Evolutionary Portfolio Selection With Liquidity Shocks
Evolutionary Portfolio Selection With Liquidity Shocks
Evolutionary Portfolio Selection With Liquidity Shocks
Abstract
Insurance companies invest their wealth in financial markets. The wealth evolution
strongly depends on the success of their investment strategies, but also on liquidity
shocks which occur during unfavourable years, when indemnities to be paid to the
clients exceed collected premia. An investment strategy that does not take liquidity
shocks into account, exposes insurance companies to the risk of bankruptcy, when
liquidity shocks and low investment payoffs jointly appear. Therefore, regulatory au-
thorities impose solvency restrictions to ensure that insurance companies are able to
face their obligations with high probability. This paper analyses the behaviour of in-
surance companies in an evolutionary framework. We show that an insurance company
that merely satisfies regulatory constraints will eventually vanish from the market. We
give a more restrictive no bankruptcy condition for the investment strategies and we
characterize trading strategies that are evolutionary stable, i.e. able to drive out any
mutation.
1
positive. In particular, if an investor is the unique survivor at some point in time, the no
bankruptcy condition is sufficient but also necessary to avoid almost surely going bankrupt.
Nevertheless, while investors with strategies satisfying the no bankruptcy condition will not
almost surely go bankrupt, we also show that investors who use the simple strategy that
corresponds to the no bankruptcy boundary, will eventually disappear from the market.
Moreover, we characterize trading strategies that are evolutionary stable, if they exist, fol-
lowing the idea first introduced by Hens and Schenk-Hoppé (2002a). We give the condition
on the dividend process and liquidity shocks factor for the existence of evolutionary stable
strategies, when the state of the world follows an i.i.d. process. We show that the condition
for the existence of evolutionary stable strategies is related to the growth rate of the trad-
ing strategies in a neighbourhood of the strategy investing according to the no bankruptcy
boundary. If this growth rate is strictly positive, then an investor putting more than the no
bankruptcy boundary on the risky assets is able to further increase her market share, when
asset prices are dominated by the strategy corresponding to the no bankruptcy boundary.
However, this is true only as long as liquidity shocks do not force the investor to use all her
wealth, or else she will disappear from the market. Therefore, while it can happen that the
growth rate of a strategy in a neighbourhood above the no bankruptcy boundary is positive,
this strategy cannot be evolutionary stable, since it will almost surely disappear, because of
liquidity shocks. In this case, no evolutionary stable strategies could exist. This result also
suggest that in the presence of liquidity shocks, evolutionary stability should be character-
ized in term of both the growth rate and the probability of default.
This work contributes to the development of the evolutionary portfolio theory, that started
with the seminal paper of Blume and Easley (1992), where an asset market model is first
introduced to study the market selection mechanism and the long run evolution of investors’
wealth and assets’ prices. In their model, Blume and Easley (1992) consider diagonal se-
curities1 , with no transaction costs and positive proportional saving rates are exogenously
given. In the case of complete markets with diagonal securities, Blume and Easley (1992)
show that there is a unique attractor of the market selection mechanism and prices do not
matter. With simple strategies2 and constant, identical saving rates across investors, the
unique survivor is the portfolio rule known as “betting your beliefs” (Breiman 1961), where
the proportion of wealth to be put on each asset is the probability of the corresponding state
of nature. This strategy can also be generated by maximizing the expected logarithm of
relative returns, which is know as the Kelly rule, studied in discrete-time by Kelly (1956),
Breiman (1961), Thorp (1971) and Hakansson and Ziemba (1995) (for an overview, see
also Ziemba 2002) and, in continuous-time, by Pestien and Sudderth (1985), Heath, Orey,
Pestien, and Sudderth (1987) and Karatzas and Shreve (1998), among others. Hens and
1
A system of securities is called diagonal, if for each state of nature there is exactly one asset which has
a strictly positive payoff.
2
A portfolio rule called a simple strategy, if the proportion of wealth put on each asset is constant over
time.
2
Schenk-Hoppé (2002a) proposed a more general setting, with incomplete markets, general
short-lived assets that re-born each period and constant, positive, proportional and, identical
saving rates across investors. In their evolutionary model, the equilibrium notion refers to
wealth distributions that are invariant under the market selection process. The authors show
that invariant wealth distributions are generated by a population, where only one investor
(or portfolio rule) exists (a so-called monomorphic population). Moreover, they introduce
the concept of evolutionary stable portfolio rules, that is also considered in this paper. The
main result of Hens and Schenk-Hoppé (2002a) is that, in the case of ergodic state of the
world processes and without redundant assets, there is a unique evolutionary stable portfolio
rule, which is the one that puts on each asset the proportion of wealth corresponding to the
expected relative payoff of the asset. In Evstigineev, Hens, and Schenk-Hoppé (2003) this
result is extended to a model with long-lived assets, under the assumption of Markow state
of the world. Introducing long-lived assets allows to take into account the capital gains and
losses due to assets’ prices changes. This will also be of much importance in the presence
of liquidity shocks, as we will discuss in this paper. Moreover, in Evstigneev, Hens, and
Schenk-Hoppé (2002) it is also shown that, with independent and identically distributed
state of world processes, the strategy that invests according to relative dividends is the
unique simple portfolio rule that asymptotically gathers total wealth. A generalization of
the results obtained by Blume and Easley (1992). Sandroni (2000), and Blume and Easley
(2002) have also studied the case of long-lived assets, to include market prices in the evolu-
tion of wealth shares. The main result of Blume and Easley (2002) and Sandroni (2000), is
that, with complete markets, among all infinite horizon expected utility maximizers, those
who happen to have rational expectation will eventually dominate the market and this result
holds independently of investors’ risk aversion. In his model Sandroni (2000) also includes
endogenously determined positive and proportional saving rates.
All these models assume that withdrawals and savings are a positive proportion of the cur-
rent wealth, so that bankruptcy is excluded in their setup. Moreover, e.g. in Evstigineev,
Hens, and Schenk-Hoppé (2003), the withdrawal rates are assumed to be identical among
investors. Under these assumptions, the only criterion that matters for a trading strategy to
be evolutionary stable, is its exponential growth rate in the presence of a mutant strategy.
This paper shows that with non-proportional and maybe negative withdrawal rates, a second
criterion has to be considered, since in fact, even if a strategy has the maximal exponential
growth rate in the presence of any mutant, it can disappear because exogenously determined
liquidity shock occurs.
In the classical finance approach with exogenously given price dynamics, asset-liability
management models already assume that investors maximize the investment’s expected pay-
off less penalties for bankruptcy or targets not meet (see Carino, Myers, and Ziemba 1998,
Carino and Ziemba 1998). Liu, Longstaff, and Pan (2003) consider a price dynamic for the
risky asset with jumps (event risk) and take utility functions identical to −∞ for strictly
negative terminal wealth, so that no portfolio rule, that has a strictly positive probability
3
of going bankrupt, will be optimal. They obtain lower (since they do not exclude short-
selling) and upper bounds for the proportion of wealth to be put on the risky asset and they
provide optimal portfolio weights. Alternatively, Browne (1997) distinguishes between the
survival problem and the growth problem. He first looks at portfolio rules that maximize
the probability of surviving in the so-called danger-zone (where bankruptcy has strictly pos-
itive probability to occur) and second, he considers portfolio rules that maximize the growth
rate in the safe-zone, where bankruptcy is almost surely excluded. Browne (1997) identifies
wealth-level dependent strategies, but in his time-continuous setup, no optimal strategy is
found for the danger-zone, and a weaker optimality criterion is introduced. The optimal
strategy for the safe-zone corresponds to a generalization of the Kelly criterion previously
discussed. Zhao and Ziemba (2000) propose a model with a reward function on minimum
subsistence, i.e. the objective function to maximize equals the sum of the expected final
wealth and a concave increasing function on the supremum over the wealth levels that are
almost surely smaller than final wealth. In this way, the optimal portfolio rule solves a
trade-off between expected payoff and minimum subsistence.
