed only in positive amount (fraud of either trader in this paper, buying or selling in
portfoio choice). As a result, the inaction region arises when each action is counterproductive
for its agent, while the action regions are visited at their boundaries because costs are linear in
the action performed (bankruptcy probability in this paper, trading costs in portfolio choice).
The rest of this paper is organized as follows. Section 1 describes the model of rogue trading
and its rationale. Section 2 constructs a Nash equilibrium with two traders and states the main
result. Section 3 discusses the interpretation of the results and their implications. Concluding
remarks are in section 4. All proofs are in the Appendix.
1 A Model of Rogue Trading

Krawiec (2000) offers the following definition: “A rogue trader is a market professional who
engages in unauthorized purchases or sales of securities, commodities or derivatives, often for a
financial institution’s proprietary trading account.”
Most episodes of fraudulent trading share some distinctive features. First, they violate a
firm’s internal rules or external regulations. Second, fraud often remains concealed and results
in modest (relative to the firm’s size) gains that are ascribed to the skill of the perpetrator.
Third, fraud generates substantial risk without expected return for the firm, and is revealed
only when catastrophic losses eventually materialize.
To reproduce these features, it is useful to think of a small fraud as a (forbidden) bet that
a trader wages on the whole firm’s capital. With a small chance (say, ε), the bet bankrupts the
firm, but most of the time (with probability 1 − ε) it results in a return of 1/(1 − ε) ≈ 1 + ε,
which the trader can take credit for. Of course, the bet’s overall return for the firm is zero, as
(1 − ε) · 1/(1 − ε) + 0 · ε = 1. Such asymmetric outcomes (likely small gains against unlikely
large losses) are in fact common in both illicit and licit trading strategies (for example, selling
deep out-of-the money options), and have attracted the label of “picking up nickels in front of
a steamroller” (Duarte et al. (2006)).
Thus, the dilemma of an unscrupulous but profit-driven and risk-averse trader is to what
degree to engage in fraud, as cheating too little may forego some easy profits, but cheating too

much may result in likely bankruptcy. If one imagines the small fraud above as the outcome of a
(heavily biased) coin-toss, the trader essentially ponders how many coins to toss. For example,
tossing two coins would generate a likely payoff of (1 − ε)−2 but may also lead to bankruptcy
with probability 2ε − ε2 .
If the trader is the sole firm’s owner, it is not hard to see that fraud does not pay: when one
bears both gains and losses in full, waging fair bets on one’s capital merely replaces a payoff
with another one, more uncertain but with the same mean – an inferior choice by risk aversion.
In this sense, fraud arises from social interactions, both through the incentives implied by
traders’ compensation contracts, or by each trader’s awareness that others traders in the firm
may engage in fraud. The present model focuses on the latter motive by assuming that each
trader receives a fixed fraction of individual profits and losses, which is a common arrangement
for bonuses with clawback provisions. The model envisages multiple traders: each of them has a
mandate to invest a share of the firm’s capital in some risky asset with a positive risk premium
and is paid with a fraction of the terminal payoff. Thus, each trader’s objective is aligned with
the firm’s. For the sake of tractability and clarity, the paper focuses on the case of two traders.
The moral hazard stems from the asymmetric effects of fraud on a trader’s reward: as
long as the fraudulent activity is successful, the trader can disguise its revenues as the fruit
of personal skill in performing the investment mandate. In reality, such additional revenues
merely compensate for the fraudulent bets that the trader wages on the capital of the whole
firm, rather than personal capital (e.g. exceeding risk limits by either collateralizing firm’s
asset or assuming excess liabilities). Of course, such bets are possible exactly because they are
fraudulent, and are explicitly forbidden by the firm’s regulation: they nonetheless exist, due
to “inadequate or failed internal processes, people and systems” embodied in the definition of
operational risk (Committee et al. (2011)).
The appeal of fraud – privatizing gains while socializing losses – thus varies with a trader’s
share of the firm’s capital: intuitively, the temptation of fraudulently enriching oneself is much
stronger for a small trader, who has little to lose and much to gain from gambling with others’
wealth, than for a large trader who has significant skin in the game. For this reason, in the
present continuous-time model each trader can cheat with varying intensity in response to
changes in one’s and others’ wealth.
After this informal description, the precise definition of the model follows.
1.1 Investment and Fraud in Continuous Time

Fix a complete stochastic basis (Ω, F, P) equipped with the natural filtration F = (Ft )t≥0 of
an N -dimensional (N ≥ 1) Brownian motion B = (Bt )t≥0 , satisfying the usual hypotheses of
S
right-continuity and saturatedness. Denote by F∞ := σ t≥0 F t ⊂ F.
Assuming a zero safe rate to ease notation, in the absence of fraud the capital Yti of the i-th
trader (i ∈ {1, . . . , N }) evolves as
dYti,x = µi Yti,x dt + σi Yti,x dBti , Y0i,x = xi > 0,
reflecting the trader’s average ability µi > 0 to deliver excess returns with the volatility σi > 0
that the firm’s risk management is willing to accept. For simplicity, assume that B i and B j
are independent for i 6= j, which means that traders take uncorrelated risks (for example, one
invests in stocks and the other in bonds).
To describe how each trader may engage in fraud by endangering the firm’s capital, define
the class of increasing processes
n o
A := A = (At )t≥0 : F-adapted, right-continuous, non-decreasing, with A0− = 0 ∈ R .
For A ∈ A, At represents the cumulative amount of “bets” waged by a traderR on the firm’s
t
capital up to time t. To understand such a representation, suppose that At = 0 λs ds, which

means that in the interval [s, s + ds] the trader wages a fair bet that has the probability λs ds
of bankrupting the firm. Because the fraud is illicitly waged on the firm’s capital (thereby
exceeding the capital Y i,x that the trader has beenPassigned), if bankruptcy does not occur
such fraud yields a profit of YsS λs ds, where Y S,x := N k=1 Y
k,x is the total capital of the firm.
Although this description is intuitive, it has a twofold limit: First, it encompasses only the
case of fraud with a finite rate λs , in that it excludes bursts of rogue trades at any instant.
Second, such bursts do arise in the model, therefore cannot be excluded a priori for the sake of
simplicity. For this reason, a more careful but also more technical description is necessary.
To make precise the intuition that how dAs drives the bankruptcy rate, note first that
any A ∈ APis right-continuous and of finite variation. Therefore, it has the representation
At = Act + 0≤s≤t ∆As for any t ≥ 0, where ∆As = As − As− and Ac is the continuous part
of the process A with Ac0 = 0. 1 N N
PNFor ak set of fraud process A = (A , . . . , A ) ∈ A , denote the
S
total fraud process by A· = k=1 A· . The bankruptcy time is then defined as
τA = inf t ≥ 0 : ASt ≥ θ ,

(1.1)
where θ is an F-measurable exponential random variable with rate 1, independent of the filtra-
tion F. (Recall the convention that τA = +∞ on the empty set.)
Before bankruptcy occurs, the wealth of each trader follows the dynamics
dYti,x = µi Yti,x dt + σi Yti,x dBti + Yt−

S
dÃit , i,x
Y0− = xi > 0, 1 ≤ i ≤ N, (1.2)
where the integral with i

respect
i
to Ã in (1.2) is understood in the Lebesgue-Stieltjes sense, and
Ãit := Ai,c ∆As − 1 reflects the fact that the simple return of a jump in fraud is
P
t + 0≤s≤t e
not ∆ itself but rather e∆ − 1. Such a distinction is immaterial with continuous fraud because
e∆ − 1 ∼ ∆ for ∆ close to zero. The final expression for wealth, which includes the effect of
bankruptcy after τA , is
Xti,x = 1{t<τA } Yti,x with X0−i,x
= xi . (1.3)
Lemma A.2 in Appendix A.1 formally verifies that pre-bankruptcy wealth in (1.2) is well defined
by showing that Y x = Y 1,x , . . . , Y N,x is the unique strong solution to the N -dimensional

linear stochastic differential equation (SDE) in (1.2). (Upon bankruptcy (on the set t ≥ τA (ω)),
the wealth of all traders vanishes and remains null thereafter, hence the dynamics of the fraud
processes beyond τA (ω) is irrelevant for the model. Effectively, fraud is described by the stopped
process Ai·∧τA .)
Note that the bankruptcy time τA is not an F-stopping time. Thus, to accommodate the
wealth process X x = X 1,x , . . . , X N,x , it is necessary to make the minimal enlargement of the
filtration F that makes τA a stopping time. To this end, let HA = HtA t≥0 be the natural
filtration of the indicator process 1{·≥τA } t≥0 and define the enlarged filtration GA = GtA t≥0

as GtA = s>t Fs ∨ HsA , which is the smallest right-continuous filtration containing F such
T
that τA is a stopping time.2

As waging bets on one’s own wealth means bearing their risks in full, thus yielding a zero
risk-premium, a trader who owns the whole firm (N = 1) has a wealth process that is a GA -
martingale, in the absence of investment skill.3 (See Proposition A.4, which additionally justifies
the choice of the return from jump fraud.)
2
Such an extension is known as ‘progressive filtration enlargement’ (cf. Jeulin (2006) and Jeanblanc and
Le Cam (2009)). Moreover, the bankruptcy time τA is GA -predictable if and only if frauds occur only at time 0
or never occur (Lemma A.6).
3
In principle, one could consider the case of fraud with a negative risk premium. This model focuses on
the parsimonious case of zero-risk premium, which maximizes the propensity for a trader to cheat. If the risk
premium were positive, the bet would be a legitimate investment opportunity, for which the label of “fraud”
would not be justified.

The goal of each trader is to maximize expected utility over a random horizon τ , which is
an F-measurable exponential random variable with rate λ > 0, independent of both F∞ and θ
(hence of the bankruptcy time τA ). This random horizon models a trader with an open-ended
contract, whose mandate is to maximize profits in the long-term. The arrival rate λ summarizes
both the traders’ impatience and the likelihood that business may end for exogenous reasons
(that is, independently of traders’ performance).
A trader’s attitude to risk is represented by a utility function of power type
i x1−γ i
U (xi ) = i , with 0 < γi < 1.
1 − γi
In particular, the relative risk aversion parameter γi is below one, so that the utility is finite
also upon bankruptcy (xi = 0), and the problem is nontrivial. If it were greater or equal to one,
then zero wealth would be completely unacceptable (U (0) = −∞) and fraud would disappear.
In fact, as shown below (Remark 2.10), fraud does vanish as γi converges to one.
As anticipated in the description, an important implication of this model is that a rational
and strictly risk-averse trader abstains from fraud if no other trader is present. Its significance
is to confirm that in this model fraud stems from the ability to share losses but not gains,
hence disappears when such sharing disappears. (Note that for this result the assumption of an
exponential horizon, made in the rest of the paper, can be dropped.)
Proposition 1.1 (No fraud for N = 1). Let N = 1 and h τ1 be an F-measurable,

2 /2)τ
i a.s. finite
random horizon (independent of F∞ and θ) such that E e (1−γ1 )(µ1 −γ1 σ 1 1 < ∞.
If the sole trader maximizes
E U 1 (Xτ1,x

1
1
)
over all fraud processes A1 ∈ A, then A1,∗ is optimal if and only if A1,∗
t = 0 a.s. for any t ≥ 0
such that P (τ1 ≥ t) > 0. In particular:
(i) If τ1 is unbounded, then A1,∗

t = 0 a.s. for all t ≥ 0.
(ii) If τ1 ≤ T 1 a.s. for some T 1 > 0, then AT1,∗1 = 0. If P τ1 = T 1 > 0, then also AT1,∗1 = 0

−
a.s.
S,x
Note that this proposition fails if N ≥ 2 because the coupling term Yt− of (1.2) rescinds the
martingale property (Proposition A.4) for each trader’s wealth in the absence of drift (µi = 0).
For example, if all but the i-th trader abstain from fraud, then X i,x can become a sub-martingale
if the i-th trader cheats: in this case, the wealth processes of other traders become super-
martingales as they share the bankruptcy risk from the i-th trader action. As shown in Section
2, engaging in fraud may be optimal depending on traders’ shares of capital, risk-aversions,
drifts, and volatilities.
2 Main Result
While the presentation in the previous section considered an arbitrary number N of traders,
the main result in this section focuses on two traders to simplify both the exposition and the
proofs.4 Thus, henceforth N = 2 and, for clarity, the indexes {a, b} replace {1, 2} to identify
traders. The wealth processes are denoted by either X x (Aa , Ab ) or X x (respectively, Y x (Aa , Ab )
or Y x ), depending on the need to specify the fraud process A = (Aa , Ab ) in context.
4
A model with N traders implies that relative shares of capital follow a N − 1 dimensional diffusion, which is
considerably simpler in one dimension.

2.1 Definition of Nash equilibrium
For any i, j ∈ {a, b} such that i 6= j (henceforth abbreviated as ‘for any i 6= j ∈ {a, b}’ ), the
goal of trader i is to maximize expected utility on a random horizon τ given the other trader
j’s fraud process Aj , i.e.,
J i (x; Ai , Aj ) := E U i Xτi,x (Ai , Aj )

(2.1)
over any Ai ∈ A, where the random horizon τ is independent of F and θ, and exponentially
distributed with rate λ (meaning that λ1 represents traders’ average horizon or, equivalently, λ
is their time-preference (or impatience) rate). Let thus
V i (x; Aj ) := sup J i (x; Ai , Aj )

Ai ∈A
be the value function for i-th trader given trader j’s fraud process Aj and initial wealth x ∈ R2+ .
The next assumption is required to ensure that the optimization problem is well-posed: the
impatience rate λ needs to be large enough in relation to risk aversion and skill, so that the
value function remains finite. Note that this parameter restriction depends only on minimum
risk aversion and maximum skill.
Assumption 2.1. λ > (1 − γa ∧ γb ) (µa ∨ µb ).
The well-posedness of the value function V i is confirmed by the next result. (The proof is
in Section A.4.)
Lemma 2.2. For any i 6= j ∈ {a, b}, x ∈ R2+ and (Ai , Aj ) ∈ A2 ,
Z ∞
i i j −λt−AS i i,x i j
(i) J (x; A , A ) = λE e t U (Yt (A , A ))dt .
0
(ii) For any c > 0, J i (cx; Ai , Aj ) = c1−γi J i (x; Ai , Aj ).

λU i (xa + xb )
(iii) 0 < V i (x; Aj ) ≤ .
λ − (1 − γi )(µa ∨ µb )
By Lemma 2.2 (i), we take the approach to focus on the pre-bankruptcy wealth Y x as the
state processes of the optimization problem; By Lemma 2.2 (ii), we exploit the scale-invariance of
the value function to reduce the resulting Hamilton-Jacobi-Bellman (HJB) equations to ordinary
differential equations (see Section B for the derivation and simplification of the HJB equations).
For t ≥ 0, let C ([0, t]) (resp. C ([0, ∞))) denote the set of R-valued continuous functions on
[0, t] (resp. [0, ∞)). Analogously, let D↑ ([0, t]) (resp. D↑ ([0, ∞))) be the set of [0, +∞)-valued
non-decreasing, right-continuous functions on [0, t] (resp. [0, ∞)). Define the set of strategies
n
Λ = Ψ = {Ψt : t ≥ 0} such that Ψt : R2+ × C ([0, t])2 × D↑ ([0, t]) → [0, ∞)
is measurable and the path t 7→ Ψt (x, f[0,t) , g[0,t] ) is in D↑ ([0, ∞)) ,
o
for all x ∈ R2+ , f ∈ C ([0, ∞))2 , g ∈ D↑ ([0, ∞)) ,
where f[0,t) and g[0,t] denote the restrictions of the paths f and g to the time interval [0, t].
For any A ∈ A, the measurability of Ψt ensures that the process Ψt x, B[0,t) , A[0,t] t≥0 is
F-adapted, and since its path is in D↑ ([0, ∞)), then Ψ· x, B[0,·) , A[0,·] ∈ A.

For any i 6= j ∈ {a, b}, trader jis aware that the other trader
i is rational and hence chooses
i,∗ i i j i
the strategy A ∈ argmaxAi ∈A J x; A , Ψ· x, B[0,·) , A[0,·] . Thus, trader j chooses the max-

imizer Aj,∗ = Ψj· x, B[0,·) , Ai[0,·] , as if observing Ai,∗ , but with no need to actually observe it.
Trader i behaves analogously, leading to the following definition of a Nash equilibrium.

Definition 2.3 (Nash equilibrium). A pair (Ψa , Ψb ) ∈ Λ2 is a Nash equilibrium if, for all
i 6= j ∈ {a, b}:

(i) There exists a unique pair (Aa,∗ , Ab,∗ ) ∈ A2 such that Ai,∗ i j,∗
· = Ψ· x, B[0,·) , A[0,·] .
(ii) Non-cooperative optimality holds, i.e.,

J i x; Ai , Ψj· x, B[0,·) , Ai[0,·] ≤ J i x; Ai,∗ , Aj,∗ for any Ai ∈ A.