The rest of this paper is organized as follows. In the next section we present the model
setup. In Section 3 we derive the no bankruptcy condition on investment strategies, that
ensures that liquidity shocks do not cause bankruptcy. In Section 4 we present the main
results of the paper. Section 5 concludes. Technical results and proofs are given in the
Appendix.
4
i is given by
wti = mit + qt ait . (1)
The wealth of investor i evolves as follows3
i
wt+1 = (1 + r) mit + (Dt+1 + qt+1 ) ait − Ct+1
i
. (2)
We say that investor i goes bankrupt during period (t, t + 1] (or simply period t + 1) iff
i
wt+1 ≤ 0. In this case she uses all her wealth to pay the indemnities and vanishes from
the market, i.e. we arbitrarily write mis = ais = 0 for all s ≥ t + 1 (and thus we also set
wsi = 0 for all s ≥ t + 1). Note that the investor’s wealth at time t + 1 also depends on the
price qt+1 of the risky asset, which is determined at equilibrium by investors’ demand for the
risky asset and supply. Thus, time t + 1 investors’ strategies may cause a bankruptcy. Let
It = {i | wti > 0} be the set of investors, who survive period t. Obviously, It ⊆ It−1 and thus
mit = ait = 0 for all i 6∈ It−1 . Investor j is said to be the unique survivor at time t if and
only if It = {j}.
i
The next period amount Ct+1 is determined by the following. At time t investor i can
i
decide to sell δt ≥ 0 insurance contracts on one single future stochastic claim Xt+1 ≥ 0
(which is identical for all investors). The premium Pt+1 of each contract is F t -measurable
(depends only on information available up to time t), determined by the market clearing
condition on the insurance market at time t, and is paid at time t + 1 by the buyer of the
insurance contract, who is supposed to be external to the economy just defined, i.e. buyers
of insurance contracts do not participate to the financial market. The amount collected or
withdrawn by investor i at time t + 1 is then given by the difference
i
Ct+1 = δti (Xt+1 − Pt+1 )
where for x ∈ R, x+ = max(0, x). We prefer to keep the notation simpler and since we are essentially
looking at strategies that survive in the long run, the wealth evolution of those strategies is correctly given
by equation (2).
5
The parameters α and ǫ are exogenously given by regulatory authorities. Equation (3)
defines the pricing rule for insurance contracts and is called quintile principle and it has
been discussed in Schnieper (1993) and Embrechts (1996). Moreover, it corresponds to
the proportional value-at-risk constraints studied by Leippold, Vanini, and Trojani (2003)
with time independent proportional factors, and sameness between investors in their general
equilibrium consideration. The parameters αti and ǫit are fixed and can be interpreted as the
“loss acceptability” of investor i and, in the general setting of the model, we assume that
they can vary between investors. Other simplifying assumptions will be introduced later. For
a given premium Pt+1 and parameters αti and ǫit , equation (3) serves to compute the number
δti of insurance contracts that investor i can sell, in order to satisfy the solvency constraint.
The premium Pt+1 is determined endogenously when the insurance market clears.
We should bear in mind that for an investor, going bankrupt means vanishing from
the market and thus should be avoided! They can further decrease their insurance risk by
choosing a smaller αi or a smaller ǫi . As we will see below, an investor with a small αi , who
is a “safer investor” with respect to minimal solvency requirement, is also forced to reduce
her exposure to the insurance market, “losing” in this way growth opportunities when claims
are less than premia. The amount αti wti represents the technical reserve or the proportion of
current wealth to be invested prudently by investor i to make the risky insurance business
acceptable in the future (see Norberg and Sundt 1985). If the amount δti (Xt+1 − Pt+1 ) is
strictly greater than the technical reserves, we say that investor i faces a liquidity shock .
From equation (3), investor i faces liquidity shocks with probability ǫit during period t + 1.
αi
In our setting, the solvency constraint essentially imposes that mit ≥ Rt wti , for all i, or
equivalently
mit αti
≥ , ∀i ∈ It . (5)
wti R
£ ¤
Let µt = E Xt | F t−1 and σt2 = Var(Xt | F t−1 ) be the conditional expectation and the con-
ditional variance of Xt , given F t , respectively, and let Ft be the conditional cumulative
distribution function of Yt = Xtσ−µ
t
t
, i.e.
£ ¤
Ft (y) = P Yt ≤ y | F t−1 .
Moreover, Ft−1 denotes the generalized inverse of Ft . To avoid the premium Pt+1 fully
covering the insurance risk, we impose the following restrictions
Assumption 1 (Insurance market). For t ∈ Z and i ∈ It , let (αti , ǫit ) and (α̃ti , ǫ̃it ) be two
possible choices for the loss acceptability parameters of investor i. Let αti = α̃ti . Then for all
premia Pt+1
δti > δ˜ti ⇒ ǫit > ǫ̃it .
This assumption says that for given technical reserves, the probability of having liquidity
shocks strictly increases with the number of insurance contracts sold. If this is not satisfied,
6
then it would be possible to cover additional insurance risk only through collected premia,
which is not a fair pricing rule. Since δti = 0 solves equation (3) with αti = 0 and ǫit = 0,
Assumption 1 also implies that an insurance company without technical reserves that sells a
strictly positive number of contracts, faces liquidity shocks with a strictly positive probability.
Assumption 1 indirectly imposes restrictions on equilibrium premia, as shown in the following
lemma.
Lemma 1. If Assumption 1 holds, then for all t ∈ Z and i ∈ It :
−1
Pt+1 < µt+1 + σt+1 Ft+1 (1 − ǫit ).
or
αti wti
δti = −1 . (7)
µt+1 + σt+1 Ft+1 (1 − ǫit ) − Pt+1
Lemma 1 ensures that δti ≥ 0. Equation (7) says that investor i supply for insurance con-
tracts is proportional to her technical reserve and decreases with increasing probability ǫit .
For a fixed supply of insurance contracts, an investor can therefore decrease her technical
reserve by decreasing her liquidity shock probability ǫit . Naturally, the solvency constraints
(3) and (4) do not take into account the magnitude of a liquidity shock! This is a well know
critique of quintile constraints (see e.g. Artzner, Delbaen, Eber, and Heath 1997).
P i
We assume that demand for insurance contracts is normalized to 1, i.e. i δt = 1 for all
t. It follows: X X
−1
Pt+1 = µt+1 + σt+1 δti Ft+1 (1 − ǫit ) − αti wti . (8)
i i
7
P −1
P
σt+1 i δti Ft+1 (1 − ǫit ) − i αti wti is the so-called loading factor and is supposed to be strictly
positive. In fact, it is well known from the ruin theory, that if Pt+1 ≤ µt+1 , i.e. if the
premium at time t is less or equal to the conditional expectation of next period claims given
all information available at time t, then for any value for the initial wealth (without financial
market) the probability of going bankrupt is equal one (see Feller 1971, page 396). From
the last equation, we see that the premium of the insurance contract increases with in-
creasing
PI conditional variance, as one would expect, and decreases when the weighted wealth
i i i i
i=1 αt wt increases. Moreover, a safer investor, with a smaller α or a smaller ǫ than a
riskier investor, contributes to an increase of the premium, from which all investors benefit.
This behaviour has also been described by Ceccarelli (2002). Equations (7) and (8) can be
solved for δti and Pt+1 : they provide a unique solution with a strictly positive premium (this
will become clear for the special case considered below; however, we give a general proof of
the existence and uniqueness of a solution in the Appendix 6.1).
Since the goal of this paper is to analyse investors’ long-run wealth evolution with respect
to their investment strategies on financial markets, we assume that their profiles on insurance
markets are identical, meaning that they possess the same loss acceptability parameters.
Here, we do not address the question of investors’ strategies (choice of the loss acceptability
parameters) on the insurance market. It is not clear whether an investor who has higher
loss acceptability, will growth faster or not. In fact, while it is true by equation (7) that
higher loss acceptability means greater liquidity shocks (for both the probability and the
amount), it must also be said that investors who sell a larger number of contracts benefit
from growth opportunities when premia are greater than claims. Moreover, less technical
reserves means a smaller exposure to liquidity shocks (as discussed above), but also less
restrictive constraints for the investment strategies, meaning that those investors can put
less money into the risk-free asset and profit from growth opportunities on the financial
market. We address these issues in other works. Here, as in Leippold, Vanini, and Trojani
(2003) we make the following assumption.
Note that by equations (7) and (8), when all investors possess the same ǫit , δti does not depend
on ǫit anymore and therefore the magnitude of liquidity shocks is minimized for all investors
if ǫit = ǫ. Moreover, by Assumption 2 and equations (7) and (8) it follows
X
−1
Pt+1 = µt+1 + σt+1 Ft+1 (1 − ǫ) − α wti , (9)
i
wti
δti = P , (10)
j wtj
8
and therefore investor’s i supply for insurance contracts corresponds to her relative wealth.
Now, we introduce a precise structure for the claim Xt+1 . In particular, we assume that
the total claim Xt+1 is proportional to the aggregate wealth available at time t, meaning
that the amount of insured claims increases or decreases depending on the aggregate success
of the investors (a similar assumption will be also made for the dividend process). This
assumption also prevents a shock from destroying the economy. The proportional factor is
supposed to be independent of the history up to time t and can be interpreted as the liquidity
shock factor for the economy. Mathematically we have
Xt+1 = ηt+1 Wt , (11)
P P
where ηt+1 ∈ [0, 1] is independent of F t and Wt = i∈It wti = Ii=1 wti is the aggregate wealth
£ ¤
available in the economy at time t. From equation (11) it follows that µt+1 = Wt E ηt+1
2
and σt+1 = Wt2 Var(ηt+1 ). Moreover, ηt+1 ∼ Gt+1 where Ft+1 (y) = Gt+1 ( Wy t ), ∀y. Thus
¡ ¢
Pt+1 = µ(ηt+1 ) + σ(ηt+1 ) G−1
t+1 (1 − ǫ) − α Wt .
Therefore, the premium Pt+1 is strictly positive for all t, if the loading factor (σ(ηt+1 ) G−1
t+1 (1−
ǫ) − α) Wt is greater than zero for all t. Moreover, for the sake of simplicity, we make the
following assumption:
Assumption 3 (Liquidity shocks).
Liquidity shocks (ηt )t≥1 are independent and identically distributed, i.e. Gt = G for all t,
ηt ∼ η ∼ G, where G is a continuous cumulative distribution function.
£ ¤
Let µ = E η and σ 2 = Var(η), then by Assumptions 2 and 3,
Pt+1 = µ Wt + σ G−1 (1 − ǫ) Wt − α Wt = (β − α) Wt , (12)
i
Ct+1 = (ηt+1 − β + α) wti , (13)
where β = µ + σ G−1 (1 − ǫ). As discussed above for the general case, we impose that the
loading factor (σ G−1 (1 − ǫ) − α) Wt is strictly positive, i.e. α < min{α, σ G−1 (1−ǫ)}. Then
β − α > β − σ G−1 (1 − ǫ) = µ > 0 and thus Pt+1 > 0 for all t.
We now turn our attention to the financial market. We suppose that the risky asset
is in fixed supply, normalized to one. Instead, the supply of cash is exogenously given by
cumulated dividends and collected premia less withdrawals. The market clearing conditions
are
I
X X
ait = ait = 1 (14)
i=1 i∈It
à !
X X X
Mt = R mit−1 + Dt ait−1 − qt 1− ait−1 − Cti (15)
i∈It i∈It i∈It
9
P P P P
where Mt = i∈It mit and Ct = i∈It Cti . Note that i∈It mit−1 ≤ Mt−1 = i∈It−1 mit−1
since It ⊆ It−1 . Moreover, if no bankruptcy occurs during period t, then It = It−1 and the
usual equation for Mt follows, i.e. Mt = R Mt−1 + Dt − Ct . Note that Mt ≥ 0 for all t. In
fact if for some t, Mt < 0, then there exists at least one investor, say j ∈ It , with mjt < 0.
But since borrowing is not allowed, investor j is forced into bankruptcy during period t, a
contradiction to j ∈ It .
To be consistent with Assumption 3, and in order to avoid that dividends become very small
as compared to insurance shocks, we make the following assumption for the dividend process:
(ii) £ ¤
P d > 0 = 1 − H(0) ∈ (0, 1),
i.e. at each time dividends have strictly positive probability of being zero and of being
strictly positive.
This assumption, together with Assumption 3, solves the difficulty encountered by Hens
and Schenk-Hoppé (2002b), where the rate of return on the long-lived asset eventually dom-
inates that of the numéraire, so that the strategy that invests only in long-lived asset is able
to drive out any other strategy. Hens and Schenk-Hoppé (2002b) suggest to base evolution-
ary finance model on Lucas (1978), where assets’ payoffs are in term of a single perishable
consumption good. In this way, the consumption rate is at least as the growth rate of the
total payoff of the market. In our model, also without relaying on Lucas (1978), the pricing
rule for insurance contracts (that also determines Ct ) and, Assumption 3 and 4, ensure that
the rate of “consumption” increases proportionally to the growth rate of the total payoff.
Moreover, as will discuss later, if assets’ payoffs were in term of perishable consumption
goods, it would not be possible to find a trading strategy that preserves the wealth (the
reserve capital in the insurance business) and have positive growth rate.
Let λit ∈ [0, 1] be the proportion of wealth invested in the risky asset by investor i ∈ It
at time t. We have
i λit wti
at = and mit = (1 − λit ) wti .
qt
We call the sequence (λit ){t>0 | i∈It } the trading strategy of investor i and λit the strategy of
investor i at time t. We use the convention that λit = 0 if i ∈ / It . Note that λit is a random
variable, i.e. it depends on the state of the world up to time t, st . Other assumptions on the
10
process defining the trading strategy (λit )t≥0 will be introduced later. Here, we just impose
the following restriction on the strategies at time t, (λit )i∈It , to prevent the price of the risky
asset from becoming zero.
Assumption 5 essentially states that if more than one investor survives period t, then there
exists at least one survivor with a strictly positive proportion of her wealth invested in
the risky asset and one survivor with a strictly positive proportion of her wealth invested
in the risk-free asset. Naturally, when a survivor has a mixed strategy4 λit ∈ (0, 1), then
Assumption 5 is obviously satisfied with i = j. If |It | = 1, then it might occur that the
unique survivor uses a strategy investing all her wealth in the risk-free asset. The strategy
λit = 1 is excluded by the solvency constraint. In fact, the solvency condition stated by
equation (5) is equivalent to
α α
1 − λit ≥ ⇔ λit ≤ 1 − =: λ ∈ (0, 1), (16)
R R
i.e., for each investor, the proportion of wealth invested in the risky asset is bounded from
above by λ. It seems to be a natural restriction for an insurance company (or a pension
fund), as shown e.g. in Davis (2001, Tables 5 and 6) for life insurances and pension funds
of several countries. Let λt = (λ1t , . . . , λIt )′ , then the market clearing condition for the risky
asset (14) implies
qt = λ′t wt .