The pair (Aa,∗ , Ab,∗ ) ∈ A2 is the equilibrium fraud processes. (See Carmona (2016) for a
discussion of Nash equilibria in stochastic settings.)
Remark 2.4. Note that, in general, a Nash equilibrium does not necessarily coincide with a
best-response map, defined as a pair (Ψa , Ψb ) such that, for any i 6= j ∈ {a, b} and Ai ∈ A,

J x; Ψ· x, B[0,·) , A[0,·] , A = sup J j x; Aj , Ai .
j j i i

(2.2)
Aj ∈A
Put differently, in a Nash equilibrium the response Ψj of trader j does not have to be optimal
in the sense of (2.2), but merely enough to deter trader i from deviating from Ψi .
Note also that condition (ii) in Definition 2.3 is equivalent to J i x; Ai , A
j ≤ J i x; Ai,∗ , Aj,∗

0
0

for any Ψi, ∈ Λ such that there exists (Ai , Aj ) ∈ A2 with Ai· = Ψi,· x, B[0,·) , Aj[0,·] and

Aj· = Ψj· x, B[0,·) , Ai[0,·] . In other words, given trader j’s strategy Ψj , trader i chooses only
among the strategies that lead to the existence of a joint fraud process.
Finally, traders do not need to directly observe each other’s fraud process, but rather infer
it from common knowledge of each other’s rationality. This phenomenon is the same as in the
prisoners’ dilemma, in which each prisoner is questioned separately hence does not observe the
other’s action, and yet rationally anticipates it.
2.2 Construction of Nash equilibrium

In the Nash equilibrium described below, each trader cheats as little as necessary so as to keep
the personal share of wealth above a certain threshold. To rigorously define this behavior, it is
necessary to recall the notion of Skorokhod reflection. Define the set Λ(0,1) similar to Λ by
n
Λ(0,1) = ψ = {ψt : t ≥ 0} where ψt : (0, 1) × C ([0, t])2 × D↑ ([0, t]) → [0, ∞)
is measurable and the path t 7→ ψt (w, f[0,t) , g[0,t] ) is in D↑ ([0, ∞)) ,
o
for all w ∈ (0, 1), f ∈ C ([0, ∞))2 , g ∈ D↑ ([0, ∞)) .
x
For any x = (x1 , . . . , xN ) ∈ RN
+ and any i ∈ {1, . . . , N }, define ri (x) = PN i , so that ri (Ytx )
k=1xk
is trader i’s share of the firm’s capital at time t. Let Wti,wi (Ai , Aj ) = ri Yt (Ai , Aj )
x

for any
i,wi
t ≥ 0 with W0− (Ai , Aj ) = ri (x) = wi .
Definition 2.5 (Skorokhod reflection). Let i 6= j ∈ {a, b} and mi ∈ (0, 1). A ψ i,mi ∈
Λ(0,1) is the solution of the (one-sided) Skorokhod reflection problem (henceforth SPm
i
i+
) if
j i,∗ i,∗
for any A ∈ A andany wi ∈ (0, 1), there exists a unique process A ∈ A, given by A· =
i,mi
ψ· wi , B[0,·) , Aj[0,·] that satisfies
(i) mi ≤ Wti,wi (Ai,∗ , Aj ) < 1 a.s. for all t ≥ 0.

Z
(ii) 1nW i,wi (Ai,∗ ,Aj )>m o dAi,∗
t = 0 a.s..
t i
[0,∞)
By (i), Wti,wi (Ai,∗ , Aj ) almost surely does not take values below mi for any t ≥ 0, while (ii)
means that, as Ai,∗ t increases, Wt
i,wi
(Ai,∗ , Aj ) can reach mi but without spending any positive
amount of time at this point. Note that the uniqueness of Ai,∗ implies the uniqueness of
i,m 0 ,i,m i i,mi j
ψ i in the sense that: if ψ i is another solution of SPmi + , then ψ· wi , B[0,·) , A[0,·] and
0 ,i,m

ψ· i wi , B[0,·) , Aj[0,·] are indistinguishable for any Aj ∈ A and wi ∈ (0, 1).
Clearly, for any i 6= j ∈ {a, b}, W i,wi being reflected at mi towards 1 is the same as the
other trader’s fraction of wealth W j,1−wi being reflected at 1 − mi towards 0, because W i,wi and
1 − W j,1−wi are indistinguishable. Proposition A.10 below proves the existence of the solution
to SPmi .5
i+
At this point, it is necessary to introduce some notation.
Definition 2.6. Suppose Assumption 2.1 holds. For any i 6= j ∈ {a, b}, define the parameter-
dependent threshold
−αi (1 − γi )
ŵi := , (2.3)
γi − αi
where

1
q
αi := 2 2 2
ki − ki + 2σ pi , σ 2 := σa2 + σb2 ,
σ
!
γi σj2 σ2
pi := λ − (1 − γi ) µj − , ki := µj − µi − γi σj2 + .
2 2
Furthermore, set
γi σi2

qi := λ − (1 − γi ) µi − , ai := 1 − γi − αi ,
2

1
q
2
βi := 2 ki + ki + 2σ pi ,2 bi := 1 − γi − βi .
σ
Let ∆ := {(wa , wb ) ∈ (0, 1)2 : wa + wb < 1}, for any i 6= j ∈ {a, b} define F i : ∆ → R by
bi −βi
i wi wj
F (wi , wj ) := ai (αi (1 − γi − wi ) + γi wi )(wj − bi )
1 − wi 1 − wj
ai −αi
wi wj
− bi (βi (1 − γi − wi ) + γi wi )(wj − ai )
1 − wi 1 − wj
+ (ai − bi )(wi (αi + βi − 1) − αi βi )wjγi (1 − wj )1−γi .
Observe that F i depends on the arrival rate of random horizon and all of both traders’
parameters but trader j’s risk-aversion γj . The next result identifies the fraud thresholds used
in Theorem 2.8 below to construct the Nash equilibrium.
Lemma 2.7. There exists (w̃a , w̃b ) ∈ ∆ such that
F a (w̃a , w̃b ) = F b (w̃b , w̃a ) = 0. (2.4)
Moreover, such pair (w̃a , w̃b ) satisfies w̃k < ŵk for all k ∈ {a, b}.
5
In fact, the proposition establishes conditions under which the individual Skorokhod reflections becomes a
two-sided Skorokhod reflection, which ultimately allows to derive a Nash equilibrium.
10

For any k ∈ {a, b} and any mk ∈ (0, 1), let Ψk,mk (x, ·, ·) := ψ k,mk (rk (x), ·, ·) for all x ∈ R2+
where ψ k,mk solves SPm
i
k+
.
Theorem 2.8 (Nash equilibrium). For (w̃a , w̃b ) as in Lemma 2.7, the pair (Ψa,w̃a , Ψb,w̃b ) is
a Nash equilibrium and the corresponding game values satisfy for any i 6= j ∈ {a, b} and any
x ∈ R2+ ,
V i (x; Aj,∗ ) = λ(xa + xb )1−γi ϕi (ri (x)) ,
where
c0 (1 − w)−γi ,
 i

 w ∈ (0, w̃i ), (2.5a)
U i (w)


ϕi (w) = ci1 wαi (1 − w)ai + ci2 wβi (1 − w)bi + w ∈ [w̃i , 1 − w̃j ), (2.5b)

 qi

 i −γi
c3 w , w ∈ [1 − w̃j , 1), (2.5c)
with the constants
−ai bi (1 − w̃i )γi U i (w̃i )
ci0 = > 0,
qi ((αi + βi − 1)w̃i − αi βi )
−bi w̃iai (1 − w̃i )−ai (γi w̃i + (1 − γi − w̃i )βi )
ci1 = > 0,
(1 − γi )qi (βi − αi ) ((αi + βi − 1)w̃i − αi βi )
ai w̃ibi (1 − w̃i )−bi (γi w̃i + (1 − γi − w̃i )αi )
ci2 = > 0,
(1 − γi )qi (βi − αi ) ((αi + βi − 1)w̃i − αi βi )
U i (1 − w̃j )

i γi i αi ai i βi b i
c3 = (1 − w̃j ) c1 (1 − w̃j ) w̃j + c2 (1 − w̃j ) w̃j + > 0.
qi
The proof of this theorem is in Appendix C.
For the purpose of comparative statics, it is useful to consider the case when only one trader
can commit fraud. Indeed, depending on circumstances, access to fraud may be uneven. For
instance, Nick Leeson was able to conceal his unauthorized trades because he was allowed to
settle his own trades (controlling both the front- and the back-office), a privilege that other
traders of the firm did not share. In this regard, assuming that one of the two traders cannot
cheat, the other trader maximizes expected utility by cheating in a similar way, but with a
different fraud threshold:
Theorem 2.9 (Solo Rogue Trader). For any i 6= j ∈ {a, b}, if Aj ≡ 0, then the optimal fraud
process for trader i is Ai,∗ = Ψ·i,ŵi x, B[0,·) , 0 and the corresponding value function satisfies

V i (x; 0) = λ(xa + xb )1−γi ϕ̂i (ri (x))

for any x ∈ R2+ , where
 i −γ
 s0 (1 − w) i on (0, ŵi ), (2.6a)
i
ϕ̂ (w) =
 si1 wαi (1 − w)ai + U i (w)
on [ŵi , 1), (2.6b)
qi
with
ai
ai (1 − ŵi )γi i

1 − γi − ŵi ŵi
si0 = U (ŵi ) > 0 and si1 = > 0.
qi (ŵi − αi ) (1 − γi )qi (ŵi − αi ) 1 − ŵi
Remark 2.10. Because ŵi > w̃i (Lemma 2.7), a sole cheater is less tolerant of decreasing
wealth than if there was another cheater present. The fraud region of trader i is indeed smaller
in the Nash equilibrium, where both cheat as little as possible (Theorem 2.8), so as to keep their
proportion of wealth above w̃i . Furthermore, limγi ↑1 ŵi = 0 follows by (2.3) and, due to ŵi > w̃i
then limγi ↑1 w̃i = 0, which shows that fraud is undesirable under log-utility for both a sole cheater
and dual cheaters because bankruptcy becomes totally unacceptable (i.e. log(0) = −∞).
11

1 1 1 1
0.9 0.9 0.9 0.9
0.8 0.8 0.8 0.8
0.7 0.7 0.7 0.7
0.6 0.6 0.6 0.6
0.5 0.5 0.5 0.5
0.4 0.4 0.4 0.4
0.3 0.3 0.3 0.3
0.2 0.2 0.2 0.2
0.1 0.1 0.1 0.1
0 0 0 0
0% 10% 20% 30% 40% 50% 60% 20% 40% 60% 80% 100%
a a
1 1 1 1
0.9 0.9 0.9 0.9
0.8 0.8 0.8 0.8
0.7 0.7 0.7 0.7
0.6 0.6 0.6 0.6
0.5 0.5 0.5 0.5
0.4 0.4 0.4 0.4
0.3 0.3 0.3 0.3
0.2 0.2 0.2 0.2
0.1 0.1 0.1 0.1
0 0 0 0
0 0.2 0.4 0.6 0.8 1 5 10 15
a 1/ (Average horizon)
Fig. 1: Fraud thresholds for trader a (blue) and b (red), as trader a’s share of wealth (vertical
axis), in Nash equilibrium (solid line), and when the other trader is honest (dashed line),
against trader a’s expected return (upper-left, 0% ≤ µa ≤ 60%), volatility (upper-right, 0% <
σa ≤ 100% ), risk-aversion (bottom-left, 0 < γa < 1), and average horizon (bottom-right,
0 < 1/λ ≤ 20 ). Other parameters are µa = µb = 10%, σa = σb = 20%, γa = γb = 0.5, λ = 1/3.
3 Discussion
This section brings to life the theoretical results in Section 2.2 by examining the properties of
the Nash equilibrium for concrete parameter values.
Figures 1 and 2 display the dependence of the fraud thresholds and the average amount of
fraud of each trader on model parameters, respectively. A trader’s fraud threshold is relatively
insensitive to the profitability of personal investments (Figure 1, upper left), even as such
profitability increases from 10% to 60%. The flatness of the threshold, however, does not imply
flatness of average fraud, which instead declines rapidly as profitability increases (figure 2, upper
left). The explanation of this phenomenon lies in the dynamics of relative wealth shares: when
one trader’s profitability is high, that trader’s wealth share tends to increase over time, thereby
reaching the fraud threshold less often, hence generating lower fraud.
By contrast, the fraud threshold of the other trader (whose profitability remains constant)
12

rapidly withdraws upwards, meaning that this trader cheats when the respective wealth share
falls below a lower threshold. Again, this fact does not imply a decline in the amount of fraud,
because such trader’s typical wealth share also tends to decline. In fact, figure 2 shows that
the amount of fraud first increases (up to µa ≈ 30%) and then decreases: The initial rise is
understood as a short-term appropriation, whereby the less skilled trader’s higher fraud pilfers
the other’s profits. The subsequent decline is more akin to a long-term appropriation: the less
skilled trader recognizes that the other’s skill is so high that it is overall more profitable to limit
the amount of fraud per unit of time, as to let the other’s wealth grow faster, so that future
fraud can be even more profitable. Put differently, the less skilled trader establishes a sort of
parasite-host relationship with the more skilled trader, thereby avoiding to cheat excessively,
lest the host perish. Note also that the threshold of the more skilled trader is more sensitive to
the honesty (or lack thereof) of the other trader, while the less skilled trader becomes indifferent
to the other’s honesty when the profitability is sufficiently high.
As the volatility of a trader’s investments increases (upper right, figures 1 and 2), that
trader’s fraud threshold recedes aggressively, but fraud increases significantly. Increased volatil-
ity is qualitatively similar to lower skill, which makes the trader more reliant on fraud to generate
profits. Vice versa, the other trader can still rely on a personal payoff with lower volatility, which
would be significantly degraded by the additional asymmetry generated by more fraud.
Risk aversion (lower left, figures 1 and 2) has a major impact on propensity for fraud.
Holding the opponent’s risk aversion constant at 0.5, as a trader’s risk aversion increases from
zero to one, the fraud threshold declines very rapidly from one (incessant fraud) to zero (no
fraud). Note that, as a fraud threshold declines, the other threshold also declines, not to
zero, but to the threshold that assumes the other’s honesty. Put differently, a fearless trader’s
propensity for fraud forces the other, more prudent, trader to withdraw from fraud, as overall
risk is already too high. The implication is that, when the two traders have very different risk
aversion but similar investment opportunities, it is the least risk averse that has most potential
for fraud. Vice versa, when risk aversions are similar, the overall potential for fraud is much
lower and is evenly distributed between traders.
Fraud completely disappears with unit risk aversion (i.e. logarithmic preferences). In this
case, the dread of bankruptcy is so high that traders abstain from fraud regardless of its potential
rewards. Note that this phenomenon stems from the fraud’s inherent discontinuity, which always
implies a probability, however small, that wealth may vanish. Put differently, for the logarithmic
investor the marginal utility of any amount of fraud is infinitely negative, regardless of expected
profits.
The average horizon is also an important determinant of fraud (lower right, figures 1 and 2).
Fraud thresholds recede as the horizon increases (λ decreases) and with it the expected reward
for delaying fraud. In fact, the average amount of fraud tends to increase up to a horizon
of about ten years, declining thereafter. The implication is that increasing the horizon has an
ambiguous effect on fraud: while reducing it per unit of time, its overall amount in fact increases
in the medium term, and its reduction requires rather long horizons, higher than typical staff
turnover in financial institutions.
4 Conclusion
This paper develops a structural model of rational rogue trading. Self interested, risk averse
traders can deliberately engage in fraudulent trading activity that can be concealed as superior
performance while successful, but may lead to a firm’s bankruptcy if unsuccessful. Traders ab-
stain from fraud when they have sufficient skin in the game, suggesting that effective mitigation
of rogue trading episodes should not focus on large traders alone.
13

0.06
0.15
0.05
0.04
0.1
0.03
0.02 0.05
0.01
0 0
0% 10% 20% 30% 40% 50% 60% 20% 40% 60% 80% 100%
a a
0.035
1.2
0.03
1
0.025
0.8
0.02
0.6
0.015
0.4
0.01
0.2 0.005
0 0
0.2 0.4 0.6 0.8 5 10 15
a 1/ (Average horizon)
Fig. 2: Equilibrium average fraud (vertical axis), up to horizon or bankruptcy, of traders a

(blue) and b (red) against trader a’s expected return (upper-left, 0% ≤ µa ≤ 60%), volatility
(upper-right, 0% < σa ≤ 100% ), risk-aversion (bottom-left, 0.1 < γa < 0.9), and average
horizon (bottom-right, 0 < 1/λ ≤ 20). Results obtained from simulation of 104 paths, each
with step size 5 · 10−4 . Other parameters are µa = µb = 10%, σa = σb = 20%, γa = γb = 0.5,
wa = wb = 0.5, λ = 1/3.
A Appendix
A.1 The wealth of Rogue Traders
We start by recalling the definition of the stochastic exponential (cf. (Jacod and Shiryaev, 2013,
Eq. I.4.62)) of general semimartingales:
Definition A.1. For any R-valued semimartingale S, the stochastic exponential of S is the
process Y
1 c
E(S)t := exp St − S0 − [S ]t e−∆Ss (1 + ∆Ss ) t≥0
2
0≤s≤t
where E(S)0− = 1 and Sc denotes the continuous part of S.
14

All stochastic exponentials in this article are P-a.s. strictly positive because the jump sizes
are bounded away from −1 (cf. (Jacod and Shiryaev, 2013, Theorem I.4.61 (c)). For finite-
variation jumps (as in this paper), note that
X
St = Stc + ∆Ss ,
0≤s≤t
c = 0, therefore the stochastic exponential simplifies to

with S0−
Y
c 1 c
E(S)t = exp St − [S ]t (1 + ∆Ss )
2
0≤s≤t
The following result shows that the pre-bankruptcy wealth (1.2) is well-defined and provides
an expression in terms of the stochastic exponential.
Lemma A.2.
x 1,x N,x to the SDE (1.2) and for all

(i) There exists a unique strong solution Y = Y , . . . , Y
i ∈ {1, . . . , N }, P Yti,x > 0 for all t ≥ 0 = 1.
(ii) For all i ∈ {1, . . . , N } and any t ≥ 0