P
Note that for i ∈ / It , wti = 0 by assumption and thus λ′t wt = i i
i∈It λt wt . We rewrite
equation (2) as follows
· ¸
λit
wt+1 = R (1 − λt ) + (dt+1 Wt + qt+1 ) − (ηt+1 − β + α) wti .
i i
(17)
qt
11
then, given that investors satisfy those conditions, we analyses the long-term wealth evolu-
tion. In our setting, analogously to Liu, Longstaff, and Pan (2003), we obtain upper bounds
for the λit ’s (a lower bound is given by the no short sale restriction). We will show below
that an investor with a strategy that does not prevent bankruptcy at each period, has a
strictly positive probability of vanishing from the market, even if she is the unique survivor.
Moreover, if an investor uses a simple strategy that does not prevent bankruptcy, she has
probability 1 of vanishing from the market, even if at some point in time she is the unique
survivor and thus dominates assets’ prices. In particular, an investor holding only the risky
asset (i.e. λit = 1) becomes extinct with probability 1. This result shows that Theorem 1 in
Hens and Schenk-Hoppé (2002b) does not hold when bankruptcy can occur.
We first consider the case |It | = 1 for some t > 0, i.e. It = {j} for some j ∈ {1, . . . , I}.
We restrict ourself to strategies λjt > 0. If λjt = 0, as is clearly excluded since R >
sup supp(η) − β + α5 ! The price of the risky asset at time t is given by qt = λjt wtj and
the aggregate wealth at time t is Wt = wtj : from equation (17) it follows immediately that
Let η = inf supp(η), η = sup supp(η) and d = sup supp(d) and K be the continuous
multivariate cumulative distribution of (η, d) on [η, η] × [0, d], i.e.
£ ¤
K(x, y) = P η ≤ x, d ≤ y .
£ ¤ R η R x+z
Moreover, let K̃(z) = P d − η ≤ z = η 0 dK(x, y) be the cumulative distribution
function of d − η. Then
£ ¤ £ ¤
P j ∈ It+1 = P dt+1 − ηt+1 > −R (1 − λjt ) − β + α
= 1 − K̃(−R (1 − λjt ) − β + α)
and thus
£ ¤ R+k+β−α
P j ∈ It+1 = 1 ⇔ λjt ≤ =: λ, (18)
R
where k = inf supp(K̃). We call this latter equation the no bankruptcy condition. Note that
β+k
λ=λ+ .
R
Therefore, the solvency constraint (16) is a stronger condition on the strategies than the
no bankruptcy condition (18), if β > −k, i.e. if higher shocks (greater than β) on the
5
supp(η) denotes the support of η.
12
insurance market and small dividends (less the η − β) in the financial market do not occur
simultaneously, which is not a realistic assumption. This is due to the fact that the solvency
constraint does not care about dividends, and thus does not take into consideration the
(positive) correlation between shocks and dividends, such that higher shocks will have a
smaller impact on the wealth evolution since they correspond to higher dividends. If β < −k
(which is the most common case, as for example when insurance shocks and dividends are
considered independent), the no bankruptcy condition (18) is stronger than the solvency
constraint and thus investors just care about the no bankruptcy condition (18). In this
case, the solvency constraint (16) does not eliminate bankruptcy! In the sequel we make the
following assumption on the joint distribution of (η, d):
Assumption 6 (Shocks and dividends joint distribution).
For all δ1 > 0 and δ2 > 0, £ ¤
P η > η − δ1 , d ≤ δ2 > 0,
i.e., big shocks and very small dividends have strictly positive probability to jointly occur.
Assumption 6 implies the following Lemma on the distribution of d − η.
Lemma 2. For all δ > 0, £ ¤
P d − η ≤ −η + δ > 0
and thus k = −η, i.e. maximal shocks and zero dividends have strictly positive probability to
jointly occur.
Proof.
£ ¤ £ ¤
P d − η ≤ −η + δ = P d ≤ η − η + δ
Z
£ ¤ £ ¤
= P d ≤ η − η + δ | η − η > −δ1 d P η − η > −δ1
Z0<δ1 <δ
£ ¤ £ ¤
≥ P d ≤ −δ1 + δ | η − η > −δ1 d P η − η > −δ1
0<δ1 <δ
Z £ ¤
P d ≤ −δ1 + δ, η > η − δ1 £ ¤
= £ ¤ d P η − η > −δ1
0<δ1 <δ P η > η − δ1
| {z }
>0
> 0
13
Let us now consider a single survivor j with a simple strategy λj > λ. Then£ at each period
¤
she will have a strictly positive probability of going bankrupt and therefore P j ∈ ∩t It = 0,
meaning that she will vanish almost surely from the market. We state these results in the
following Lemma.
Lemma 3. Let It = {j} for some t and j ∈ {1, . . . , I}, i.e. investor j is the unique survivor
at time t. The following holds:
(i) If λjt > λ, then investor j has strictly positive probability of going bankrupt during
period t + 1.
(ii) If λjs > λ for all s ≥ t, then investor j will almost surely eventually vanish from
the market. In particular, if investor j uses a simple strategy λj > λ, then she will
eventually almost surely vanish from the market almost surely.
Let us now consider the case |It | > 1. Without loss of generality we set It = {1, 2}:
if It = {i1 , . . . , in } with n = |It | > 2, then we can still reduce the original setting to a
2-investors setting by defining a “new investor” with strategy ξs ∈ [0, 1] at time s ∈ {t, t + 1}
and wealth ws , where Pn i i n
l=2 λs ws
l l X
ξs = P n il , w s = wsil .
l=2 ws l=2
The price of the risky asset at time s ∈ {t, t + 1} is then given by qs = λis1 wti1 + ξs ws . Thus
let us assume that It = {1, 2}. Then from the wealth evolution (17) it follows immediately
that for i = 1, 2
λit j λi
i ∈ It+1 ⇔ R (1 − λit ) + dt+1 Wt + wt+1 λjt+1 t − (ηt+1 − β + α) > 0, (19)
qt qt
where j 6= i.
i
Proof. (i) Suppose that i ∈ It+1 . Then wt+1 > 0 and by equation (17)
i λit wti
wt+1 = R (1 − λit ) wti + dt+1 Wt +
qt
¡ 1 ¢ λi w i
+ wt+1 λ1t+1 + wt+1
2
λ2t+1 t t − (ηt+1 − β + α) wti ,
qt
and thus
µ ¶
λiwti λit wti
i
wt+1 1− λit+1 t = R (1 − λit ) wti + dt+1 Wt
qt qt
j λit wti
+wt+1 λjt+1 − (ηt+1 − β + α) wti ,
qt
14
³ ´
λi wi
where j 6= i. Since λit+1 6= 1 (solvency restriction), then 1 − λit+1 tqt t > 0, and thus
i
from wt+1 > 0 it follows that
λit wti j λi w i
R (1 − λit ) wti + dt+1 Wt + wt+1 λjt+1 t t − (ηt+1 − β + α) wti > 0.
qt qt
Since i ∈ It , then wti > 0 and therefore dividing the last inequality by wti we obtain
λit j λi
R (1 − λit ) + dt+1 Wt + wt+1 λjt+1 t − (ηt+1 − β + α) > 0.
qt qt
λit j λi
R (1 − λit ) + dt+1 Wt + wt+1 λjt+1 t − (ηt+1 − β + α) > 0,
qt qt
where j 6= i. Then for i ∈ It ,
· ¸
i i λit j j λit
wt R (1 − λt ) + dt+1 Wt + wt+1 λt+1 − (ηt+1 − β + α) > 0,
qt qt
i
and thus wt+1 > 0, since
+
λit j λit
i
R (1 − λit ) + dt+1 Wt qt
+ wt+1 λjt+1 qt
− (ηt+1 − β + α)
wt+1 = wti λi wi
1− λit+1 tqt t
λit wti
and 1 − λit+1 qt
> 0 by equation (16).
The necessary and sufficient condition (19) for avoiding bankruptcy for investor i also
j λi
depends on other investors’ wealths and strategies, through the term wt+1 λjt+1 wti . Spec-
t
ulating on other investors’ behaviour, investor i could essentially put less wealth on the
risk-free asset than allowed under the no bankruptcy condition (18). While this would imply
a strictly positive probability of going bankrupt when investor i dominates assets’ prices,
the no bankruptcy condition is not necessary for avoiding almost surely bankruptcy in the
presence of competitors, when they significantly invest in the risky asset. However, the no
bankruptcy condition is the minimal condition on investment strategies that almost surely
eliminates bankruptcy in the presence of each type of competitor. In fact, an investor who
systematically violates the no bankruptcy condition (18), will eventually disappear from the
market with probability one, if her opponents are investing all their wealth on the risk-free
asset, i.e. an investment strategy that systematically violates the no bankruptcy condition
15
is almost surely driven out by the risk-free strategy. Thus the no bankruptcy condition
is the minimal condition that ensures that each investor will not go bankrupt with prob-
ability 1, regardless from other investors’ behaviour. In the sequel, because of the long
horizon perspective considered here, and following the approach of Liu, Longstaff, and Pan
(2003), we use the no bankruptcy condition to ensure that investors almost surely do not
face bankruptcy. In their setting, bankruptcy is penalized with minus infinity utility, so that
no optimal strategy will allow final negative wealth with strictly positive probability.
Following Hens and Schenk-Hoppé (2002b), we rewrite the wealth dynamics. We define
λit wti
Bti = ,
λ′t wt
and
By Assumption 7, we have
16
or
(I − Bt λ′t+1 ) wt+1 = At
where At = (A1t , . . . , AIt )′ , Bt = (Bt1 , . . . , BtI )′ and I is the identity on RI . Note that for
i∈/ It , Ait = Bti = 0. The inverse of I − Bt λ′t+1 is given by I + (1 − λ′t+1 Bt )−1 Bt λ′t+1 ,
provided that λ′t+1 Bt 6= 1 (see Horn and Johnson 1985, Sec. 0.7.4). It can be easily checked
that λ′t+1 Bt < 1 if there exits an investor i ∈ It+1 with λit < 1 and λit+1 > 0 and this is still
the case when |It | > 1, by Assumptions 5 and 7. If It = {j} for some j, then investor j is
already the unique survivor and the wealth evolution is easily obtained. Therefore, in the
sequel we only consider the case |It | > 1. Under the assumption of no default during period
t + 1, the wealth evolution can then be written as
¡ ¢−1
wt+1 = I − Bt λ′t+1 At
µ ′ ¶
Bt λt+1
= I+ At , (20)
1 − λ′t+1 Bt
and the i-th component is given by
i wti
wt+1 =P j j j×
j (1 − λt+1 ) λt wt
" Ã !#
£ ¤ X
× dt+1 Wt λit + R(λ − λit ) + (η − ηt+1 ) λ′t wt + (λit − λjt )λjt+1 wtj .
j6=i
i rti Wt
wt+1 =P j j j×
j (1 − λ ζt+1 ) ζt rt
" Ã !#
£ ¤ X
× dt+1 ζti + R λ (1 − ζti ) + (η − ηt+1 ) ζ ′t rt + λ (ζti − ζtj ) ζt+1
j
rtj .
j6=i
17
Let θt+1 be defined by
θt+1 = ζ ′t rt dt+1 +
à !
X³ ´ X j j j
′
+ R λ (1 − ζtk ) + (η − ηt+1 ) rtk ζ t rt + λ (ζtk − ζt ) ζt+1 rt .
k j
Then
θt+1
Wt+1 = P j j j Wt .
j (1 − λ ζt+1 ) ζt rt
θt+1
The ratio P j j
rtj
is the growth rate of the economy. From the wealth evolution of
j (1−λ ζt+1 ) ζt
equation (20), we obtain the evolution of wealth shares:
ri
i
rt+1 = t ×
θ
" t+1 Ã !#
£ ¤ X
× dt+1 ζti + R λ (1 − ζti ) + (η − ηt+1 ) ζ ′t rt + λ (ζti − ζtj )ζt+1
j
rtj .
j
(21)
i
From this last equation it follows directly that rt+1 = 0 if rti = 0 and therefore also, rt+1 i
=1
i
if rt = 1.
Without any additional assumption on the dividend process, the liquidity shock factor and
the investment strategies we are now able to prove that a trading strategy that corresponds
to the no bankruptcy boundary λ, is almost surely driven out by any strategy that is bounded
away from λ. From equation (21) it follows that for i, k ∈ It ,
i
µ i¶
rt+1 rt
k
= ×
rt+1 rtk
³ P ´
dt+1 ζti + [Rλ(1 − ζti ) + (η − ηt+1 )] ζ Tt rt + λ j6=i (ζti − ζtj )ζt+1
j
rtj
× £ ¤³ P ´.
dt+1 ζtk + Rλ(1 − ζtk ) + (η − ηt+1 ) ζ Tt rt + λ j6=k (ζtk − ζtj ) ζt+1 j
rtj
Let us now suppose that only two investors exist. The first investor is using a simple strategy
corresponding to the no bankruptcy boundary, i.e. λ1t = λ for all t. The second investor is
using a strategy which is bounded away from the no bankruptcy condition, as well as from
the strategy putting the wealth only on the risk-free asset, i.e. δ̃ < λ2t < λ − δ̃ for all t > 0
and for some δ̃ > 0. Using the notation introduced above, we have ζt1 = 1 for all t and
ζt2 ∈ (δ, 1 − δ) for all t and δ = λδ̃ > 0. We obtain the following result.
Theorem 1. Under Assumptions 3-7, and given an investor with ζt1 = 1 for all t > 0 and
an investor with ζt2 ∈ (δ, 1 − δ) for all t > 0 and some δ > 0, the investor with the simple
18
strategy corresponding to the no bankruptcy boundary, will almost surely vanish from the
market.
Proof. On {dt+1 − ηt+1 > −η} (by Assumptions 4 and Assumption 6, this set has probability
one) we have
µ 2¶ ¡ ¢
2
rt+1 rt dt+1 ζt2 + [R λ (1 − ζt2 ) + (η − ηt+1 )] ζ Tt rt − (1 − ζt2 ) λ rt1
1
= ¡ ¢
rt+1 rt1 dt+1 + (η − ηt+1 ) ζ Tt rt + (1 − ζt2 ) λ ζt+1
2
rt2
µ 1¶
rt
= ×
rt2
dt+1 ζt2 + [R λ(1 − ζt2 ) + (η − ηt+1 )] [1 − (1 − ζt2 ) λ − (1 − ζt2 )(1 − λ)rt2 ]
× £ ¤
dt+1 + (η − ηt+1 ) 1 − (1 − ζt2 )(1 − λζt+1 2
)rt2
µ 1¶ µ
rt dt+1 ζt2 + [1 − (1 − ζt2 ) λ] [R λ (1 − ζt2 ) + (η − ηt+1 )]
≥ min ,
rt2 dt+1 + (η − ηt+1 )
!