Z !
Yti,x = xi E µi · +σi B·i + x −1
ri (Ys− ) dÃis a.s. (A.1)
[0,·]
t
and
N N Z
! !
· Z ·
YtS,x
X X
= xk E µk rk (Ysx )ds + σk rk (Ysx )dBsk + ÃS· a.s. (A.2)
k=1 k=1 0 0
t
PN
where ÃS· = k
k=1 Ã· .
Proof of Lemma A.2. Denote by IN an N × N identity matrix. The SDE (1.2) can be written
in the vector form,
dYt = diag(Yt )dRt + trace (diag(Yt− )) IN dÃt , Y0− = x ∈ Rn+ , (A.3)
where R = (R1 , . . . , RN ) with Rti = µi t + σi Bti for any i ∈ {1, . . . , N } and Ã = (Ã1 , . . . , ÃN ).
The linearity of the coefficients of (A.3) implies uniform Lipschitz continuity, hence the existence
and uniqueness of a strong solution (cf. (Cohen and Elliott, R 2015,
P Theorem 16.3.11)).
j
For any i ∈ {1, . . . , N }, let Zti = Rti + Ãit and Hti = xi + [0,t] N Y
j6=i s− dÃ i for all t ≥ 0 with
s
i = 0 and H i = x . Rewriting (1.2) yields
Z0− 0− i
Z
i i i
Yt = Ht + Ys− dZti .
[0,t]
By (Jacod, 2006, Theorem 6.8), it follows that

Z !
Yti = E(Z i )t xi + E(Z i )−1 i
s− dH̄s , (A.4)
[0,t]
where
X ∆H i ∆Ãi X ∆H i Z N
= Hti,c + j
X
s s s
H̄ti = Hti − = xi + Ys− dĀis (A.5)
i
1 + ∆Ãs i
1 + ∆Ãs
0≤s≤t 0≤s≤t [0,t] j6=i
15

∆Ãi
and Āit = Ai,c with Āi0− = 0. Substituting (A.5) into (A.4) yields
P
t + 0≤s≤t 1+∆Ãi
 
Z N
j
X
Yti = E(Z i )t xi + E(Z i )−1
s− Ys− dĀis  . (A.6)
[0,t] j6=i
Let τ0 = inf t ≥ 0 : mink∈{1,...,N } Ytk ≤ 0 be the first exit time of the process mink∈{1,...,N } Y k

from R+ . Suppose P (0 ≤ τ0 < ∞) > 0, then for any ω ∈ {ω : 0 ≤ τ0 (ω) < ∞}, there exists
q ∈ {1, · · · , N } such that Yτq0 (ω) (ω) ≤ 0 and Yτq0 (ω)− (ω) ≥ 0 as Y q is càdlàg. Since xq > 0, (A.6)
implies
N
X
Ysj (ω) < 0
j6=q
for some s < τ0 (ω), which contradicts the definition of τ0 . This finishes the proof of (i).
We thus may rewrite (1.2) and the firm’s total pre-bankruptcy wealth Y S as
S
dYti i Yt−
i
= dRt + i
dÃit , i
Y0− = xi ,
Yt− Yt−
N N
dYtS X Yk
t
X
S
= S
dRtk + dÃSt , S
Y0− = xk .
Yt− Y
k=1 t k=1
An application of (Jacod, 2006, Theorem 6.8) yields (ii).
The following lemma characterizes the conditional probability of bankruptcy in relation to

total frauds.
Lemma A.3. The following hold for all t ≥ 0, P-almost surely:
(i)
S
P (τA > t|Ft ) = e−At , (A.7)
(ii)
P (τA > t|F∞ ) = P (τA > t|Ft ) . (A.8)
Proof. First we show that

{t < τA } = {ASt < θ}. (A.9)
On one hand, {t < τA } ⊂ {ASt < θ} follows by the definition of τA . On the other hand, let
ω ∈ Ω be such that ASt (ω) < θ(ω). If τA (ω) < +∞, then θ(ω) ≤ ASτA (ω) (ω). Hence, t < τA (ω)
because ASt (ω) < ASτA (ω) (ω) and AS is non-decreasing. If τA (ω) = +∞, then trivially t < τA (ω).
Because P(θ > x) = e−x for any x ≥ 0 and θ is independent of F∞ ,
S
P (τA > t|F∞ ) = P ASt < θ|F∞ = e−At .

Finally, (A.7) and (A.8) follow by the tower property of conditional expectation and AS· being
F-adapted.
The following result shows that the treatment of jumps in the definition of Ã is the only one
consistent with the martingale property for wealth in the absence of skill.
Proposition A.4. Let N = 1, µ1 = 0 and let Ã1t := A1,c 1

P
t + 0≤s≤t g ∆As where g : [0, +∞) →
[0, +∞) is measurable such that g(0) = 0.
(i) If A1t = A1,c a

t a.s. for all t ≥ 0, or if g(a) = e − 1, then X
1,x1 is a GA -martingale.
16

(ii) If X 1,x1 is a GA -martingale for any A1 ∈ A, then the jump size function g(a) = ea − 1
for any a ≥ 0.
Proof of Proposition A.4. Proof of (i): By (Aksamit and Jeanblanc, 2017, Lemma 3.8), (A.8)
implies the immersion property, that is, an F-martingale remains a martingale on the enlarged
filtration GA . It then follows by Lévy’s characterization theorem, that B 1 remains a Brownian
motion on GA . Lemma A.2 (ii) and (Cohen and Elliott, 2015, Corollary 15.1.9) yield that for
any t ≥ 0

Xti,x = xi 1{t<τA } E σ1 B·1 + Ã1· = xi 1{t<τA } E Ã1· E σ1 B·1 t a.s.

t t

1{t<τA } eAt is a GA -
1 1
If A1 has a.s. continuous paths or g(a) = ea − 1, then E Ã1· = eAt , and
t
martingale by (Bielecki and
Rutkowski, 2004, Lemma 5.1.7.). Because the covariation between
1{t<τA } e and E σ1 B· t is zero and Xt = Xt∧τ
A 1 1 1,x1 1,x1
t
A
a.s. for all t ≥ 0, it follows that X i,x is a
A
G -martingale.
Proof of (ii): Consider the family Aξ of strategies indexed by ξ ≥ 0, defined for t ≥ 0 by
ξ
At = 1{t≥1} ξ. By construction Aξ ∈ A for all ξ ≥ 0. Denote the corresponding wealth as X 1,x1 ,ξ .
By assumption, it is a GA -martingale for any ξ ≥ 0. One can factorize it as X 1,x1 ,ξ = Mt Ut ,
where for any t ≥ 0
Mt = x1 1{t<τA } eAt E σ1 B·1 t
1
(A.10)
and
1
Y
e−∆As 1 + g(∆A1s ) = 1t<1 + e−ξ (1 + g(ξ))1{t≥1} .

Ut =
0≤s≤t∧τA
Note that M is a GA -martingale and by Lemma A.5, the finite variation process U is GA -
predictable. Integration by parts (Aksamit and Jeanblanc, 2017, Proposition 1.16) yields
Z Z
1,x1
Xt = Mt U t = x + Us dMs + Ms− dUs .
[0,t] [0,t]
R· R·
The process 0 Us dMs is a GA -(local)martingale. As the process 0 Ms− dUs is the limit of
Riemann-Stieltjes sums, it inherits the G A
1,x ,ξ A
R · -predictability Aof finite variation from its integrator
U . Because X 1 is a G -martingale, R · 0 Ms− dUs is a G -local martingale. Then, by (Cohen
and Elliott, 2015, Lemma 10.3.9.), 0 Ms− dUs is a constant. Since M1− > 0 with positive
probability6 and (
Z ·
0, t<1
Ms− dUs = −ξ
,
0 M1− (e (1 + g(ξ)) − 1), t≥1
we conclude that e−ξ (1 + g(ξ)) = 1, for all ξ ≥ 0.
A.2 Proof of Proposition 1.1

By Lemma A.2, trader 1’s wealth has the expression
Xt1,x1 = 1{t<τA } x1 eAt +(µ1 −σ1 /2)t+σ1 Bt ,
1 2 1
t ≥ 0.
Then, by Lemma A.3, it follows that
h i h i
E U 1 (Xt1,x1 )|Ft = E 1{t<τA } U 1 (Yt1,x1 )|Ft = e−At U 1 (Yt1,x1 )
1
x1−γ 1
1 2 1 x1−γ1 (1−γ1 )((µ1 −σi2 /2)t+σ1 Bt1 )
= 1
e−γ1 At +(1−γ1 )((µ1 −σi /2)t+σ1 Bt ) ≤ 1 e .
1 − γ1 1 − γ1
6
In (A.10) all quantities
except
the indicator function are strictly positive; and P (τA ≥ 1) > 0 because, in
ξ

view of (A.9), 0 < P A1 < θ ≤ P ∩ε∈(0,1) {A1−ε < θ} = P ∩ε∈(0,1) {τA > 1 − ε}
17

Therefore, by the tower property of conditional expectation,
h i x1−γ1 (1−γ1 )(µ1 −γ1 σ12 /2)t
E U 1 (Xt1,x1 ) ≤ 1 e (A.11)
1 − γ1
and
h ix1−γ 1
2
E U 1
(Xt1,x1 )
= 1 e(1−γ1 )(µ1 −γ1 σ1 /2)t if and only if At = 0 a.s.. (A.12)
1 − γ1
Let Pτ1 be the law of τ1 , i.e. Pτ1 (U ) = P(τ1 ∈ U ) for any τ1 -measurable set U ⊂ R+ . Then, by
the law of total probability and the independence of τ1 from B and θ,
Z ∞ Z ∞ h i
E U 1 (Xt1,x1 ) dPτ1 (dt).
1 1,x1 1 1,x1
E U (Xτ1 ) = E U (Xτ1 )|τ1 = t dPτ1 (dt) = (A.13)
0 0
Thus, (A.11) implies that
x1−γ x1−γ1 h (1−γ1 )(µ1 −γ1 σ12 /2)τ1 i
1
Z ∞
2
e(1−γ1 )(µ1 −γ1 σ1 /2)t Pτ1 (dt) = 1
1 1,x1 1
E U (Xτ1 ) ≤ E e (A.14)
1 − γ1 0 1 − γ1
and, due to (A.13) and (A.12), the equality
x1−γ 1 h
(1−γ1 )(µ1 −γ1 σ12 /2)τ1
i
E U 1 (Xτ1,x 1

1
) = E e
1
1 − γ1
1

holds if and only if P At = 0 = 1 for all t > 0 for which P (τ1 ≥ t) > 0. (Suppose, by
1
contradiction, that there exists some t0 ≥ 0 for which P (τ1 ≥ t0 ) > 0, but P At0 > 0 > 0.
Because A1 is non-decreasing a.s., for all t ≥ t0 , {At ⊇ At0 } is an event of probability one, and
(A.12) implies that
h i x1−γ1 (1−γ1 )(µ1 −γ1 σ12 /2)t
E U 1 (Xt1,x1 ) < 1 e , t ≥ t0 ,
1 − γ1
and since P (τ1 ≥ t0 ) > 0, we have upon integration (cf. (A.13)) indeed strict inequality in
(A.14).
A.3 Doob-Meyer decomposition

Let ĀSt = AS,c −∆ASs ) for all t ≥ 0 and note that ĀS ∈ A. In the absence of
P
t + 0≤s≤t (1 − e
fraud jumps, the total fraud process AS is the GA -compensator of the bankruptcy time τA . In
the presence of jumps, the compensator is in fact ĀS , as shown in the following lemma.
Lemma A.5 (Doob-Meyer decomposition). The process M A = (MtA )t≥0 defined by
MtA = 1{t≥τA } − ĀSt∧τA
is a uniformly integrable GA -martingale. Furthermore, ĀS·∧τA is the unique GA -predictable,
integrable and non-decreasing process such that M is a GA -martingale and ĀS0− = 0.
Proof. Note that the non-decreasing process 1{·≥τA } is a GA -submartingale. Define the stochas-
tic process Z = (Zt )t≥0 by Zt := P(t < τA |Ft ). Because F ⊂ GA and all F-martingales are
continuous (by the Martingale Representation Theorem (Karatzas and Shreve, 1998, Theorem
4.2)), the dual F-predictable projection of 1{·≥τA } is 1 − Z· by (Aksamit and Jeanblanc, 2017,
Proposition 3.9 (b)). It follows by (Aksamit and Jeanblanc, 2017, Proposition 2.15) that the
A −1
R
G -compensator of τA is [0,·∧τA ] Zs− d (1 − Zs ). Then by Lemma A.3 and the Itô formula,
Z Z
−1 S S
Zs− d (1 − Zs ) = eAs− d −e−As−
[0,t∧τA ] [0,t∧τA ]
X S

= AS,c
t∧τA + 1 − e−∆As a.s. for all t ≥ 0.
0≤s≤t∧τA
18

Lemma A.6. The bankruptcy time τA is GA -predictable if and only if ASt∧τA = AS0 a.s. for any
t ≥ 0.
Proof. Suppose τA is a GA -predictable stopping time. It follows that the process 1{·≥τA } is GA -
predictable. Lemma A.5 implies that the GA -martingale Mt is GA -predictable. A predictable
martingale of finite variation must be constant, hence, MtA = M0A for any t ≥ 0. It follows that
1{t≥τA } − 1{0=τA } = ĀSt∧τA − ĀS0 a.s.

for any t ≥ 0. On the event {0 < τA < +∞}, for any t < τA , ĀSt∧τA − ĀS0 = 0 a.s. and then
S
∆ĀSτA = 1 a.s., which contradicts the fact that ∆ĀSt = 1 − e−At < 1. Hence, P(0 < τA <
+∞) = 0. On the events {τA = 0} and {τA = ∞}, it is clear that ĀSt∧τA = ĀS0 a.s. (and thus
ASt∧τA = AS0 .)
Conversely, let for any t ≥ 0, ASt∧τA = AS0 a.s. Then, by definition of bankruptcy, τA ∈
{0, +∞} a.s. As the events {τA = 0} and {τA = ∞} are in G0A , the sequence τn of increasing
GA -stopping times defined by τn = 1{τA =0} + n1{τA =+∞} announces τA .
Remark A.7. The preceding Lemma implies that the bankruptcy time is announced by a strictly
increasing sequence of stopping times if and only if either in the absence of fraud or if any
fraudulent trades are performed only at the initial time t = 0.
A.4 The Value Function (Proof of Lemma 2.2)

Lemma A.8. For any p ∈ (0, 1) and t ≥ 0,
N
!p
h p i
E XtS,x
X
≤ xk ept maxk∈{1,...,N } µk . (A.15)
k=1
Proof. Using the expression of Y S,x as in (A.2), since the covariation between the terms inside
E(·) in (A.2) is zero, by (Cohen and Elliott, 2015, Corollary 15.1.9) it follows that
N N N
! Z · k,x ! Z · k,x !
S,x
X
S
X Ys X Ys k
Yt = xk E Ã· E µk S,x
ds E σk S,x
dBs
k=1
t
k=1 0 Ys t k=1 0 Y s t
and, upon rearranging and estimating terms,

N N N
! Z · k,x ! Z · k,x !
−1 Ys Ys
S,x
X X X
S k
Yt E Ã· = xk E µk S,x
ds E σk S,x
dBs
t
k=1 k=1 0 Y s t k=1 0 Y s t
N N Z · N
! ! !
X k,x Z · k,x
X Ys X Ys
≤ xk E max µk S,x
ds E σk S,x
dBsk
k∈{1,...,N } Y Y
k=1 k=1 0 s t k=1 0 s t
N N
! !
Z · k,x
X X Ys
= xk e(maxk∈{1,...,N } µk )t E σk S,x
dBsk . (A.16)
k=1 k=1 0 Y s t
For any k ∈ {1, . . . , N } and any t ≥ 0, as

 
k,x 2
Z t !
2
σk
2
σk
Ys
exp  ds ≤ exp
 t <∞
2 0 YsS,x 2
a.s. so the Novikov’s criterion holds (Revuz and Yor, 2013, Corollary 1.16) and hence,
N Z · k,x !
X Ys k
E σk S,x
dBs
k=1 0 Ys t≥0
19

is a true martingale. Taking expectation on both sides of (A.16) yields
N
!
−1
S,x
X
E Yt E ÃS· ≤ xk e(maxk∈{1,...,N } µk )t . (A.17)
t
k=1
The stochastic exponential E(ÃS )t satisfies

N
!
AS,c

ÃS S ∆Aks
Y Y X
S
E(Ã )t = e t (1 + ∆ÃSs )e−∆Ãs =e t 1+ e −1
0≤s≤t 0≤s≤t k=1
AS,c
PN k AS
Y
≤e t e k=1 ∆As =e t , (A.18)
0≤s≤t
PN
where the inequality (A.18) follows from e k=1 yk −1 ≥ N yk N
P
k=1 (e −1) for any (y1 , . . . , yN ) ∈ R+ .
1 N
Note that the equality of (A.18) is strict if and only if the fraud processes (A , . . . , A ) satisfy
P ∆Ais ∆Ajs = 0 = 1, 0 ≤ s ≤ t, for all i 6= j ∈ {1, . . . , N }.