2
dt+1 + [R λ (1 − ζt ) + (η − ηt+1 )]
ζt2 £ ¤
dt+1 + (η − ηt+1 ) ζt2 (1 − λ ζt+1
2 2
) + λ ζt+1
µ 1¶ µ
rt dt+1 ζt2 + [1 − (1 − ζt2 ) λ] [R λ (1 − ζt2 ) + (η − ηt+1 )]
≥ min ,
rt2 dt+1 + (η − ηt+1 )
2
¶
2 dt+1 + [R λ (1 − ζt ) + (η − ηt+1 )]
ζt
dt+1 + η − ηt+1
µ 1¶
rt R λ (1 − δ) δ
≥ 2
> 0.
rt dt+1 + η − ηt+1
d ζ + [R λ (1 − ζ) + (η − η)] [1 − (1 − ζ) λ − (1 − ζ) (1 − λ) r]
r 7→ h i
d + (η − η) 1 − (1 − ζ) (1 − λ ζ̃) r
19
Let ǫ < R (1 − δ) δ λ, then
2 t+1
rt+1 X r2
log 1 ≥ C 1{dτ +η−ητ ≤ǫ} + log 01
rt+1 τ =1
r0
= C K̃(d − η ≤ −η + ǫ) = γ > 0,
rt2 rt2
by Assumption 6. Thus 1−rt2
= rt1
≈ exp(t γ) → ∞ as t → ∞, i.e rt2 → 1 almost surely.
The theorem states that, while being at the boundary of the no-bankruptcy condition
means that bankruptcy is excluded with probability one, the market selection mechanism still
forces such an investor to vanish from the market, if other investors are using strategies that
are bounded away from λ. Therefore, the trading strategy λit = λ cannot be evolutionary
stable as defined by Evstigineev, Hens, and Schenk-Hoppé (2003). In fact, even if this
strategy possesses almost the entire wealth, an investment strategy that is bounded away
from λ is able to drastically perturb the distribution of wealth shares and to drive out λit .
We next ask the question about investment strategies that are evolutionary stable, referring
to Hens and Schenk-Hoppé (2002a) and Evstigineev, Hens, and Schenk-Hoppé (2003). The
evolution of wealth shares from equation (21) can be written as follows. For i = 1, . . . , I let
"
i ³ ´
i rt i i
f (rt , t) = dt+1 ζt + Rλ(1 − ζt ) + (η − ηt+1 ) ×
θt+1
à !#
X
× ζ ′t rt + λ (ζti − ζtj ) ζt+1
j
rtj . (22)
j
Then
i
rt+1 = f i (rt , t)
or
rt+1 = f (rt , t), (23)
where f = (f 1 , . . . , f I )′ . Although it does not appear explicitly in the definition of fti , the
function ft also depends on the state of the world st+1 up to time t + 1, through investors’
strategies at time t + 1, the dividend dt+1 and the liquidity shock factor ηt+1 . We make the
following additional assumption on the trading strategies to make f independent from t and,
therefore, the market selection mechanism stationary, also because Assumptions 3 and 4 on
(ηt )t∈Z and (dt )t∈Z , respectively.
20
Assumption 8 (Stationary trading strategies).
The trading strategies are stationary, i.e. for all t ∈ Z and all i ∈ It
λit (st ) = λi (st ).
The market selection process (23) generates a random dynamical system (see Arnold
1998) on the simplex ∆I−1 . Given a vector of initial wealth shares r ∈ ∆I−1 and t > 0, the
map
φ(t, ω, r) = f (st , ·) ◦ f (st−1 , ·) ◦ · · · ◦ f (s1 , r), (24)
on N × Ω × ∆I−1 gives the investors’ wealth shares at time t, if the state of the world is
ω = (st )t∈Z , and φ(0, ω, r) = r. In the sequel we characterize vectors of wealth shares that
are invariant under φ. We introduce the following definition.
Definition 1 (Fixed point). The vector of relative wealth shares r ∈ ∆I−1 is called a
deterministic fixed point of φ, if and only if
φ(t, r, ·) = r
almost surely for all t. The distribution of market shares r is said to be invariant under the
market selection process (23).
Clearly, r is a deterministic fixed point of φ if and only if f (r) = φ(1, r, ·) = r almost
surely. Therefore, the vectors of wealth shares r = ei for i = 1, . . . , I are deterministic fixed
points of φ, where ½
i,j 1 if j = i
e = .
0 else
The following lemma shows that ei are the unique deterministic fixed points of φ. This result
also holds if Assumption 8 is not satisfied.
Lemma 4. Let r be a deterministic fixed point of φ. Then r = ei for some i = 1, . . . , I.
Proof. Let assume that ri = rt+1
i
= rti ∈ (0, 1). Then
³ ´³ X ´
θt+1 = dt+1 ζti + R λ (1 − ζti ) + (η − ηt+1 ) ζ ′t rt + λ (ζti − ζtj ) ζt+1
j
rtj ,
j
or equivalently
X
dt+1 ζtk rtk +
k6=i
à !
X³ ´ X
+ R λ (1 − ζtk ) + (η − ηt+1 ) rtk ζ ′t rt + λ (ζtk − ζtj ) ζt+1
j
rtj
k6=i j
= dt+1 ζti
(1 − + rti )
³ ´³ X ´
i i ′ i j j j
+(1 − rt ) R λ (1 − ζt ) + (η − ηt+1 ) ζ t rt + λ (ζt − ζt ) ζt+1 rt .
j
(25)
21
P
Since 1 − rti = k
k6=i rt , the right-hand side of equation (25) corresponds to
X X³ ´ ³ X j j j
´
′
dt+1 ζti rtk + i k
R λ (1 − ζt ) + (η − ηt+1 ) rt ζ t rt + λ i
(ζt − ζt ) ζt+1 rt
k6=i k6=i j
P k k
k6=i ζt rt
P
Let us first suppose that k (ζti − ζtk ) rtk = 0. Then ζi = ξt , where ξt = 1−rt i . Moreover,
the last equation is equivalent to
X
(ζti − ζtk ) ζtk rtk = 0
k
and thus X
ξt2 = (ζtk )2 rtk .
k6=i
This last equation implies ζtk = 0 for all k, or rtk = 0 for k 6= i. In the first case we have a
contradiction to Assumption
P 5. i In the second case we have a contradiction to rti ∈ P (0, 1).
Let us now suppose that k (ζt − ζt ) rt 6= 0. Without loss of generality, we take k (ζti −
k k
22
P
ζtk ) rtk > 0 (the same argument can also be used for the case k (ζti − ζtk ) rtk < 0) . Then
X j j X j
0 = dt+1 + λ (η − ηt+1 ) ζt+1 rt − R λ (1 − λ ζt+1 ) ζtj rtj
j j
à !
R λ2 X
i 2 k 2
X
i l l k
X j j
−P i l l
(ζt ) − (ζt ) − (ζt − ζt ) r t r t ζt+1 rt
l (ζt − ζt ) rt k l j
X j j X j j j
= dt+1 + λ (η − ηt+1 ) ζt+1 rt − R λ (1 − λ ζt+1 ) ζt rt
j j
2
Rλ X X j j
−P i l l
ζtk (1 − ζtk ) rtk ζt+1 rt .
l (ζt − ζt ) rt k6=i j
nP o
i j j
Since rt+1 = rti ∈ (0, 1), the set ζ r
j t+1 t = 0 has probability zero by Assumption 5.
P j j
Thus j ζt+1 rt > 0 almost surely. Let δ > 0, then by Assumptions 4 and 6, the set
{st+1 | dt+1 (st+1 ) = 0, η − ηt+1 (st+1 ) < δ} has strictly positive probability independently
from st . Thus
X λ X
(1 − λ) ζtj rtj + P i l l
ζtk (1 − ζtk ) rtk < δ.
j l (ζt − ζt ) rt k6=i
Since this is true for all δ > 0, ζtj = 0 for all investors with strictly positive wealth share at
time t, a contradiction to Assumption 5 or rti = rt+1i
∈ (0, 1). Therefore, rti = 0 or rti = 1.
The Lemma implies that we can restrict ourselves to monomorphic populations of investors
(all investors with a strictly positive market share possess the same trading strategy), to
analyse invariant wealth share distributions. In particular, we are looking at deterministic
fixed points that are stable, such that a small perturbation of the vector of wealth shares does
not change the long-run evolution. Since invariant wealth share distributions correspond to
monomorphic populations, the stability of investment strategies is related to the stability of
the associated fixed point. Therefore, we consider a population of trading strategies with
an incumbent strategy λi (with market share rti ) and a distinct mutant strategy λj (with
market share rtj = 1 − rti ).
23
Let us consider a population of only two investors, where the incumbent is investor 1 with
market share rt1 and the mutant is investor 2 with market share 1 − rt1 . The wealth share
dynamic for investor 1 is obtained from (21) by
∂ψ(rt1 ) ¯¯
¯
∂rt1 rt1 =1
where
g̃ζ 1 (ζ 2 (s0 ), s0 ) =
Z n £ ¤−1 h 2 0
= µ(ds) log ζ 1 (s0 )−1 d(s) + R λ (1 − ζ 1 (s0 )) + η − η(s) ζ (s ) d(s)+
S
³ ´³ ´io
+ R λ (1 − ζ 2 (s0 )) + (η − η(s)) ζ 1 (s0 ) + (ζ 2 (s0 ) − ζ 1 (s0 )) λ ζ 1 (s0 , s) .
The function ζ 2 (s0 ) 7→ g̃ζ 1 (ζ 2 (s0 ), s0 ) is continuous, strictly concave on [0, 1] for all s0 and,
obviously g̃ζ 1 (ζ 1 (s0 ), s0 ) = 0 for all s0 . Moreover,
∂g̃ζ 1 (ζ 2 (s0 ), s0 ) ¯¯ 1
2 0
¯2 0 1 0 = 1 0 ×
∂ζ (s ) ζ (s )=ζ (s ) ζ (s )
Z
d(s) − Rλζ (s ) + Rλ2 ζ 1 (s0 , s)(1 − ζ 1 (s0 )) + (η − η(s))λζ 1 (s0 , s)
1 0
× µ(ds) ³ ´ .
S d(s) + Rλ(1 − ζ 1 (s0 )) + (η − η(s))
24
Theorem 2. Let the state of nature (st )t∈Z be determined by an i.i.d. process. Suppose that
all investors use simple strategies λi (ω) ≡ λi ∈ (0, 1).
(i) If
£ d ¤ £ η−η ¤ £ 1 ¤
E + λE ≥ RλE ,
d+η−η d+η−η d+η−η
then there is no evolutionary stable investment strategy.
(ii) If
£ d ¤ £ η−η ¤ £ 1 ¤
E + λE < RλE ,
d+η−η d+η−η d+η−η
then there is a unique evolutionary stable investment strategy λ⋆ = λ ζ ⋆ where ζ ⋆ sat-
isfies: £ ¤
Z
d(s) − R λ ζ + R λ2 ζ (1 − ζ) + (η − η(s)) λ ζ
µ(ds) = 0. (26)
S d(s) + R λ (1 − ζ) + η − η(s)
Proof. An investment strategy ζ ⋆ is evolutionary stable if gζ ⋆ (ζ) < 0 for all ζ ∈ [0, 1]. From
Theorem 1, we have ζ ⋆ 6= 1. Moreover, if investors strategies are simple, we have
gζ 1 (ζ 2 ) = g̃ζ 1 (ζ 2 ).
Let £ ¤
d − R λ ζ + R λ2 ζ (1 − ζ) + (η − η) λ ζ
h(ζ; d, η) = .
d + R λ (1 − ζ) + η − η
Then ∂h(ζ;d,η)
∂ζ
< 0 for ζ ∈ [0, 1], thus h is strictly decreasing on [0, 1] for all d and η and,
h(0; d, η) = d+R λd(η−η) ≥ 0 for all d and η, and strictly positive for d > 0. Thus
Z
µ(ds) h(0; d(s), η(s)) > 0
S
25
since by Assumption 4 the set {s ∈ S | d(s) > 0} has strictly positive probability. Moreover,
h(1; d, η) = d−Rd+η−η
λ+(η−η) λ
and
Z Z
d(s) − R λ + (η − η(s)) λ
µ(ds) h(1; d(s), η(s)) = µ(ds)
S S d(s) + η − η(s)
Z Z
d(s) η − η(s)
= µ(ds) +λ µ(ds)
S d(s) + η − η(s) S d(s) + η − η(s)
Z
µ(ds)
−R λ
S d(s) + η − η(s)
£ d ¤ £ η−η ¤ £ 1 ¤
= E + λE − RλE .
d+η−η d+η−η d+η−η
Therefore Z
µ(ds) h(1; d(s), η(s)) < 0
S
if and only if
£ d ¤ £ η−η ¤ £ 1 ¤
E + λE − RλE < 0.
d+η−η d+η−η d+η−η
£ d ¤ £ η−η ¤ £ 1 ¤
Thus, if the condition E d+η−η + λ E d+η−η − R λ E d+η−η ≥ 0 holds, then the function
R
S
µ(ds)h(ζ; d(s), η(s)) is strictly positive on [0, 1) and therefore no evolutionary stable strat-
egy exists. Note that since gζ 1 is continuous and strictly concave, if no evolutionary stable
strategy exists, then no local evolutionary stable strategy can exist either. In fact, in a small
neighborhood of ζ1 there exists an investment strategy ζ 2 > ζ 1 with gζ 1 (ζ 2 ) > 0. This proves
(i). £ ¤ £ η−η ¤ £ 1 ¤
d
If E d+η−η + λ E d+η−η − R λ E d+η−η < 0, then there exists exactly one ζ ⋆ ∈ (0, 1) such
R
that S µ(ds)h(ζ ⋆ ; d(s), η(s)) = 0 and therefore gζ ⋆ (ζ) < 0 for ζ ∈ (0, 1), ζ 6= ζ ⋆ . Obviously
ζ ⋆ solves Z
µ(ds) h(ζ; d(s), η(s)) = 0,
S
26
this kind of strategy only delivers growth as long as it doesn’t lead to bankruptcy, which
will almost surely occur. Thus, if the growth rate at ζ = 1 is positive, no evolutionary stable
strategy can exist.
Example £
Let
¤ us suppose that d and η are independent, d is distributed on [0, 0.1] with H(0) = P d=
0 = 0.01, (d | d > 0) ∼ Beta(2, 2) and, η ∼ Uniform(0.05, 0.1). Moreover, we take α = 0.04
and R = 1.025. Then λ = 0.93796 and ζ ⋆ = 0.213528. Thus, the unique evolutionary stable
investment strategy consists in investing only the 20.03% of the wealth on the risky asset.
The evolution of market shares is shown Figure 1.
⋆ ⋆
If (d | d > 0) ∼ Beta(2, 3), then
£ ¤ the unique evolutionary stable strategy is λ = 17.66%; ζ
decreases with decreasing E d .
5 Conclusion
In this paper we have proposed an evolutionary portfolio model with bankruptcy. The in-
vestors are insurances companies, thus with insurance market exposures. The amount of
wealth that is withdrawn or collected at each period, corresponds to the difference between
indemnities, which must be paid out and premia paid in. If this difference is negative, in-
surance companies face liquidity shocks that could force investors to withdraw their entire
wealth, thus forcing the company into bankruptcy. We introduce the no bankruptcy condi-
tion on investment strategies that ensures that bankruptcy is excluded with probability one
and we analyse the set of simple strategies that are evolutionary stable if the state of nature
is generated by an i.i.d. process. In fact, if an investment strategy does not almost surely
eliminate bankruptcy, the company will almost surely disappear from the market eventually.
We prove that invariant wealth shares distribution only corresponds to monomorphic popu-
lations. Moreover, we show that, depending on the dividend and liquidity shock processes,
there exists a unique evolutionary stable strategy or, for all strategies there exists at least
one mutant that is able to drastically perturb the distribution of wealth shares. We give the
condition that characterize existence and, if this condition is satisfied we also characterize
the unique evolutionary stable strategy.