(A.19)
By Lemma A.3 (i), for any t ≥ 0
h i h i h i h i
E XtS,x = E 1{t<τA } YtS,x = E YtS,x E 1{t<τA } |Ft = E e−At YtS,x .
S
(A.20)
Therefore, by the estimate (A.18),

h i −1
S,x S,x
E Xt ≤ E Yt E ÃS· (A.21)
t
and the strict equality holds if and only if condition (A.19) holds.
Finally, it follows by Jensen’s inequality, (A.21) and (A.17) that for any 0 < p ≤ 1,
N
!p
h p i h ip
S,x S,x
X
E Xt ≤ E Xt ≤ xk ept maxk∈{1,...,N } µk .
k=1
Proof of Lemma 2.2. By the independence between τ and F∞ ∨ σ(θ), the tower property of
the conditional expectation, and Lemma A.3, it follows that for any i 6= j ∈ {a, b} and any
Ai , Aj ∈ A,
E U i Xτi,x (Ai , Aj ) = E E U i Xτi,x (Ai , Aj ) |F∞ ∨ σ(θ)

Z ∞
−λt i i,x i j
=E λe U (Xt (A , A ))dt
0
Z ∞
e E 1{t<τA } |Ft U (Yt (A , A ))dt
−λt
i i,x i j
= λE
Z0 ∞
−λt−AS i i,x i j
= λE e t U (Y
t (A , A ))dt
0
which proves (i).

Furthermore, the scale-invariance (ii) is an immediate result of Y i,cx and cY i,x being indis-
tinguishable (implied by Lemma A.2 (ii)).
Finally, Tonelli’s theorem and Lemma A.8 yield
Z ∞ 1−γi
i i j λ −λt i,x i j
J (x; A , A ) = e E Xt (A , A ) dt
1 − γi 0
λ (xa + xb )1−γi ∞ −λt+(1−γi )(µa ∨µb )t λU i (xa + xb )
Z
≤ e dt = .
1 − γi 0 λ − (1 − γi )(µa ∨ µb )
Taking the supremum over Ai ∈ A, the proof of (iii) follows.
20

A.5 Skorokhod Reflection
First we study the SDE that identifies the fractions of wealth of each trader. For this application,
assume for the rest of this chapter N = 2 traders. For any i 6= j ∈ {a, b}, define the coefficient
functions (of a single variable)
b̄i (w) := w(1 − w) σj2 (1 − w) − σi2 w + µi − µj ,

(
(σa w(1 − w), −σb w(1 − w)) if i = a,
σ̄i (w) :=
(−σa w(1 − w), σb w(1 − w)) if i = b,
respectively. Introduce also the processes
Q̃it = Ai,c
X
t + qi (∆Ãs ) and Q̃S = Q̃1 + Q̃2 ,
0≤s≤t
ai
where qi : R2+ → R+ is defined as qi (a1 , a2 ) = 1+a1 +a2 .
Lemma A.9. The following statements hold:
(i) For any i 6= j ∈ {a, b}, the traders’ pre-bankruptcy wealth shares ri (Y·x ) is the unique
strong solution to the SDE
Z t Z t Z Z
Wti,wi = wi + b̄i (Wsi,wi )ds + σ̄i (Wsi,wi )dBs + i,wi
(1 − Ws− )dQ̃is − i,wi
Ws− dQ̃js .
0 0 [0,t] [0,t]
(A.22)
with wi = ri (x).
(ii) For all t ≥ 0, Wti,wi ∈ (0, 1) a.s.
Proof. (i): An application of Itô formula shows that the fractions ri (Y·x ) satisfy the SDE (A.22)
and the uniqueness follows by local Lipschitz continuity of its coefficients (cf. (Cohen and Elliott,
2015, Theorem 16.3.11.)). Furthermore, the strict positivity of the pre-bankruptcy wealth Y x
(Lemma A.2) proves (ii).
The next statement establishes the existence of solution to the Skorokhod reflection problem
i
SPm :
i+
Proposition A.10 (Skorokhod reflection problem). Let i 6= j ∈ {a, b} and mi ∈ (0, 1). There
i,∗ j
exists ψ i,mi ∈ Λ(0,1) that solves SPm
i
i+
. The unique process A · := ψ i,mi w , B
i [0,·) , A[0,·] is:
for any t ≥ 0, P-a.s.,
P i,c,∗ i+

1 h ∆Ajt
Ai,c,∗
t = t , ∆Ai,∗
t = ln 1 + mi e i,wi i,∗ j
− Wt− (A , A ) (A.23)
1 − mi 1 − mi
where P0i,c,∗ = 0 and for t ≥ 0,

Z t
i,c,∗
b̄i Wsi,wi (Ai,∗ , Aj ) ds

Pt = sup mi − wi −
0≤s≤t 0
Z t Z +
i,wi
σ̄i Wsi,wi (Ai,∗ , Aj ) dBs + (Ai,∗ , Aj )dĀjs

− Ws−
0 [0,t]
i+
−∆Ajs
X h i,wi
− mi − e Ws− (Ai,∗ , Aj ) . (A.24)
0≤s≤t
21

a,∗ b,∗ 2
Furthermore, if ma + mb < 1 then there exists a unique pair (A , A ) ∈ A such that
for any i 6= j ∈ {a, b}, Ai,∗ = ψ·i,mi wi , B[0,·) , Aj,∗
[0,·] with wi + wj = 1 (known as two-sided
Skorokhod reflection problem). In this case, the expression of Ai,∗ simplifies to
Pti,c,∗
+
1 − wi
Ai,c,∗
t = , ∆Ai,∗
t = 1{t=0} ln , (A.25)
1 − mi 1 − mi
where
Z t
Pti,c,∗ b̄i Wsi,wi (Ai,∗ , Aj,∗ ) ds

= sup mi − min {wi , 1 − mj } −
0≤s≤t 0
Z t +
σ̄i Wsi,wi (Ai,∗ , Aj,∗ ) mj )Aj,c,∗ − [mi − wi ]+

− dBs + (1 − t (A.26)
0
for any i 6= j ∈ {a, b}.

Proof. Note that if Ai ≡ 0 then the process Wti,wi (0, Aj ) i,wi
with W0− (0, Aj ) = wi ∈ (0, 1)
t≥0
satisfies
Z t Z t
Wti,wi (0, Aj ) = wi + b̄i (Wsi,wi (0, Aj ))ds + σ̄i (Wsi,wi (0, Aj ))dBs
Z 0 0
i,wi j j
− Ws− (0, A )dĀs . (A.27)
[0,t]
Slightly generalize the SDE (A.27) by adding a process P i ∈ A: for any wi ∈ (0, 1) and t ≥ 0,
Z t Z t
Wti,wi (P i , Aj ) = wi + b̄i (Wsi,wi (P i,∗ , Aj ))ds + σ̄i (Wsi,wi (P i,∗ , Aj ))dBs
Z 0 0
i,wi i,∗ j j i
− Ws− (P , A )dĀs + Pt . (A.28)
[0,t]
Note that W·i,wi (P i , Aj ) is not necessarily bounded above by 1 depending on the process P i .
By (De Angelis and Ferrari, 2018, Lemma 2.2.), there exists a map i,mi ∈ Λ(0,1) (which is the
ρ
limit of Picard iterations) such that P·i,∗ = ρi,m
·
i
wi , B[0,·) , Aj[0,·] ∈ A is the unique process in
A satisfying
(i) mi ≤ Wti,wi (P i,∗ , Aj ) < 1 a.s. for any 0 ≤ t < τ1 ,

Z
(ii) 1nW i,wi (P i,∗ ,Aj )>m o dPti,∗ = 0 a.s.,
t i
[0,τ1 )
where the exit time τ1 = inf{t ≥ 0 : Wti,wi (P i,∗ , Aj ) ≥ 1}. Moreover, P i,∗ has the expression
Z t Z t
Pti,∗ = sup mi − wi − b̄i (Wsi,wi (P i,∗ , Aj ))ds − σ̄i (Wsi,wi (P i,∗ , Aj ))dBs
0≤s≤t 0 0
Z +
i,wi
+ Ws− (P i,∗ , Aj )dĀjs
[0,t]
i,∗
a.s. for any t ≥ 0 with P0− = 0. Let P i,c,∗ denote the continuous part of the process P i,∗ . Then
there exists a unique process with continuous path (Ai,c,∗ t )t≥0 ∈ A such that
Z t
i,c,∗
Pt = (1 − Wsi,wi (P i,∗ , Aj ))dAi,c,∗
s (A.29)
0
22

or equivalently
Z t
i,c,∗
At = (1 − Wsi,wi (P i,∗ , Aj ))−1 dPsi,c,∗
0
Z t −1
= 1{W i,wi (P i,∗ ,Aj )>m } + 1{W i,wi (P i,∗ ,Aj )=m } 1 − Wsi,wi (P i,∗ , Aj ) dPsi,c,∗
s i s i
0
Pti,c,∗
= , (A.30)
1 − mi
where the third equality follows by condition (ii).
Note that for any t ≥ 0, ∆Wti,wi (P i,∗ , Aj ) = ∆Pti,∗ −Wt− i,wi
(P i,∗ , Aj )∆Ājt , a.s. Condition (ii)
i,∗ i,wi
implies that if ∆Pt (ω) > 0 for some ω ∈ Ω, then Wt (P i,∗ , Aj )(ω) = mi for any t ∈ [0, τ1 ),
which in turn implies that
j
∆Pti,∗ (ω) = mi − e−∆At (ω) Wt−
i,wi ,∗
(P i,∗ , Aj )(ω) > 0.
Otherwise, if ∆Pti,∗ (ω) = 0 for some ω ∈ Ω, then Wti,wi (P i,∗ , Aj )(ω) ≥ w̃i . Hence,
h j
i+
∆Pti,∗ = mi − e−∆At Wt−i,wi ,∗
(P i,∗ , Aj ) (A.31)
a.s. for any t ≥ 0.

ai (1+aj −w)
Let pi,j (ai , aj , w) = (1+aj )(1+ai +aj ) for any (ai , aj , w) ∈ R20,+ × (0, 1). Because ai 7→
pi,j (ai , aj , w) is strictly increasing, for any t ≥ 0 there exists a unique random variable ∆Ai,∗
t
such that

∆Pti,∗ = pi,j ∆Ãi,∗
t , ∆ Ã j
t , W i,wi
t− (P i,∗
, Aj
) , (A.32)
or, equivalently,
∆Pti,∗ (∆Ãjt + 1)
∆Ãi,∗
t = (A.33)
1 − ∆Pti,∗ − (∆Ãjt + 1)−1 Wt−i,wi
(P i,∗ , Aj )
1 h j
i+
= mi e∆At − Wti,wi (P i,∗ , Aj ) ,
1 − mi
where the second equality follows by (A.31). Finally, defining the process Ãi,∗ t = Ai,c,∗t +
P i,∗ i,wi i,∗ , Aj ))
0≤s≤t ∆ Ãs , and substituting (A.29) and (A.32) into (A.28) shows that (W t (P t≥0
solves the SDE (A.22). Define the process Ai,∗ i,c,∗ i,∗ i,∗
P
t := A t + 0≤s≤t ∆A s such that ∆ Ã t =
i,∗
e∆At − 1, then (A.23) is satisfied. As for any (Ai , Aj ) ∈ A2 the solution to the SDE (A.22)
with initial data wi ∈ (0, 1) never leaves the interval (0, 1) with probability 1, it follows that
τ1 = +∞. To complete the proof, set Wti,wi (Ai,∗ , Aj ) = Wti,wi (P i,∗ , Aj ) and define the functional
ψ i,mi that acts on ρi,mi via (A.30) and (A.33).
Concerning the second part of the claim, for any process P i ∈ A and P j ∈ A consider the
SDE:
Z t Z t
i,wi i j i,wi i j
Wt (P , P ) = wi + b̄i (Ws (P , P ))ds + σ̄i (Wsi,wi (P i , P j ))dBs
0 0
+ Pti − Ptj (A.34)
i,wi
By (Tanaka, 2002, Theorem 4.1), for any W0− (P i,∗ , P j,∗ ) = wi ∈ (mi , 1 − mj ), there exists a
unique pair (P i,∗ , P j,∗ ) ∈ A2 with continuous paths such that, with probability 1,
(a) For any t ≥ 0, mi ≤ Wti,wi (P i,∗ , P j,∗ ) < 1 − w̃j ,
23

Z
(b) 1nW i,wi (P i,∗ ,P j,∗ )>m o dPti,∗ = 0,
t i
[0,∞)
Z
(c) 1nW i,wi (P i,∗ ,P j,∗ )<1−m o dPtj,∗ = 0,
t j
[0,∞)
and W·i,wi (P i,∗ , P j,∗ ) has continuous paths, almost surely. Therefore, we can define a unique
pair (Ai,c,∗ , Aj,c,∗ ) ∈ A2 with continuous paths such that
Z t
Pti,∗ = (1 − Wsi,wi (P i,∗ , P j,∗ ))dAi,c,∗
s = (1 − mi )Ai,c,∗
t , (A.35)
0
Z t
Ptj,∗ = Wsi,wi (P i,∗ , P j,∗ )dAj,c,∗
s = (1 − mj )Aj,c,∗
t . (A.36)
0
Substituting (A.35) into (A.34) reveals that Wti,wi (P i,∗ , P j,∗ ) satisfies the SDE (A.22) with
h i+
Ak = Ak,c,∗ for any k ∈ {a, b}. For wi ∈ (0, 1), define the processes Ai,∗t = A i,∗,c
t + ln 1−wi
1−mi
h i+
j,∗ j,∗,c 1−wi i,∗ j,∗
and At = At + ln mj . Note that (A0 , A0 ) are the unique functions of wi such that
0 1 W i,wi (Ai,∗ ,Aj,∗ )>m
W0i,wi (Ai,∗ , Aj,∗ ) ∈ [mi , 1−mj ], Ai,∗ n o = 0 and Aj,∗ 1n o =
0 i,w
0 i W i (Ai,∗ ,Aj,∗ )<1−m
0 j
0. Together with the i,∗ =

properties (a)–(c), it follows by Part (i) of the claim that A
ψ·i,w̃i wi , B[0,·) , Aj,∗
[0,·] for any i 6= j ∈ {a, b}. As ∆Ai,∗t = ∆Aj,∗t = 0 a.s. for any t > 0,
the expressions of (A.23) and (A.24) simplify to (A.25) and (A.26).
A.6 Cheating Thresholds (Lemma 2.7)

ˆ i = (wi , wj ) ∈ (0, 1)2 : wi < min(ŵi , 1 − wj ) ⊂ ∆.

For any i 6= j ∈ {a, b}, let ∆
Lemma A.11 (Existence of thresholds). The following hold for any i 6= j ∈ {a, b}:
(i) αi < 0, ai > 1 − γi , βi > 1 − γi , bi < 0 and ŵi < 1 − γi .

ˆ i and
(ii) There exists a function f i whose graph satisfies (f i (wj ), wj ) : wj ∈ (0, 1) ⊂ ∆

(wi , wj ) ∈ ∆ : F i (wi , wj ) = 0 = (f i (wj ), wj ) : wj ∈ (0, 1) .

(A.37)
(iii) (a) f i is differentiable.

(b) limwj ↑1 f i (wj ) = 0 and limwj ↓0 f i (wj ) = ŵi .
(iv) There exists (w̃a , w̃b ) ∈ ∆ such that
F a (w̃a , w̃b ) = F b (w̃b , w̃a ) = 0. (A.38)
Moreover, such pair (w̃a , w̃b ) satisfies w̃k < ŵk for all k ∈ {a, b}.
Figure 4 displays the functions f a , f b and the solution (w̃a , w̃b ) to (A.38), where one can
observe that f a and f b satisfy (ii) and (iii). In this case, the pair (w̃a , w̃b ) satisfying (iv) is
unique.
24

wb wb
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 w 0 w
0 0.2 0.4 0.6 0.8 1 a 0 0.2 0.4 0.6 0.8 1 a
wb wb
1 1
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 w 0 w
0 0.2 0.4 0.6 0.8 1 a 0 0.2 0.4 0.6 0.8 1 a
Fig. 3: The functions f a (blue) and f b (red) (described in Lemma A.11) satisfy-
ing F (f (wb ), wb ) = 0 and F b (f b (wa ), wa ); and the fraud threshold (w̃a , w̃b ) satisfying
a a
F a (w̃a , w̃b ) = F b (w̃b , w̃a ) = 0 against trader a’s expected return (upper-left, 0% ≤ µa ≤ 60%),
volatility (upper-right, 0% < σa ≤ 100% ), risk-aversion (bottom-left, 0.1 < γa < 0.9), and
average horizon (bottom-right, 0 < 1/λ ≤ 20). Other parameters are µa = µb = 10%,
σa = σb = 20%, γa = γb = 0.5, wa = 0.5, λ = 1/3.
Proof of Lemma A.11. Starting with (i), Assumption 2.1 implies that
!
γi σj2 γi σi2

+ +
pi > (1 − γi ) [µi − µj ] + and qi > (1 − γi ) [µj − µi ] + , (A.39)
2 2
and inequality (A.39) yields
r !
1
αi < 2 ki − ki2 + σ 2 (1 − γi ) 2[µi − µj ]+ + γi σj2 < 0, ai > 1 − γi ,
σ
r !
1 2 2

+ 2

βi > 2 ki + ki + σ (1 − γi ) 2[µi − µj ] + γi σj =: β̂i (γi ),
σ
ŵi < 1 − γi .
25

wb
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0 w
0 0.2 0.4 0.6 0.8 1 a
Fig. 4: Approximation of the implicit curves f a (wb ) and f b (wa ) in Lemma A.11 that satisfy
F a (f a (wb ), wb ) = 0 for any 0 < wb < 1 (vertical axis) and F b wa , f b (wa ) = 0 for any 0 <
wa < 1 (horizontal axis) such that wa + wb < 1 with parameters γa = γb = 0.5, µa = µb = 10%,
σa = σb = 20% and λ = 1/3.
Note that ki also depends on γi . Because (β̂i (γi ) − (1 − γi ))0 > 0 for any 0 < γi < 1, it follows
that β̂i (γi ) − (1 − γi ) ≥ β̂i (0) − 1 ≥ 0, whence βi > 1 − γi and bi < 0.
Proof of (ii): First, show the inclusion ’⊃’ in (A.37). For any wj ∈ (0, 1),
1 − wj βi 1 − wi −bi

i
lim F (wi , wj ) = lim αi ai (1 − γi ) (wj − bi ) = −∞ (A.40)
wi ↓0 wi ↓0 wj wi
and 1−γi
1 − wj
lim F i (wi , wj ) = −ai bi (ai − bi )γi > 0. (A.41)
wi ↑1−wj wj
Next, to show that
F i (ŵi , wj ) > 0 for any wj < 1 − ŵi . (A.42)
decompose Fi as
w
αi bi (ai − bi )( 1−wj j )−αi
i
F (ŵi , wj ) = li (wj ), (A.43)
(1 − αi )(γi − αi )
where
γi +αi γi +αi
i 2 wj 2 γi (1 − αi )
l (wj ) = (1 − wj )(1 − αi ) − (1 − γi ) (ai − wj ) .
1 − wj −αi (1 − γi )
j w
αi bi (ai −bi )( 1−w )−αi
> 0, so sgn F i (ŵi , wj ) = sgn li (wj ) . The first and second
j

Note that (1−αi )(γi −αi )
derivatives of li are
γi +αi γi +αi
wj γi (1 − αi )
i
lw (wj ) = (αi − 1) 2
wj−1 (αi + γ i − wj ) + (1 − γi ) 2
,
j
1 − wj −αi (1 − γi )
26

and
γi +αi
ai (γi + αi )(αi − 1)2

i wj
lw j wj
(wj ) =− ≥0 iff αi + γi ≤ 0. (A.44)
wj2 (1 − wj ) 1 − wj
Distinguish now the two cases:

(i) If αi + γi ≤ 0, it follows by (A.44) that
γi +αi
γi (1 − αi )
i
lw i
(wj ) ≤ lw (1 − ŵi ) = −ai γi−1 (γi − αi )2 < 0,
j j
−αi (1 − γi )
and therefore,
γi +αi
i i γi (1 − αi )
l (wj ) > l (1 − ŵi ) = ai (1 − γi )(γi − αi ) > 0.
−αi (1 − γi )
(ii) If αj + γj > 0, then

lim li (wj ) = 0.
wj ↓0
The concavity of li (implied by (A.44)) in combination with li (1 − ŵi ) > 0 yield
li (wj ) > 0, wj < 1 − ŵi ,
whence (A.42) follows.