This work shows that when savings and withdrawals are not a positive percentage of the
investor’s wealth, then one should also take bankruptcy into account. Thus, an additional
dimension has to be considered as compared to the case where bankruptcy is excluded al-
most surely (Blume and Easley 2002, Evstigineev, Hens, and Schenk-Hoppé 2003). Here we
take a long-horizon perspective, so that a trading strategy with strictly positive probability
of going bankrupt at each period cannot be evolutionary fit, since it will disappear from the
market almost surely. This paper also suggest that near the growth rate, one should also
take the risk dimension into account, which is the probability of going bankrupt. Thus we
introduce the risk dimension in the Evolutionary Portfolio Theory.
27
1.0
0.8
0.6
Wealth shares
(a)
0.4
0.2
0.0
0 20 40 60 80 100
Time
1.0
0.8
0.6
Wealth shares
(b)
0.4
0.2
0.0
0 20 40 60 80 100
Time
Figure 1: Evolution of market shares, for the λ⋆ strategy of Theorem 2 (full line), the strategy
λ (dotted line), the risk free strategy (dashed-dotted line), and a randomly chosen strategy
in (0, λ) (dashed line). In figure (a) all investors have the same initial wealth, while in figure
(b) the strategy λ⋆ initially possesses only the 2% of the market capital.
28
References
Arnold, L. (1998): Random Dynamical Systems. Springer Verlag, Berlin.
Artzner, P., F. Delbaen, J.-M. Eber, and D. Heath (1997): “Thinking Coherently,”
Risk, 10(11), 68–71.
Blume, L., and D. Easley (1992): “Evolution and Market Behavior,” Journal of Eco-
nomic Theory, 58(1), 9–40.
(2002): “If You’re So Smart, Why Aren‘t You Rich? Belief Selection in Complete
and Incomplete Markets,” Working paper, Department of Economics, Cornell University.
Breiman, L. (1961): “Optimal Gambling Systems For Favorable Games,” in Fourth Berke-
ley Symposium on Mathematical Statistic and Probability, vol. 1, pp. 65–78, Berkeley.
University of California Press.
Browne, S. (1997): “Survival and Growth with a Liability: Optimal Portfolio Strategies
in Continuous Time,” Mathematics of Operation Research, 22(2), 468–493.
Carino, D., D. Myers, and W. Ziemba (1998): “Concepts, Technical Issues, and Uses
of the Russel-Yasuda Kasai Financial Planning Model,” Operations Research, 46, 449–462.
Carino, D., and W. Ziemba (1998): “Formulation of the Russel-Yasuda Kasi Financial
Planning Model,” Operations Research, 46, 433–449.
Davis, P. (2001): “Portfolio Regulation of Life Insurance Companies and Pension Funds,”
Working paper No. PI-0101, The Pensions Institute, Birkbeck College.
29
Evstigneev, I., T. Hens, and K. Schenk-Hoppé (2002): “Market Selection of Financial
Trading Strategies: Global Stability,” Mathematical Finance, 12(4), 329–339.
Hens, T., and K. Schenk-Hoppé (2002b): “Markets Do Not Select For a Liquidity
Preference as Behaviour Towards Risk,” Working paper No. 139, Institute for Empirical
Research in Economics, University of Zurich.
Horn, R., and C. Johnson (1985): Matrix Analysis. Cambridge University Press, Cam-
bridge UK.
Liu, J., F. A. Longstaff, and J. Pan (2003): “Dynamic Asset Allocation With Event
Risk,” Journal of Finance, 58(1), 231–259.
Norberg, R., and B. Sundt (1985): “Draft on a System for Solvency Control in Non-Life
Insurance,” ASTIN Bulletin, 15(2), 150–169.
Pestien, V., and W. Sudderth (1985): “Continuous-Time Red and Black: How to
Control a Diffusion to a Goal,” Mathematics of Operations Research, 10(4), 599–611.
30
Sandroni, A. (2000): “Do Markets Favor Agents Able to Make Accurate Predictions?,”
Econometrica, 68(6), 1303–1341.
Schnieper, R. (1993): “The Insurance of Catastrophe Risks,” SCOR Notes, April 1993.
Thorp, E. (1971): “Portfolio Choice and the Kelly Criterion,” in Stochastic Models in
Finance, ed. by W. Ziemba, and R. Vickson, pp. 599–619. Academic Press, New York.
Zhao, Y., and W. Ziemba (2000): “A Dynamic Asset Allocation Model with Downside
Risk Control,” The Journal of Risk, 3(1), 91–113.
Ziemba, W. (2002): “The Capital Growth Theory of Investment: Part I,” Willmott Maga-
zine, December, 16–18.
6 Appendix
6.1 Existence and uniqueness of δti and Pt+1
For the sake of simplicity we drop the index t from the notation of equations (7) and (8).
Using the expression (7) for δ i in (8), we obtain
Xµ σ θi
¶
P =µ+ i−P
− 1 αi w i ,
i
µ + σ θ
since µ > 0. Thus f (P ) ≥ 0 and it possesses exactly two fixed points: P = 0 and P ⋆ ∈
(0, mini {µ + σ θi }). Therefore, there is a unique premium P ∗ > 0 which satisfies equations
(7) and (8). Moreover, by equation (7), δ i is also uniquely defined for all i.
31
6 ´
´
b ´
´
´
´
´
´
´
´
´
´
´
´
´
´
´ f
´
´ -
∗
0 P b
Figure 2: Proof of the existence and uniqueness of equilibrium insurance premium. The
demand function f (P ) has one strictly positive fixed point P ∗ .
Bi X j
i
wt+1 = Ait+1 + P tj j λt+1 Ajt
1 − j λt+1 Bt j
λi w i X j
= Ait+1 + P j j tPt j j j λt+1 Ajt
j λ t wt − j λt+1 λt wt j
· i
λ
= wti R (1 − λit ) + T t dt+1 Wt − (ηt+1 − β + α)
λt wt
P j j ³ j λjt
´
w λ
j t t+1 R (1 − λ t ) + T
λt wt
d t+1 W t − (η t+1 − β + α)
+λit P j j j
j (1 − λ t+1 ) λ w
t t
32
Thus
wti £
i
wt+1 = T
R (1 − λit ) λTt wt + λit dt+1 Wt − (ηt+1 − β + α)λTt wt
λt wt
P j j ¡ j T j T
¢#
i j wt λt+1 R (1 − λt )λt wt + λt dt+1 Wt − (ηt+1 − β + α)λt wt
+λt P j j j
j (1 − λt+1 )λt wt
wti 1
= P ×
λt wt j (1 − λjt+1 )λjt wtj
T
"
X ¡ ¢
(1 − λjt+1 )λjt wtj R(1 − λit )λTt wt + λit dt+1 Wt − (ηt+1 − β + α)λTt wt
j
à !#
X ¡ ¢
+λit wtj λjt+1 R(1 − λjt )λTt wt + λjt dt+1 Wt − (ηt+1 − β + α)λTt wt
j
wti
= P ×
j (1 − λjt+1 )λjt wtj
" Ã !#
£ ¤ X
dt+1 Wt λit + R(1 − λit ) − (ηt+1 − β + α) λTt wt + (λit − λjt )λjt+1 wtj
j
and therefore
wti
wt+1 = P j j j×
j (1 − λt+1 )λt wt
" Ã !#
£ ¤ X
dt+1 Wt λit + R(λ − λit ) + (η − ηt+1 ) λTt wt + (λit − λjt )λjt+1 wtj .
j6=i
We use that
R−η+β−α
λ= ⇐⇒ R + β − α = R λ + η.
R
33