Due to (A.40),–(A.42) the intermediate value theorem implies that for any wj ∈ (0, 1), there
exist ui ∈ min(ŵi , 1 − wj ) such that F i (ui , wj ) = 0. Thus, there exists a function f i with graph
in ∆ ˆ i such that the ‘⊃’ relation of (A.37) holds.
Next, we prove the ’⊂’ relation of (A.37) by first showing that, for any fixed wj ∈ (0, 1),
if ui ∈ (0, min(ŵi , 1 − wj )) is such that F i (ui , wj ) = 0, then Fwi i (ui , wj ) > 0. The equation
F i (ui , wj ) = 0 expands to
1 − ui −bi 1 − wj βi

ai (αi (1 − γi − ui ) + γi ui )(wj − bi ) =
ui wj
− (ai − bi )(ui (αi + βi − 1) − αi βi )wjγi (1 − wj )1−γi
ai −αi
ui wj
+ bi (βi (1 − γi − ui ) + γi ui )(wj − ai ) (A.45)
1 − ui 1 − wj
and differentiating F i over wi yields for any (wi , wj ) ∈ ∆
− (1 − wi )wi Fwi i (wi , wj ) =

1 − wi −bi 1 − wj βi

ai (wj − bi ) (γi wi (wi − bi − 1) + bi γi αi + (1 − wi )(wi − bi )αi )
wi wj
ai −αi
wi wj
− bi (ai − wj ) (γi wi (1 + ai − wi ) − ai γi βi − (1 − wi )(wi − ai )βi )
1 − wi 1 − wj
− (ai − bi )(αi + βi − 1)(1 − wi )wi (1 − wj )1−γi wjγi . (A.46)
Substituting (A.45) into (A.46) yields that
sgn ∂wi F i (ui , wj ) = sgn ρi (ui , wj )li (ui ) ,

where
ai ai
i wi wj
ρ (wi , wj ) = −(1 − γi − wi )(ai − wj ) + (wi − αi )wj ,
1 − wi 1 − wj
27

and
li (wi ) = −(1 − γi )αi βi + (αi βi + γi (1 − αi − βi ))wi
for any wi ∈ (0, min(ŵi , 1 − wj )). If αi βi + γi (1 − αi − βi ) ≥ 0, then li (ui ) > 0; if αi βi + γi (1 −
αi − βi ) < 0 then, by the inequality ui < ŵi ,
αi (αi − 1)(1 − γi )γi
li (ui ) > > 0.
γi − αi
Hence,
sgn ∂wi F i (ui , wj ) = sgn ρi (ui , wj ) .

It is clear that ρi (ui , wj ) > 0 if wj ≥ ai . Thus, assume wj < ai . Since wj + ui < 1, wj ≥ ai is

satisfied if and only if wj < min{ai , 1 − ui }. Note that
lim ρi (ui , wj ) = 0, (A.47)

wj ↓0
and (
i ai (wi − αi ) > 0, if ai < 1 − ui
ρ (ui , min(ai , 1 − ui )) = , (A.48)
γi ai > 0, if ai ≥ 1 − ui
and ρiwj wj (ui , wj ) has the same sign as
ai ai
ui wj
−(ai − 1)ai (1 − γi − ui )(ai + wj ) ≤0 iff ai ≥ 1. (A.49)
1 − ui 1 − wj
If ai ≥ 1, then the concavity of ρi , (A.47) and (A.48) yield
ρi (ui , min{ai , 1 − ui })
ρi (ui , wj ) ≥ wj > 0.
min{ai , 1 − ui }
ai ai
ui ui w w
If ai < 1, then 1−wj > 1−wj and 1−uj i > 1−uj i give

i ui wj
ρ (ui , wj ) > −(1 − γi − ui )(ai − wj ) + (ui − αi )wj
1 − ui 1 − wj
wj ρ̂(ui , wj )
= ,
(1 − ui )(1 − wj )
where, for any wi ∈ (0, min(ŵi , 1 − wj )),
ρ̂i (wi , wj ) := −wi (1 − γi − wi )(ai − wj ) + (1 − wi )(1 − wj )(wi − αi ).
Taking the partial derivatives with respect to wi yields
ρ̂iwi (ui , wj ) = (1 − ai )(2 − γi − 2ui − wj ) > 0.
(The inequality follows from ui + wj < 1 and ui < ŵi < 1 − γi .) Thus,
∂wi F i (ui , wj ) > 0 (A.50)
for any ui ∈ (0, min(ŵi , 1 − wj )) is such that F i (ui , wj ) = 0. Define f i as follows:
f i (wj ) = inf wi ∈ (0, min(ŵi , 1 − wj )) : F i (wi , wj ) = 0 ,

which is the minimal zero of wi 7→ F i (wi , wj ). Suppose, by contradiction, that there exist
wi > f i (wj ) such that F i (wi , wj ) = 0 and let
vi = inf wi ∈ f i (wj ), min (ŵi , 1 − wj ) : F i (wi , wj ) = 0 ,

(A.51)
28

which is the first zero of wi 7→ F i (wi , wj ) after f i (wj ). Then by (A.50), Fwi i (f i(wj ), wj ) >
v −f i (w )

0 and Fwi i (vi , wj ) > 0. By the smoothness of F i , there exists ∈ 0, i 2 j such that
F i (f i (wj ) + , wj ) > 0 and F i (vi − , wj ) < 0. However, the intermediate value theorem implies
there exists zi ∈ (f i (wj ) + , vi − ) such that F i (zi , wj ) = 0, which is impossible in view of
definition (A.51).
To establish the claim, it remains to show that f i (wj ) is the unique solution in the larger
domain (0, 1 − wj ) (⊃ (0, min (ŵi , 1 − wj ))) such that F i (f i (wj ), wj ) = 0. This facts follows
by showing that there does not exist (wi , wj ) ∈ ∆\∆ ˆ i such that F i (wi , wj ) = 0. Note that
∆\∆ ˆ = {(wi , wj ) ∈ ∆ : wi ≥ ŵi }. By virtue of (A.42), it suffices to consider wi ∈ (ŵi , 1).
i
Differentiating F i with respect to wj reveals that Fwi j (wi , wj ) has the same sign as
1 − wj βi 1 − wi −bi

ai (bi βi + (1 − wj − βi )wj )((1 − γi − wi )αi + γi wi )
wj wi
−αi ai
wj wi
− bi (ai αi + wj (1 − wj − αi ))(γi wi + (1 − γi − wi )βi )
1 − wj 1 − wi
− (ai − bi )(γ + (1 − 2γi )wj )((αi + βi − 1)wi − αi βi )(1 − wj )1−γi wjγi . (A.52)
Let (ui , uj ) ∈ ∆ be such that F i (ui , uj ) = 0. Substituting F i (wi , uj ) = 0 into (A.52) yields

sgn Fwi j (ui , uj ) = sgn Gi (ui , uj ) ,

(A.53)
where for any (wi , wj ) ∈ ∆,
Gi (wi , wj ) := (γi + αi )(βi (1 − γi − wi ) + γi wi )

1 − wi βi −αi 1 − wj βi −αi

− (γi + βi )(αi (1 − γi − wi ) + γi wi ) . (A.54)
wi wj
Note that for any wi ∈ (ŵi , 1), αi (1 − γi − wi ) + γi wi > 0, and it follows that

sgn Giwj (wi , wj ) = sgn(αi (1 − γi − wi ) + γi wi ) > 0. (A.55)
Suppose, by contradiction, that for any fixed wi ∈ (ŵi , 1) there exists uj ∈ (0, 1 − wi ) such that
F i (wi , uj ) = 0. Let vj be the smallest uj , i.e.,
vj = inf wj ∈ (0, 1 − wi ) : F i (wi , wj ) = 0 .

(A.56)
Then by (A.53) and (A.55), Fwi j (wi , vj ) > 0, implying there exists an ∈ (0, vj ) such that
F i (wi , vj − ) < 0. Thus, as for any wi ∈ (ŵi , 1),
lim F i (wi , wj ) = +∞,

wj ↓0
there exists an intermediate point zj ∈ (0, vj − ) such that F i (wi , zj ) = 0, which contradicts
definition (A.56), and shows the non-existence of the solution in ∆\∆ ˆ i , hence the inclusion in
(A.37).
Proof of (iii)a: By (A.50) the implicit function theorem implies that, in the neighborhood
of any point (wi0 , wj0 ) ∈ ∆ such that wi < min(ŵi , 1 − wj ) that solves F i (wi0 , wj0 ) = 0, there
exists a unique smooth (in fact, analytic) function f0i (wj ) satisfying f0i (wj0 ) = wi0 such that
f0i (wj ) ∈ (0, min(ŵi , 1 − wj )) and F i (f0i (wj ), wj ) = 0. Suppose, by contradiction, that f i is
not analytic. Then, then there exists wj0 ∈ (0, 1), where the local function f0i (wj0 ) 6= f i (wj0 ).
But this implication contradicts uniqueness, and thus wj 7→ f i (wj ) is indeed analytic on the
29

open domain (0, 1). Proof of (iii)b: Since 0 < f i (wj ) < 1 − wj for all wj ∈ (0, 1), the limit
limwj ↑1 f i (wj ) = 0 follows. Moreover, for any wi ∈ (0, ŵi ),
lim F i (wi , wj ) = −∞, (A.57)

wj ↓0
wi 1−γi
lim F i (wi , wj ) = −γi ai bi (ai − bi ) > 0. (A.58)
wj ↑1−wi 1 − wi
Thus, by the intermediate value theorem, for any wi ∈ (0, ŵi ) there exists uj < 1 − wi such that
F i (wi , uj ) = 0 and so there exists a function g i : (0, ŵi ) → (0, 1) such that
n o
ˆ i : F i (wi , wj ) = 0 .
(wi , g i (wi )) : wi ∈ (0, ŵi ) ⊂ (wi , wj ) ∈ ∆

Property (iii)b yields
(wi , g i (wi )) : wi ∈ (0, ŵi ) ⊂ (f i (wj ), wj ) : wj ∈ (0, 1) ,

which implies supwj ∈(0,1) f i (wj ) ≥ ŵi , equivalently,
max f i (wj ) ≥ ŵi or lim f i (wj ) ≥ ŵi .

wj (0,1) wj ↓0
On the other hand, f i (wj ) < ŵi for all wj ∈ (0, 1) implies maxwj (0,1) f i (wj ) < ŵi and by the
continuity of f i , limwj ↓0 f i (wj ) ≤ ŵi . Thus, limwj ↓0 f i (wj ) = ŵi .
Proof of (iv): First we establish the existence of (w̃a , w̃b ) ∈ ∆ such that F a (w̃a , w̃b ) =
F (w̃b , w̃a ) = 0. For any i 6= j ∈ {a, b}, by (iii)b we extend f i continuously to 0 by setting
a
f i (0) = limwj ↓0 f i (wj ) = ŵi and to 1 by setting f i (1) = limwj ↑1 f i (wj ) = 0. Let D = (0, 1)2
and define the function H : D̄ → D̄ such that H(wa , wb ) := f a (wb ), f b (wa ) . Property (ii)

implies for any i 6= j ∈ {a, b} and for any wj ∈ [0, 1], f i (wj ) ∈ [0, ŵi ] ⊂ [0, 1]. (Therefore, H is
well-defined.) Since D̄ is compact and H is continuous (due to (iii)a), by Brouwer’s fixed point
theorem there exists (w̃a , w̃b ) ∈ D̄ such that f a (w̃b ), f b (w̃a ) = (w̃a , w̃b ). Next, we show that
/ ∂D. Note that ∂D = D1a,b ∪ D1b,a ∪ D2a,b ∪ D2b,a where
(w̃a , w̃b ) ∈
D1a,b = {[0, wb ] : wb ∈ [0, 1]}, D1b,a = {[wa , 0] : wa ∈ [0, 1]},

D2a,b = {[1, wb ] : wb ∈ [0, 1]}, D2b,a = {[wa , 1] : wa ∈ [0, 1]}.
For any i 6= j ∈ {a, b}, on D1i,j , f i f j (0) = f i (ŵj ) 6= 0 since f i (wj ) ∈ (0, ŵi ) for any wj ∈

(0, 1); and on D2i,j , f i f j (1) = f i (0) 6= 1 because f i (0) = ŵi . Hence, (w̃a , w̃b ) ∈

/ ∂D and so
i
(w̃a , w̃b ) ∈ D. To this end, by (ii) it follows that w̃i = f (w̃j ) < 1 − w̃j for any i 6= j ∈ {a, b},
thus, (w̃a , w̃b ) ∈ ∆.
To conclude the proof, note that if a pair (w̃a , w̃b ) ∈ ∆ is such that F a (w̃a , w̃b ) = F a (w̃b , w̃a ) =
0, then by (ii), (w̃a , w̃b ) = f a (w̃b ), f b (w̃a ) ∈ ∆ ˆa ∩ ∆ˆ b , meaning that w̃k < ŵk for any
k ∈ {a, b}.
In the following result, we first obtain some properties of the function Gi (given by (A.54));
then we find that the function f i is strictly decreasing.
Lemma A.12 (Monotonicity of f i ). The following statements hold for any i 6= j ∈ {a, b},
(i) If αi < −γi , then

(a) There exists a function g i : 0, 1+γ
2
i
→ (0, ŵi ) such that
n o
1 + γi

ˆ i i i
(wi , wj ) ∈ ∆ : G (wi , wj ) = 0 = (g (wj ), wj ) : wj ∈ 0, . (A.59)
2
30

wi
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0 w
0 0.2 0.4 0.6 0.8 1 j
Fig. 5: The function f i : (0, 1) i f i (w ), w

→ (0, ŵ i ) (Lemma A.11, in blue) such that F j j = 0;
1+γ
and the function g i : 0, 2 i → (0, ŵi ) (Lemma A.12, in green) such that Gi g i (wj ), wj = 0,

with the domain of each in dashed horizontal lines.
ˆ i,
(b) For any (wi , wj ) ∈ ∆
 1+γi
> 0
 if wj ≥ 2 , (A.60a)
Gi (wi , wj ) > 0 if wj < 1+γi
2 and wi < g i (wj ), (A.60b)

1+γi
g i (w

<0 if wj < 2 and wi > j ). (A.60c)
dg i (w )

(c) g i is differentiable and dwj j > 0 for all wj ∈ 0, 1+γ 2
i
.

(d) For any wj ∈ 0, 1+γ 2
i
, g i (wj ) > 1−γi
2 with limwj ↑ 1+γ
i
i g (wj ) =
1−γi
2 and limwj ↓0 g i (wj ) =
2
ŵi .
(ii) f i is strictly decreasing on (0, 1).

ˆ i,
(iii) Denote by f i,−1 the inverse of f i . For any (wi , wj ) ∈ ∆
(
> 0, if wj > f i,−1 (wi )
F i (wi , wj ) .
< 0, if wj < f i,−1 (wi )
1−γi
Proof. Proof of (i)a and (i)b: Note that αi < −γi if and only if ŵi > 2 . A direct calculation
31

reveals that
lim Gi (wi , wj ) = +∞, (A.61)
wi ↓0
lim Gi (wi , wj ) = (γi + αi )(βi (1 − γi − wi ) + γi wi ) < 0 wj ∈ (0, 1 − ŵi ) , (A.62)

wi ↑ŵi

1 + γi
lim Gi (wi , wj ) = γi (αi − βi )(1 + γi − 2wj ) < 0 wj ∈ 1 − ŵi , , (A.63)
wi ↑1−wj 2

1 + γi
lim Gi (wi , wj ) = γi (αi − βi )(1 + γi − 2wj ) ≥ 0 wj ∈ ,1 , (A.64)
wi ↑1−wj 2
and Gi has the partial derivative
Giwi (wi , wj ) = (αi + γi )(βi + γi )
βi −αi
(1 − wi )(1 − wj )
+ (γi + βi ) (αi − γi + (βi − αi ) (γi wi + (1 − γi − wi )αi ))
wi wj
< 2γi (αi − βi ) < 0, (A.65)
where the inequality follows from wi < 1 − wj and γi wi + (1 − γi − wi )αi < 0 (which, in turn,
follows from wi < w̃i ).
Let first wj ≥ (1 + γi )/2: Strict monotonicity (A.65) and the limits (A.61) and (A.64) of
equal sign imply inequality (A.60a). Let now wj < (1 + wi )/2. The map wi 7→ Gi (wi , wj )
changes sign on (0, min(ŵi , 1 − wj ) due to (A.61) and (A.62) resp. (A.63). Strict monotonicity
(A.65) therefore implies the existence of a unique ui = g i (wj ) ∈ (0, min(ŵi , 1 − wj )) such that
Gi (ui , wj ) = 0, and therefore also the inequalities (A.60b)–(A.60c) must hold. This settles both
the proof of (i)b and of (i)a.
Proof of (i)c: The differentiability follows by similar arguments as in the proof of Lemma
A.11 (iii)a. To check the monotonicity, note that the implicit function theorem implies for any
1+γi
wj ∈ 0, 2 ,
dg i (wj ) Giw (wi , wj )
=− ij .
dwj Gwi (wi , wj ) wi =f i (wj )

Checking the partial derivative Giwj and using wi < ŵi it follows that

sgn Giwj (wi , wj ) = sgn(γi wi + (1 − γi − wi )αi ) < 0,
which together with (A.65) yields

!
Giwj (wi , wj )
i
dg (wj )
sgn = sgn − i < 0.
dwj Gwi (wi , wj ) wi =f i (wj )

Proof of (i)d: wi = 1−γ 1+γi

2 satisfies wj < 1 − wi = 2 , which implies
i
!
1 + γi βi −αi 1 − wj βi −αi

i 1 − γi 1 − γi
G , wj = (γi + αi )(γi + βi ) 1 −
2 2 1 − γi wj
> 0.
Hence, g i (wj ) > 1−γ
2
i
by the monotonicity of wi 7→ Gi (wi , wj ) (see (A.65)). Together with the
continuity of g i and g i (wj ) < min(ŵ, 1 − wj ) ((i)a), it follows that limw ↑ 1+γi g i (wj ) = 1−γi
2 .
j 2

Finally, as for any wi ∈ 1−γ i
2 , ŵi ,
lim Gi (wi , wj ) = +∞,

wj ↓0
lim Gi (wi , wj ) = γi (βi − αi )(1 − γi − 2wi ) < 0,

wj ↑1−wi
32

1−γi
and by the intermediate value theorem for any wi ∈ 2 , ŵi there exists uj ∈ (0, 1 − wi ) such
that Gi (w
i , uj ) = 0. By the inclusion ⊆ in (i)a, g i (uj )
= wi . Because this statement holds for
1−γi i i
any wi ∈ 2 , ŵi , also supw ∈ 0, 1+γi g (wj ) = ŵi . As g is strictly bounded from below by

j 2
ŵi , it follows that limwj ↓0 g i (wj ) = ŵi .

Proof of (ii): The implicit function theorem yields
df i (wj ) Fwi (wi , wj )

= − ij , (A.66)
dwj Fwi (wi , wj ) wi =f i (wj )

hence by (A.50), (A.53) and (A.66),

i
df (wj )
= −sgn Gi (f i (wj ), wj ) .

sgn (A.67)
dwj
Next, to show that f i (wj ) is strictly decreasing, distinguish two cases:
(i) αi ≥ γi : Lemma A.11 (i) shows that Gi (wi , wj ) > 0 on ∆ ˆ i . As Lemma A.11 (ii) implies the
i
graph of f i satisfies (f i (wj ), wj ) : wj ∈ (0, 1) ⊂ ∆ ˆ i , by (A.67) it follows that df (wj ) < 0

dwj
for all wj ∈ (0, 1).
n o
(ii) αi < γi : By (A.60a), Gi (wi , wj ) > 0 on ∆ ˆ i ∩ (wi , wj ) ∈ ∆ ˆ i : wj ≥ 1+γi and by Lemma
2
A.11 (i) it follows that
\
1 + γi ˆi ˆ i : wj ≥ 1 + γ i .
(f i (wj ), wj ) : wj ∈ 0, ⊂∆ (wi , wj ) ∈ ∆
2 2
df i (wj )
Hence, by (A.67) dwj < 0 for any wj ∈ [ 1+γi
2 , 1). It remains to to check the monotonicity
of f i on the interval (0, 1+γi i
2 ) (note that this coincides with the whole domain of g ).
We proceed by showing z i (wj ) := f i (wj ) − g i (wj ) ≤ 0 for all wj ∈ (0, 1+γ i
2 ). Note that
since f i (wj ) < 1 − wj for any wj (0, 1), then f i 1+γ 2
i
< 1−γ i
2 . Then by (i)d, it follows
that
i i 1 + γi
lim z (wj ) = f − lim g i (wj ) < 0. (A.68)
1+γ
wj ↑ 2 i 2 1+γ
wj ↑ 2 i

Suppose, by contradiction, that there exist vj ∈ 0, 1+γ 2
i
such that z i (vj ) > 0. Then, the

1+γi
intermediate value theorem implies that there exists wj ∈ vj , 2 such that z i (wj ) = 0.
∗
Let wj be the first such point, i.e.,

1 + γi
wj∗ = inf wj ∈ vj , i
: z (wj ) = 0 .
2
Note that this definition implies
z i (wj ) > 0 wj ∈ vj , wj∗ .

for any (A.69)

By the mean value theorem, there exists u∗j ∈ vj , wj∗ such that
dz i (wj )
i
df (wj ) dg i (wj )

= − < 0.
dwj wj =u∗j dwj dwj

wj =u∗j
33

Then, by (i)c it follows that
df i (wj ) dg i (wj )
< <0
dwj wj =u∗j dwj wj =u∗j

and (A.67) implies Gi f (u∗j ), u∗j > 0, which in turn by (A.60b) yields f (u∗j ) < g i (u∗j ), in
contradiction to (A.69). Hence, z i (wj ) ≤ 0 for all wj ∈ (0, 1+γi
2 ).

Next, we show that z i (wj ) < 0 almost everywhere (a.e.) on 0, 1+γ 2
i
. Suppose, by

contradiction, that there exists an interval (a, b) ⊂ 0, 1+γ
2
i
such that z i (wj ) = 0 for any
wj ∈ (a, b) or, equivalently,
f i (wj ) = g i (wj ) on (a, b). (A.70)
By (i)a it follows that Gi (f i (wj ), wj ) = 0 for any wj ∈ (a, b). Then (A.67) yields that
f i is a constant on (a, b), which by (A.70) in turn implies g i is a constant on (a,b), an
impossibility due to (i)c. Hence, f i (wj ) < g i (wj ) almost everywhere on 0, 1+γ 2
i
. By
df i (wj )

A.60b and (A.67), it follows that dwj < 0 a.e. also on 0, 1+γ 2
i
, completing the proof.
Proof of (iii): The strict monotonicity (ii) shows the inverse function f i,−1 of f i is well-
defined and differentiable. By Lemma A.11 (ii), for any wi ∈ (0, ŵi ), f i,−1 (wi ) is the unique
point in (0, 1 − wi ) such that F i (wi , f i,−1 (wi )) = 0. Note that limwj ↓0 F i (wi , wj ) = −∞ and
wi 1−γi
limwj ↑1−wi F i (wi , wj ) = −γi ai bi (ai − bi ) 1−w i
> 0. Suppose, by contradiction, that there
i,−1 i,−1 (wi ), 1 − wi ) such that F i (wi , uj ) > 0 (resp.

exists uj ∈ (0, f (wi )) (resp. uj ∈ f
F i (wi , uj ) < 0). Then, the intermediate value theorem implies that there exists vj ∈ (0, uj ) ⊂
(0, f i,−1 (wi )) (resp. vj ∈ (uj , 1 − wi ) ⊂ (f i,−1 (wi ), 1 − wi )) such that F i (wi , vj ) = 0, contradict-
ing the uniqueness of f i,−1 (wi ) which is a zero of the map wj 7→ F i (wi , wj ). Hence, F i (wi , wj ) <
0 for any wj ∈ (0, f i,−1 (wi )), and F i (wi , wj ) > 0 for any wj ∈ (f i,−1 (wi ), 1 − wi ).
B Derivation of the HJB equation

This section contains the heuristic derivation of the HJB equations (B.11)-(B.15).
By the linearity of the SDE (1.2), we conjecture a singular-type Nash equilibrium in which
each trader prevents the wealth process from leaving a region. Let for any i 6= j ∈ {a, b},
C j ⊂ R2+ be an open set and C¯j := C j ∪ ∂C j its closure 2 j,C j ∈ Λ be such that for
in R+ . Let Ψ
j

any Ai ∈ A there exists a unique process Aj,∗ j,C
· = Ψ· x, B[0,·) , Ai[0,·] ∈ A satisfying a.s.,
(i) for any t ≥ 0, Ytx (Ai , Aj,∗ ) ∈ R2+ \C j ,
(ii) [0,∞) 1{Ytx (Ai ,Aj,∗ )∈∂C j } dAj,∗

R
t = 0.
In this way, trader j keeps personal wealth inside the region R2+ \C j at any time t ≥ 0 and for
any trader i’s fraud process Ai ∈ A. Moreover, if x ∈ C j , then trader j cheats instantly so as to
bring wealth at time 0 to ∂C j and if the state is at ∂C j trader j cheats as little as necessary so
to keep the wealth process in the interior of R2+ \C j (hence, C¯j is the fraud-region of trader j).
Suppose the optimal fraud process Ai,∗ ∈ A for trader i is such that a.s. for any t ≥ 0
dAi,∗ j,∗
t dAt = 0 (B.1)
(which means that Ai,∗ and Aj,∗ do not simultaneously increase) and the value function V i (x; Aj,∗ )
is smooth for any x ∈ R2+ . Then for any x = (xa , xb ) ∈ C j , Aj,∗ j,∗
0 = A0 (x) > 0 is such that
34

Y0x ∈ ∂C j . By (B.1) and Lemma 2.2 (i), the game value satisfies, for any x ∈ C j and any
0 ≤ α ≤ Aj,∗0 (x)
Z ∞
−λt−Ai,∗ −Aj,∗

i j,∗ −α i i,x α
V ((xi , xj ); A ) = e E λ e U Yt dtx = (xi , xj + (xa + xb )(e − 1))
t t

0
= e−α V i (xi , xj + (xa + xb )(eα − 1)); Aj,∗ .

Since
e−α V i (xi , xj + (xa + xb )(eα − 1)); Aj,∗ − V i ((xi , xj ); Aj,∗ )

lim = 0,
α↓0 α
it follows that
(xa + xb )Vxij − V i = 0 on C j . (B.2)
Define the associated differential operator (this is the infinitesimal generator of the uncontrolled
pre-bankruptcy process Y x (0, 0)),
X 1 X 2 2 2
Lφ(x) = µk xk ∂xk φ(x) + σk xk ∂xk xk φ(x) (B.3)
2
k∈{a,b} k∈{a,b}
for any φ ∈ C 2 (R2+ ). For any x ∈ R2+ \C j , the problem for trader i becomes an optimal (singular)
control problem. Treating (Y·x , Ai· , Aj,∗
· ) as the state process, the dynamic programming prin-
ciple (see e.g. (Fleming and Soner, 2006, Section VIII.2)) yields the following quasi-variational
inequality (QVI):
max LV i − λV i + U i , (xa + xb )Vxii − V i = 0

(B.4)
R2+ \C j
and the verification theorems (cf. (Carmona, 2016, Theorem 3.18), (Fleming and Soner, 2006,
Theorem 4.1), (Øksendal and Sulem, 2019, Theorem 8.2)) suggest that the set
C¯i = x ∈ R2+ \C j : (xa + xb )Vxii − V i = 0 ,

(B.5)
corresponds to the fraud-region of Ai,∗ , so that the optimal cheating strategy for trader i is to
only cheat in a region C i ⊂ R2+ and prevent the wealth process from leaving the region R2+ \C i
at any time t ≥ 0. More formally, and similarly to Aj,∗ , Ai,∗ is such that, a.s. for any t ≥ 0,
(i) Ytx (Ai,∗ , Aj,∗ ) ∈ R2+ \(C j ∪ C i )
(ii) [0,∞) 1{Ytx (Ai,∗ ,Aj,∗ )∈∂C i } dAi,∗

R
t =0
Here R2+ \(C j ∪ C i ) is the common no-fraud region. (Condition (B.1) shows that their fraud-
regions should not intersect, so suppose C i ∩ C j = ∅.)
Substituting (B.5) into (B.4), it follows that for any x ∈ R2+ \C j
LV i − λV i + U i = 0 on R2+ \(C i ∪ C j ), (B.6)

i i i i
LV − λV + U < 0 on C, (B.7)
(xa + xb )Vxii i
−V <0 on R2+ \(C i j
∪ C ). (B.8)
Let Li be the following differential operator acting on ϕ ∈ C 2 (R+ ),

γi
Li ϕ(w) = (1 − γi ) µi w + µj (1 − w) − (σi2 w2 + σj2 (1 − w)2 ) ϕ(w)
2
+ µi − µj + γi (σj (1 − w) − σi2 w) w(1 − w)ϕw (w)
2

σi2 + σj2 2
+ w (1 − w)2 ϕww (w).
2
35

Now, conjecture that for both traders k ∈ {a, b}, the fraud regions in a Nash equilibrium are of
the form
C k = {x ∈ R2+ : rk (x) < mk }, (B.9)
where mk ∈ (0, 1) such that 0 < ma + mb < 1. In other words, by cheating, traders prevent
their fraction of wealth from going below their critical threshold mi . The condition ma +mb < 1
ensures that C i ∩ C j = ∅. Hence, the equilibrium strategies (Aa,∗ , Ab,∗ ) can be expressed as
i,m
Skorokhod reflections such that for any i 6= j, Ai,∗ = ϕ· i ri (x), B[0,·) , Aj,∗ [0,·] where ϕi,mi
solves SPmi + (see Definition 2.5). The scale invariance of J i (Lemma 2.2 (ii)) is inherited by
the value function (i.e. for any c > 0, V i (cx; Aj,∗ ) = c1−γi V i (x; Aj,∗ )) so that for any x ∈ R2+
and ϕi (w) = V i ((w, 1 − w); Aj,∗ ) for any w ∈ (0, 1). Together with Lemma 2.2 (i) yield that V i
is of the form
V i (x; Aj,∗ ) = λ(xa + xb )1−γi ϕi (ri (x)). (B.10)
Let ri (x) = w and substitute (B.10) and (B.9) into (B.6), (B.7), (B.5), (B.8) and (B.2), yielding
the HJB equations
Li ϕi (w) − λϕi (w) + U i (w) = 0 on (mi , 1 − mj ), (B.11)

i i i i
L ϕ (w) − λϕ (w) + U (w) < 0 on (0, mi ), (B.12)
(1 − w)ϕiw (w) − γi ϕi (w) = 0 on (0, mi ), (B.13)
(1 − w)ϕiw (w) − γi ϕ (w) < 0i
on (mi , 1 − mj ), (B.14)
wϕiw (w) + γi ϕi (w) = 0 on (1 − mj , 1). (B.15)
which are the starting point of verification.
C Verification Theorem
The following result establishes the link between the HJB equations (B.13)–(B.15) and the
optimization problem.
Lemma C.1. Let (ma , mb ) ∈ ∆. For any i 6= j ∈ {a, b} let ϕi ∈ C 1 ((0, 1)) be an R+ -valued
function satisfying (B.13)-(B.15). For any α ≥ 0 and w ∈ (0, 1), set
+
˜j 1 α
f (α, w) := ln 1 + [mj e − (1 − w)] (C.1)
1 − mj
and
eα − (1 − w)

i −α−f˜j (α,w) α f˜j (α,w) 1−γi i
h̃ (α, w) := e (e + e − 1) ϕ . (C.2)
eα + ef˜j (α,w) − 1
Then:
(i) f˜j (α, w) > 0 if and only if one of the following inequalities is satisfied:
(a) w > 1 − mj ,

1−w
(b) w ≤ 1 − mj and α > ln mj .
(ii) If (α, w) is such that f˜j (α, w) > 0 then ∂α h̃i (α, w) < 0.
36

(iii) For any α ≥ 0 and for all w ∈ (0, 1),
h̃i (α, w) − ϕi (1 − w) ≤ 0,
and the equality holds for a fixed w if and only if
α = 0, if w ∈ (mi , 1), (C.3)

1−w
α ≤ ln , if w ∈ (0, mi ]. (C.4)
1 − mi
˜j
Proof. Proof We show the equivalent statement, f (α, w) = 0 if and only if w ≤ 1 − mj
of (i):
and α ≤ ln 1−w ˜j α 1−w α 1−w 1−w
mj . Note that f (α, w) = 0 if and only if e ≤ mj . If e ≤ mj , then mj ≥ 1

because eα ≥ 1, which, together with eα ≤ 1−w mj , implies that α ≤ ln
1−w
mj . The converse

implication follows from the monotonicity of the exponential, applied to α ≤ ln 1−w
mj .
Proof of (ii): If f˜j (α, w) > 0, then
eα − (1 − w)
= 1 − mj
eα + ef˜j (α,w) − 1
and thus h̃i simplifies to

γi
1 − (1 − w)e−α

i 1 − mj
h̃ (α, w) = ϕi (1 − mj ).
w + (eα − 1)mj α
e − (1 − w)
Differentiating h̃i with respect to α, and recalling that ϕi is strictly positive, it follows that
∂α h̃i (α, w) has the same sign as
g i (α, w) = (1 − w)(w − mj ) − e2α (1 + γi )mj + eα (2(1 − w)mj − γi (w − mj )) .
As ∂α g i (α, w) ≤ eα (−γi (w + mj ) − 2wmj ) < 0, g i (α, w) ≤ g i (0, w) = −w(γi +mj −(1−w)) < 0.
Hence, ∂α h̃(α, w) < 0.
Proof of (iii): All classical solutions of the linear ODEs (B.13) and (B.15) are of the form
C(1 − w)−γi and Dw−γi , C, D ∈ R,
respectively. As ϕi is positive,
C0 (1 − w)−γi for w ∈ (0, mi ),

i (C.5a)
ϕ = −γi
C1 w for w ∈ (1 − mj , 1), (C.5b)
where C0 > 0 and C1 > 0. Distinguish now three cases:
(i) w ∈ (1 − mj , 1): By (i) f˜i (α, w) > 0 for any α ≥ 0, thus by (ii), h̃i (α, w) ≤ h̃i (0, w) with
equality if and only if α = 0. Thus, in conjunction with (C.5b),
h̃i (α, w) − ϕi (w) ≤ h̃i (0, w) − ϕi (w)

1 − mj γi i

= ϕ (1 − mj ) − ϕi (w)
w
= 0,
where the equality holds if and only if α = 0.
37

(ii) w ∈ (mi , 1 − mj ]: If α > ln 1−w , by (i) f˜i (α, w) > 0. Therefore, by (ii), ∂α h̃i (α, w) < 0
mj
and thus h̃i (α, w) < h̃i ln 1−w
mj , w for any α > ln 1−w
mj . If α ≤ ln 1−w
mj , then by
(i) f i (α, w) = 0, and h̃i reduces to h̃i (α, w) = e−γi α ϕi (1 − (1 − w)e−α ). By (B.14),
∂α h̃i (α, w) = −γi h̃i (α, w) + e−(1+γi )α (1 − w)ϕiw 1 − (1 − w)e−α

≤ −γi h̃i (α, w) + γi e−(1+γi )α (1 − w)ϕi 1 − (1 − w)e−α

= γi e−γi α e−α − 1 ϕi 1 − (1 − w)e−α ≤ 0,

where equality holds if and only if α = 0. Hence, for any α ≥ 0,
h̃i (α, w) − ϕi (w) ≤ h̃i (0, w) − ϕi (w) = 0,
with equality if and only if α = 0.

(iii) w ∈ (0, mi ]: If α > ln 1−w , then by (i) and (ii) it follows that h̃i (α, w) < h̃i ln 1−w , w
mj mj
1−w 1−w i
for any α > ln mj . If α ≤ ln mj , then a similar argument as above yields h̃ (α, w) =
i
e−γi α ϕi (1 − (1 − w)e−α ). If α ∈ ln 1−m 1−w
i
, ln 1−w
mj then 1 − (1 − w)e−α ∈ [mj , 1 −

mi ). By (B.14), it analogously follows that ∂α h̃i (α, w) < 0, thus, h̃i ln 1−w mj ,w <

h̃i ln 1−m1−w
i
, w . If α ≤ ln 1−m1−w
i
, then 1 − (1 − w)e−α ∈ [1 − mi , 1) and by (C.5a) it
follows that,
h̃i (α, w) − ϕi (w) = e−γi α ϕi 1 − (1 − w)e−α − ϕi (w)

−γi
= C0 e−γi α (1 − w)e−α − C0 (1 − w)−γi = 0.
Theorem C.2 (Verification). Let (ma , mb ) ∈ ∆. For any i 6= j ∈ {a, b}, let ϕi ∈ C 1 ([0, 1]) ∩
C 2 ((0, 1 − mj )) be R+ -valued such that ϕiw is also Lipschitz continuous on (0, 1) and satisfies
(B.11)-(B.15). Let φk (x) := λ(xa + xb )1−γk ϕk (rk (x)) for any x ∈ R2+ and k ∈ {a, b}. Then the
pair (Ψa,ma , Ψb,mb ) is a Nash equilibrium and (φa , φb ) are the corresponding game values, i.e.
for any i 6= j ∈ {a, b}
φi (x) = V i (x; Aj,∗ )

j,mj
where Aj,∗· = Ψ · x, B[0,·) , Ai
[0,·] .
Proof. Let i 6= j ∈ {a, b}. We first show that
φi (x) ≥ sup J i (x; Ai , Aj,∗ ), x ∈ R2+ . (C.6)

Ai ∈A
To this end, extend ϕi to R by setting ϕi (w) := ϕi (0) for any w < 0 and ϕi (w) := ϕi (1)
for any w > 1. Let ξ ∈ C ∞ (R) be a non-negative function, compactly supported in [−1, 1] and
such that R ξ(x)dx = 1. For any m ≥ 1, let ξm (w) := ξ(mw)
R
m and define the smooth function
through convolution Z
ϕi,m (w) := ϕi (y)ξm (w − y)dy.
R
Since supp(ξm ) R⊂ [−1/m, 1/m], for any Lebesgue measurable function h on R, the value of
(h ∗ ξm )(w0 ) = R h(y)ξm (w0 − y)dy depends only on the values of h in [w0 − 1/m, w0 + 1/m].
Since ϕi is continuous on R, ϕi,m converges to ϕi as m → ∞ uniformly on any compact subsets
of R. Moreover, as ϕiw ∈ C([0, 1]), also ϕi,m
w converges to ϕiw on any compact subset of R
38

(cf. the argument in (Fleming and Soner, 2006, Appendix C)). For r > 0, define the disk
Dr (x) := {x ∈ R2 : |x| < r} and Rm,r := Rm ∩ Dr (0), where

2 1
Rm := x ∈ R+ : min{xi , xj } > .
m
Define the exit time τm,r := inf{t ≥ 0 : Ytx ∈

/ Rm,r } and the function φi,m (x) := (xa +
−λ(t∧τ )−AS
xb )1−γi ϕi,m (ri (x)). Applying Itô formula to e m,r t∧τm,r i,m
φ (Yt∧τ x
m,r
) we obtain, upon
x
taking expectations (using the abbreviation Yt = Yt (A , A )) x i j,∗
−λ(t∧τm,r )−AS
h i
φi,m (x) = E e t∧τm,r i,m
φ (Yt∧τ x
m,r
)
Z t∧τm,r
−λs−AS x,S 1−γi i i,m i,wi i,m i,wi

−E λ e s (Ys ) L ϕ (Ws ) − λϕ (Ws ) ds
0
Z t∧τm,r
S
e−λs−As (Ysx,S )1−γi (1 − Wsi,wi )ϕi,m i,wi i,m i,wi
i,c
−E λ w (W s ) − γ i ϕ (W s ) dA s
0
Z t∧τm,r
S
e−λs−As (Ysx,S )1−γi Wsi,wi ϕi,m i,wi
) + γi ϕi,m (Wsi,wi ) dAj,∗,c

+E λ w (Ws s
0X 1−γi
j,∗
S
S
i
x,S 1−γi
−E λ e−λs−As− (Ys− ) e−∆As e∆As + e∆As − 1 ϕi,m (Wsi,wi )
0≤s≤t∧τm,r

i,m i,wi
−ϕ (Ws− ) , (C.7)
where Wti,wi = ri (Ytx ) for i,wi

any t ≥ 0 with W0− = ri (x) = wi .
j,m
Since A·j,∗ = ψ· j rj (x), B[0,·) , Ai[0,·] solves SPm
j
j+
, it follows by Proposition A.10 that
0 < Wti,wi ≤ 1 − mj and 0 < ri (Ytx ) < 1 − w̃j a.s. for any t ≥ 0. Since Li ϕi is continuous on
(0, 1 − mj ), limm→∞ Li ϕi,m (w) = Li ϕi (x) for any w ∈ (0, 1 − mj ). Also, since ϕiw is Lipschitz
continuous on (0, 1), for any r > 0 there exists M > 0 such for any m ∈ N and for any x ∈ Rm,r ,
|(xi + xj )Li ϕi,m (ri (x))| < M . As
lim τm,r = τr := inf t ≥ 0 : Ytx ∈

/ R2+ ∩ Dr (0) ,

m→∞
dominated convergence implies that, with probability one,

Z t∧τm,r Z t∧τr
S S
lim e−λs−As (Ysx,S )1−γi Li ϕim (Wsi,wi )ds = e−λs−As (Ysx,S )1−γi Li ϕi (Wsi,wi )ds.
m→∞ 0 0
Using the fact that Aj,∗,c increases only at 1 − mj and the equality (B.15), letting m → ∞ and
r → ∞ in (C.7) (limr→∞ τr → ∞ a.s. as ∂R2+ is unattainable for Y x when x ∈ R2+ ), we obtain
h S
i
φi (x) = E e−λt−At φi (Ytx )
Z t
−λs−AS x,S 1−γ i i i,w i i,w

−E λ e s (Y
s )
i
L ϕ (Ws ) − λϕ (Ws ) ds
i i
(C.8)
0
Z t
−λs−AS x,S 1−γi i,wi i i,wi i i,wi
i,c
−E λ e s (Ys ) (1 − Ws )ϕw (Ws ) − γi ϕ (Ws ) dAs (C.9)
0
X 1−γi
j,∗
S
S
i
x,S 1−γi
−E λ e−λs−As− (Ys− ) e−∆As e∆As + e∆As − 1 ϕi (Wsi,wi ) (C.10)
0≤s≤t

i,wi
− ϕi (Ws− ) .
39

Note that by Lemma A.10, ∆Aj,∗ ˜j ∆Ai , W i,wi a.s., where f˜j is given by (C.1). Hence,
t = f t t−
Lemma C.1 (iii) yields
S
i j,∗
1−γi
h̃i ∆Ait , ri (Ytx ) − ϕi wi (Yt−
x
) = e−∆At e∆At + e∆At − 1 ϕi (ri (Ytx )) − ϕi ri (Yt−
x

)
≤0
a.s. for any t ≥ 0, where h̃i is given by (C.2). Together with the HJB equations (B.11)-(B.15)
and the fact that (xi + xj )1−γi U i (ri (x)) = U i (xi ) for any x ∈ R2+ , it follows that, for any t ≥ 0
h i Z t
i −λt−AS x,S 1−γ i x −λs−A S i i,x
φ (x) ≥ E λe t (Y
t ) i
ϕ (ri (Yt )) + E λ e s U (Ys )ds . (C.11)
0
h i 1−γi
1
S
Lemma A.3 implies E e−At (Ytx,S )1−γi =E {t<τA } Yt
x,S
. Using the boundedness of
ϕi , Lemma A.8, (A.18) and Jensen’s inequality yield
h i 1−γi
E e−At (Ytx,S )1−γi ϕi (ri (Ytx )) ≤ M E 1{t<τA } Ytx,S
S
h i1−γi
≤ ME 1{t<τA } Ytx,S ≤ M (xi + xj )e(1−γi )t maxk∈{a,b} µk

for some constant M ≥ max0≤w≤1 ϕi (w). Assumption 2.1 implies that
h S
i
lim E e−λt−At (Ytx,S )1−γi ϕi (ri (Ytx )) = 0.
t→∞
Letting t → ∞ for (C.11), dominated convergence theorem yields (C.6).

Next we improve (C.6) by showing that equality indeed holds: If trader i employs the
cheating strategy Ψi,m
·
i
, then

j,∗
Ai,∗
· = Ψi,mi
· x, B[0,·) , A[0,·] ,
then, by Proposition A.10, mi < ri Ytx (Ai,∗ , Aj,∗ ) < 1 − mj a.s. for almost every t ≥ 0. The

process Ai,c,∗ increases only when W i,wi is at mi ; hence the term (C.9) vanishes by (B.13). The
h i+
jump ∆Ai,∗
t = 1{t=0} ln 1−wi
1−mi is nonzero only when ri (x) < mi , and such jump brings
W0i,wi to mi , thus by Lemma C.1 (iii) the term (C.10) vanishes. Therefore, using (B.11) for the
term (C.8) equality in (C.11) follows.
Finally, letting t converge to infinity yields
φi (x) = J i (x; Ai,∗ , Aj,∗ ) = sup J i (x; Ai , Aj,∗ ).

Ai ∈A
C.1 Proofs of Theorem 2.8 and Theorem 2.9

Lemma C.3. (i) The constants cik (k = 0, 1, 2, 3) in Theorem 2.8 are strictly positive.
(ii) Let c > 0, w∗ ∈ (0, ŵi ] and suppose f ∗ (w) := c(1 − w)−γi satisfies
Li f ∗ (w∗ ) − λf ∗ (w∗ ) + U i (w∗ ) = 0, w ∈ (0, w∗ ]. (C.12)
Then Li f ∗ (w) − λf ∗ (w) + U i (w) < 0 for any w ∈ (0, w∗ ).
40

Proof. Proof of (i): First, show that for any w ∈ (0, ŵi ], (αi + βi − 1)w − αi βi > 0. If
αi + βi − 1 > 0, then clearly (αi + βi − 1)w − αi βi > 0. If αi + βi − 1 < 0, by the inequalities
w < ŵi and βi > 1 − γi (Lemma A.11 (i)),
(αi + βi − 1)w − αi βi > (αi + βi − 1)ŵi − αi βi

αi (1 − αi )(βi − (1 − γi ))
=− > 0. (C.13)
γi − αi
Since w̃i < ŵi (Lemma A.11 (iv)), it follows by (C.13) that ci0 ,ci1 ,ci2 are strictly positive, which
in turn implies ci3 > 0.
Proof of (ii): For any w ∈ (0, w∗ ), Li f ∗ (w) − λf ∗ (w) + U i (w) has the same sign as
2
1−γi γi σ
l(w) := w (1 − w) − c(1 − γi ) pi − − ki w
2
σ2
= w1−γi (1 − w)γi − c(1 − γi ) ((αi + βi − 1)w − αi βi ) . (C.14)
2
The condition C.12 implies l(w∗ ) = 0, and
1
lim l(w) = c(1 − γi )σ 2 αi βi < 0, (C.15)
w↓0 2
lww (w) = −γi (1 − γi )w−1−γi (1 − w)−2+γi < 0. (C.16)
Supposing, by contradiction, that supw∈(0,w∗ ) l(w) > l(w∗ ) = 0, by (C.15) and the strict con-
cavity (C.16), the maximum of l is attained at some z ∈ (0, w∗ ), that is, supw∈(0,w∗ ) l(w) = l(z).
Thus,
σ2
lw (z) = z −γi (1 − z)γi −1 (1 − γi − z) + c(1 − γi )(1 − αi − βi ) = 0. (C.17)
2
As z < ŵi < 1 − γi (Lemma A.11 (i)) plugging (C.17) into (C.14) yields
c(1 − γi )σ 2
l(z) = ((αi − 1)γi z + (γi z + (1 − γi − z)αi ) βi )
2(1 − γi − z)
c(1 − γi )σ 2
< (αi − 1)γi z < 0,
2(1 − γi − z)
which contradicts l(z) > 0. Hence, supw∈(0,w∗ ) l(w) ≤ 0, which implies l(w) < 0 for any
w ∈ (0, w∗ ).
C.1.1 Proof of Theorem 2.8

The proof follows by verifying that the conditions of the Verification Theorem C.2 are met:
By construction, for any i 6= j ∈ {a, b}, the function ϕi satisfies the ODEs (B.11), (B.13) and
(B.15), as well as the smooth pasting conditions
ϕi (w̃i −) = ϕi (w̃i +), (C.18)

ϕiw (w̃i −) = ϕiw (w̃i +), (C.19)
ϕiww (w̃i −) = ϕiww (w̃i +). (C.20)
As F i (w̃i , w̃j ) = 0, also the following hold:
ϕi ((1 − w̃j )−) = ϕi ((1 − w̃j )+), (C.21)

ϕiw ((1 − w̃j )−) = ϕiw ((1 − w̃j )+). (C.22)
41

By construction, ϕi ∈ C 2 ((0, w̃i )) ∩ C 2 ((w̃i , 1 − w̃j )) ∩ C 2 ((1 − w̃j , 1)). The equalities (C.18)-
(C.20) therefore imply ϕi ∈ C 2 (0, 1 − w̃j ) and equalities (C.21) and (C.22) yield ϕi ∈ C 1 (0, 1).
Due to the finiteness of the limits limw↓0 ϕi (w) = ci0 , limw↑1 ϕi (w) = ci3 , limw↓0 ϕiw (w) = γi ci0 ,
limw↑1 ϕi (w) = −γi ci3 , we may extend the function ϕi to be in C 1 ([0, 1]). Furthermore, in view
of the finite limits limw↓0 ϕiww (w) = ci0 γi (1+γi ), limw↑1 ϕiww (w) = ci3 γi (1+γi ), ϕiww ((1− w̃j )+) =
ci3 γi (1 + γi )(1 − w̃j )−2−γi ,

ϕiww ((1 − w̃j )−) = w−2 (1 − w)−2 ci1 wαi (1 − w)ai αi2 − (1 − 2γi w)αi − (1 − γi )γi w2

γ i

+ ci2 wβi (1 − w)bi βi2 − (1 − 2γi w)βi − (1 − γi )γi w2 −

qi w 1+γi

w=1−w̃j

and the continuity of ϕiww on the regions (0, 1−w̃j ) and (1−w̃j , 1), it follows that supw∈(0,1) ϕiww (w) <
∞, whence ϕiw is Lipschitz continuous.
The inequality (B.12) follows from w̃i < ŵi in conjunction with Lemma C.3 (i) and (ii). To
check the inequality (B.14), note that for any w ∈ (w̃i , 1 − w̃j ), (1 − w)ϕiw (w) − γi ϕi (w) has
thee same sign as
(wj − γi )F i (w̃i , wj ) + γi ai bi li (wj ) =: hi (wj ),
where wj := 1 − w (so wj ∈ (w̃j , 1 − w̃i )) and
−αi ai
i wj w̃i
l (wj ) := (γi w̃i + (1 − γi − w̃i )βi )
1 − wj 1 − w̃i
1 − w̃i −bi
βi
1 − wj
− (γi w̃i + (1 − γi − w̃i )αi ) .
wj w̃i
As w̃i < ŵi implies γi w̃i + (1 − γi − w̃i )αi < 0, it follows that li (wj ) > 0.
By Lemma A.11 (ii) and Lemma A.12 (ii), (w̃i , w̃j ) ∈ (wi , f i,−1 (wi )) : wi ∈ (0, ŵi ) . It

follows that f i,−1 (w̃i ) = w̃j , and Lemma A.12 (iii) yields that
F i (w̃i , wj ) > 0 for any wj ∈ (w̃j , 1 − w̃i ). (C.23)
If wj ≤ γi , by Lemma A.11 (i) and (C.23) it follows that hi (wj ) < 0. It remains to check
1−wj 1−γi

the case wj > γi : Factoring out wj from hi yields
sgn hi (wj = sgn h̄i (wj ) ,

where
1 − w̃i −bi 1 − wj −bi

i
h̄ (wj ) := ai (wj − bi − γi ) (γi w̃i + (1 − γi − w̃i )αi )
w̃i wj
ai ai
w̃i wj
+ bi (ai + γi − wj ) (γi w̃i + (1 − γi − w̃i )βi )
1 − w̃i 1 − wj
+ (ai − bi )(wj − γi )(w̃i (αi + βi − 1) − αi ).
It follows from wj > γi and Lemma A.11 (i) that wj − γi − bi > 0, γi w̃i + (1 − γi − w̃i )αi < 0
1−wj
and ai + γi − wj > 1 − wj > 0. The inequalities 1− w̃i
wj > 1 and w̃i > 1 imply that
h̄i (wj ) < ai (wj − bi − γi ) (γi w̃i + (1 − γi − w̃i )αi ) + bi (ai + γi − wj ) (γi w̃i + (1 − γi − w̃i )βi )
+ (ai − bi )(wj − γi )(w̃i (αi + βi − 1) − αi βi )
= ai bi (1 − w̃i − wj )(βi − αi ) < 0.
Therefore, (1−w)ϕiw (w)−γi ϕi (w) < 0 for any w ∈ (w̃i , 1−w̃j ), proving the inequality (B.14).
In the proof of Theorem 2.9 below, we use an auxiliary statement similar to Lemma C.1(iii).
The similar, but simpler proof, is omitted for brevity.
42

Lemma C.4. For any i 6= j ∈ {a, b} and for any mi ∈ (0, 1), let ϕi ∈ C 1 ((0, 1)) be an R+ -valued
function satisfying (B.13)–(B.14). Then for any α ≥ 0 and any w ∈ (0, 1), the function
ĥi (α, w) := e−αγi ϕi 1 − e−α (1 − w)

satisfies
ĥi (α, w) − ϕi (1 − w) ≤ 0,
where the equality holds for a fixed w, if and only if
α = 0, if w ∈ (mi , 1),

1−w
α ≤ ln , if w ∈ (0, mi ].
1 − mi
C.1.2 Proof of Theorem 2.9

A direct calculation yields that ϕ̂i satisfies
Li ϕ̂i (w) − λϕ̂i (w) + U i (w) = 0, w ∈ (ŵi , 1), (C.24)

(1 − w)ϕ̂iw (w) − γi ϕ̂i (w) = 0, w ∈ (0, ŵi ), (C.25)
and
ϕ̂i (ŵi −) = ϕ̂i (ŵi +), (C.26)

ϕ̂iw (ŵi −) = ϕ̂iw (ŵi +), (C.27)
ϕ̂iww (ŵi −) = ϕ̂iww (ŵi +). (C.28)
As ϕ̂i ∈ C 2 ((0, ŵi )) and ϕ̂i ∈ C 2 ((ŵi , 1)), it follows by (C.26), (C.27) and (C.28) that ϕ̂i ∈
C 2 ((0, 1)). Moreover, Lemma A.11 (i) implies that si0 > 0 and si1 > 0. Hence, by Lemma C.3
(ii)
Li ϕ̂i (w) − λϕ̂i (w) + U i (w) < 0, w ∈ (0, ŵi ). (C.29)
Next, we prove that
(1 − w)ϕ̂iw (w) − γi ϕ̂i (w) < 0, w ∈ (ŵi , 1). (C.30)
To this end, note that for any w ∈ (ŵi , 1),
sgn (1 − w)ϕ̂iw (w) − γi ϕ̂i (w) = sgn li (w) ,

where ai
i 1 − γi − w 1−w
l (w) = − si1 (w − αi ).
(1 − γi )qi w
Also,
γi
lim li (w) = lim lw
i
(w) = 0, and lim li (w) = − < 0,
w↓ŵi w↓ŵi w↑1 (1 − γi )qi
and
i

sgn lww (w) = sgn ((γ − αi )w + αi (1 + ai )) .
As w ∈ (ŵi , 1), it follows that
(γ − αi )w + αi (1 + ai ) ∈ ((1 − αi )αi , (1 − αi )(γi + αi )) .
Distinguish two cases:
43

i (w) < 0 and hence,
(i) If γi + αi ≤ 0, then lww
i i
lw (w) < lim lw (w) = 0.
w↓ŵi
Therefore, an ODE comparison argument yields that li < 0 on (ŵi , 1).

i
(ii) If γi + αi > 0, then lwwi (w) ≤ 0 on ŵ , −αi (1+ai ) and it follows again by an ODE
i γi −αi
i
comparison argument that li (w) < 0 for any w ∈ ŵi , −αγii(1+a
−αi
i)
. As li is strictly convex

on the interval −αγii(1+ai)
−αi , 1 , and l
i −αi (1+ai ) < 0, lim
γi −αi
i
w↑1 l (w) < 0, we conclude that

li < 0 on −αγii(1+ai)
−αi , 1 .
i
This completes the proof of subclaim (C.30). An application of Itô’s Lemma to e−λt−A φ̂i (Ytx )
yields, upon taking expectations (The dependence of Y x and W i,wi on (Ai , 0) is omitted for the
sake of brevity.)
h i
i
φ̂i (x) = E e−λt−At φ̂i (Ytx )
Z t
i
e−λs−As (Ysx,S )1−γi Li ϕ̂i (Wsi,wi ) − λϕ̂i (Wsi,wi ) ds

−E λ
0
Z t
−λs−Ais x,S 1−γi i,wi i i,wi i i,wi
i,c
−E λ e (Ys ) (1 − Ws )ϕ̂w (Ws ) − γi ϕ̂ (Ws ) dAs
0
 
i
i

x,S 1−γi i,wi 
X
− E λ e−λs−As− (Ys− ) e−γi ∆As ϕ̂i (Wsi,wi ) − ϕ̂i (Ws− ) . (C.31)
0≤s≤t
By Lemma C.4 and (C.25), (C.30), the term in the last line of (C.31) is non-negative. Using
(C.24)–(C.25) and (C.29)–(C.30), similar arguments as in the proof of Theorem C.2 yield
φ̂i (x) ≥ sup J i (x; Ai , 0), x ∈ R2+ . (C.32)

Ai ∈A
Finally, using the properties of ψ·i,ŵi ri (x), B[0,·) , 0 established in Proposition A.10, the equality

in (C.32) follows.
References
Anna Aksamit and Monique Jeanblanc. Enlargement of Filtration with Finance in View.
SpringerBriefs in Quantitative Finance. Springer International Publishing, 2017.
John Armstrong and Damiano Brigo. Risk managing tail-risk seekers: VaR and expected
shortfall vs S-shaped utility. Journal of Banking & Finance, 101:122–135, 2019.
John Bather and Herman Chernoff. Sequential decisions in the control of a spaceship. In Fifth
Berkeley Symposium on Mathematical Statistics and Probability, volume 3, pages 181–207,
1967.
Tomasz R Bielecki and Marek Rutkowski. Credit risk: modeling, valuation and hedging.
Springer-Verlag Berlin Heidelberg, 2004.
Stephen J Brown and Onno W Steenbeek. Doubling: Nick Leeson’s trading strategy. Pacific-
Basin Finance Journal, 9(2):83–99, 2001.
44

René Carmona. Lectures on BSDEs, stochastic control, and stochastic differential games with
financial applications, volume 1. SIAM, 2016.
Jennifer N Carpenter. Does option compensation increase managerial risk appetite? The
Journal of Finance, 55(5):2311–2331, 2000.
Samuel N Cohen and Robert James Elliott. Stochastic calculus and applications, volume 2.
Springer, 2015.
Basel Committee et al. Principles for the sound management of operational risk. Basel, Switzer-
land: Bank for International Settlements Communications, 2011.
Michel G Crouhy, Dan Galai, and Robert Mark. Insuring versus self-insuring operational risk:
Viewpoints of depositors and shareholders. The Journal of Derivatives, 12(2):51–55, 2004.
Mark HA Davis and Andrew R Norman. Portfolio selection with transaction costs. Mathematics
of operations research, 15(4):676–713, 1990.
Tiziano De Angelis and Giorgio Ferrari. Stochastic nonzero-sum games: a new connection
between singular control and optimal stopping. Advances in Applied Probability, 50(2):347–
372, 2018.
Jodi Dianetti and Giorgio Ferrari. Nonzero-Sum Submodular Monotone-Follower Games: Ex-
istence and Approximation of Nash Equilibria. arXiv preprint arXiv:1812.09884, 2018.
Jefferson Duarte, Francis A Longstaff, and Fan Yu. Risk and return in fixed-income arbitrage:
Nickels in front of a steamroller? The Review of Financial Studies, 20(3):769–811, 2006.
Erik Ekström, Kristoffer Lindensjö, and Marcus Olofsson. How to detect a salami slicer: a
stochastic controller-stopper game with unknown competition, 2020.
Wendell H Fleming and Halil Mete Soner. Controlled Markov processes and viscosity solutions,
volume 25. Springer Science & Business Media, 2006.
Xin Guo and Renyuan Xu. Stochastic Games for Fuel Followers Problem: N versus MFG.
arXiv preprint 1803.02925, 2018.
Xin Guo, Wenpin Tang, and Renyuan Xu. A class of stochastic games and moving free boundary
problems. arXiv preprint arXiv:1809.03459, 2018.
Rhys Gwilym and M Shahid Ebrahim. Can position limits restrain ‘rogue’trading? Journal of
Banking & Finance, 37(3):824–836, 2013.
Said Hamadène and Rui Mu. Existence of Nash equilibrium points for Markovian non-zero-
sum stochastic differential games with unbounded coefficients. Stochastics An International
Journal of Probability and Stochastic Processes, 87(1):85–111, 2015.
Toshihide Iguchi. My billion dollar education: inside the mind of a rogue trader. Toshihide
Iguchi, 2014.
Jean Jacod. Calcul stochastique et problemes de martingales, volume 714. Springer, 2006.
Jean Jacod and Albert Shiryaev. Limit theorems for stochastic processes, volume 288. Springer
Science & Business Media, 2013.
Monique Jeanblanc and Yann Le Cam. Progressive enlargement of filtrations with initial times.
Stochastic Processes and their Applications, 119(8):2523–2543, 2009.
45

Thierry Jeulin. Semi-martingales et grossissement d’une filtration, volume 833. Springer, 2006.
Ioannis Karatzas and Qinghua Li. BSDE approach to non-zero-sum stochastic differential games
of control and stopping. In Stochastic Processes, Finance and Control: A Festschrift in Honor
of Robert J Elliott, pages 105–153. World Scientific, 2012.
Ioannis Karatzas and Steven E Shreve. Brownian motion and stochastic calculus. Springer-
Verlag New York, 1998.
Daphne A Kenyon. Big bets gone bad: Derivatives and bankruptcy in orange county, 1997.
Kimberly D Krawiec. Accounting for greed: Unraveling the rogue trader mystery. Oregon Law
Review, 79:301, 2000.
Kimberly D Krawiec. The return of the rogue. Arizona Law Review, 51:127, 2009.
H Dharma Kwon and Hongzhong Zhang. Game of singular stochastic control and strategic exit.
Mathematics of Operations Research, 40(4):869–887, 2015.
Jennifer Moodie. Internal systems and controls that help to prevent rogue trading. Journal of
securities operations & custody, 2(2):169–180, 2009.
Bernt Øksendal and Agnès Sulem. Applied Stochastic Control of Jump Diffusions. Springer,
2019.
Daniel Revuz and Marc Yor. Continuous martingales and Brownian motion, volume 293.
Springer Science & Business Media, 2013.
Hiroshi Tanaka. Stochastic differential equations with reflecting boundary condition in convex
regions. pages 157–171, 2002.
Mark N Wexler. Financial edgework and the persistence of rogue traders. Business and Society
Review, 115(1):1–25, 2010.
Yuqian Xu, Michael Pinedo, and Mei Xue. Operational risk in financial services: A review and
new research opportunities. Production and Operations Management, 26(3):426–445, 2017.
Yuqian Xu, Lingjiong Zhu, and Michael Pinedo. Operational risk management: A stochastic
control framework with preventive and corrective controls. Operations Research, 68(6):1804–
1825, 2020.
46

SSRN Id3870658

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SSRN Id3870658

Uploaded by

Copyright:

Available Formats

Michael J.

Brennan Irish Finance

This paper can be downloaded without charge from

Electronic copy available at: https://ssrn.com/abstract=3870658

June 30, 2021

Electronic copy available at: https://ssrn.com/abstract=3870658

Electronic copy available at: https://ssrn.com/abstract=3870658

Electronic copy available at: https://ssrn.com/abstract=3870658

1 A Model of Rogue Trading

Electronic copy available at: https://ssrn.com/abstract=3870658

1.1 Investment and Fraud in Continuous Time

Electronic copy available at: https://ssrn.com/abstract=3870658

dYti,x = µi Yti,x dt + σi Yti,x dBti + Yt−

where the integral with i

that τA is a stopping time.2

Electronic copy available at: https://ssrn.com/abstract=3870658

Proposition 1.1 (No fraud for N = 1). Let N = 1 and h τ1 be an F-measurable,

(i) If τ1 is unbounded, then A1,∗

Electronic copy available at: https://ssrn.com/abstract=3870658

J i (x; Ai , Aj ) := E U i Xτi,x (Ai , Aj )

V i (x; Aj ) := sup J i (x; Ai , Aj )

(ii) For any c > 0, J i (cx; Ai , Aj ) = c1−γi J i (x; Ai , Aj ).

Electronic copy available at: https://ssrn.com/abstract=3870658

(ii) Non-cooperative optimality holds, i.e.,

2.2 Construction of Nash equilibrium

(i) mi ≤ Wti,wi (Ai,∗ , Aj ) < 1 a.s. for all t ≥ 0.

Electronic copy available at: https://ssrn.com/abstract=3870658

Lemma 2.7. There exists (w̃a , w̃b ) ∈ ∆ such that

F a (w̃a , w̃b ) = F b (w̃b , w̃a ) = 0. (2.4)

Electronic copy available at: https://ssrn.com/abstract=3870658

V i (x; 0) = λ(xa + xb )1−γi ϕ̂i (ri (x))

Electronic copy available at: https://ssrn.com/abstract=3870658

0.9 0.9 0.9 0.9

0.8 0.8 0.8 0.8

0.7 0.7 0.7 0.7

0.6 0.6 0.6 0.6

0.5 0.5 0.5 0.5

0.4 0.4 0.4 0.4

0.3 0.3 0.3 0.3

0.2 0.2 0.2 0.2

0.1 0.1 0.1 0.1

0.9 0.9 0.9 0.9

0.8 0.8 0.8 0.8

0.7 0.7 0.7 0.7

0.6 0.6 0.6 0.6

0.5 0.5 0.5 0.5

0.4 0.4 0.4 0.4

0.3 0.3 0.3 0.3

0.2 0.2 0.2 0.2

0.1 0.1 0.1 0.1

Electronic copy available at: https://ssrn.com/abstract=3870658

Electronic copy available at: https://ssrn.com/abstract=3870658

Fig. 2: Equilibrium average fraud (vertical axis), up to horizon or bankruptcy, of traders a

where E(S)0− = 1 and Sc denotes the continuous part of S.

Electronic copy available at: https://ssrn.com/abstract=3870658

c = 0, therefore the stochastic exponential simplifies to

(ii) For all i ∈ {1, . . . , N } and any t ≥ 0

dYt = diag(Yt )dRt + trace (diag(Yt− )) IN dÃt , Y0− = x ∈ Rn+ , (A.3)

By (Jacod, 2006, Theorem 6.8), it follows that

Electronic copy available at: https://ssrn.com/abstract=3870658

An application of (Jacod, 2006, Theorem 6.8) yields (ii).

The following lemma characterizes the conditional probability of bankruptcy in relation to