Bounded brownian motion

Carr, Peter Article Bounded brownian motion Risks Provided in Cooperation with: MDPI – Multidisciplinary Digital Publishing Institute, Basel Suggested Citation: Carr, Peter (2017) : Bounded brownian motion, Risks, ISSN 2227-9091, MDPI, Basel, Vol. 5, Iss. 4, pp. 1-24, https://doi.org/10.3390/risks5040061 This Version is available at: https://hdl.handle.net/10419/195790 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by/4.0/ risks Article Bounded Brownian Motion Peter Carr Department of Finance and Risk Engineering, Tandon School of Engineering, NYU, 12 MetroTech Center, Brooklyn, NY 11201, USA; [email protected].com Academic Editor: Albert Cohen Received: 12 February 2017; Accepted: 19 October 2017; Published: 17 November 2017 Abstract: Diffusions are widely used in finance due to their tractability. Driftless diffusions are needed to describe ratios of asset prices under a martingale measure. We provide a simple example of a tractable driftless diffusion which also has a bounded state space. Keywords: standard Brownian motion; Brownian martingale; diffusion coefficient 1. Introduction Standard Brownian motion (SBM) is the most widely studied stochastic process because it serves as a highly tractable model of both a martingale and a Markov process. In finance, the martingale property describes asset prices relative to some numeraire under the assumption of no arbitrage. The Markov property can also describe some asset prices when markets for them are thought to be semi-strong form efficient. However, for limited liability assets, such as stocks, it is well known that SBM cannot describe their prices because prices are non-negative, while the state space of an SBM is the whole real line. The observation lead Samuelson and Osborne to propose that arbitrage-free asset prices, relative to a numeraire, be modeled as a geometric Brownian martingale (GBM). As is well known, this Markovian martingale is obtained from standard Brownian motion by exponentiating. The convexity of the exponential introduces a positive drift, so one can restore martingality by introducing time decay. One can alternatively create a GBM by starting from a linear Brownian motion with constant drift and then evaluating the scale function of the drifting Brownian motion on the drifting Brownian motion. The resulting GBM has state space ( 0, ∞) , making it suitable to describe arbitrage-free prices of a limited liability asset relative to a numeraire. While the use of a GBM addresses the lower bound constraint imposed by limited liability, it does not address the upper bound. SBM and GBM can both achieve arbitrarily high positive values. This renders them unsuitable to describe the prices of assets with a finite number of payouts, each of which is bounded above. Examples would include coupon bonds and derivative securities with a finite number of bounded payouts (e.g., a binary option). Exchange rates describe the price of one currency in terms of another and are often legally manipulated by governments to lie between two positive bounds. One may also wish to describe a financial concept other than a price with a stochastic process. For example, stochastic processes are used to describe interest rates, variance rates, and hazard rates. Historical data on these observables is typically confined to a band, which often leads to the imposition of mean reversion in the dynamics. However, it would not be unreasonable to use the historical band as a guide to setting a future band that these rates cannot escape. Many concepts in probability have evolved into financial concepts. For example, a probability can be used to describe the price of a binary option Taleb (2017), while a correlation is used to describe the swap rate in a correlation swap Jacquier and Slaoui (2010). Since both of these probability concepts are bounded, it becomes natural to consider bounded processes to describe all of these concepts. In this paper, we construct a stochastic process Swith the following properties: Risks 2017,5, 61; doi:10.3390/risks5040061 www.mdpi.com/journal/risks Risks 2017,5, 61 2 of 11 1. The process S is obtained by evaluating the scale function of a mean-repelling Ornstein Uhlenbeck (OU) process. As a result, S is a time-homogeneous driftless diffusion. Furthermore, since the scale function can be evaluated in closed form, the transition probabilities of S are known in closed form. 2. The state space of the process S is ( 0, H) , for some H> 0. In words, the process S is bounded below and above and its natural boundaries are zero and a positive constant H . Since S has zero drift and can’t explode, it is a martingale. 3. The diffusion coefficient of Sis positive and bounded above by a positive constant h. 4. Aside from its starting value S0∈( 0, H) , the process S has two free parameters H and h which respectively describe the maximum value and maximum (normal) volatility of S. Analogous with the term geometric Brownian motion, we christen this process “bounded Brownian motion”. 2. Applications 2.1. Managed Currency Consider a forward exchange rate in a setting when a monetary authority is able to manage the money supply or interest rates such that a given exchange rate stays between two positive barriers over the life of the forward contract. By adding a positive constant to the bounded Brownian motion S , we synthesize the risk-neutral dynamics of the forward exchange rate process. Papers modeling the spot FX process between bands include Carr and Kakuschadze (2017), Hui et al. (2008), Ingersoll (1997) , and Rady (1997). 2.2. Correlation Swap In its simplest form, a correlation swap designates two underlying assets and a fixed maturity date. At maturity, the dollar payoff is affine in the realized correlation of returns between two underlying assets. The slope of this affine relation is the notional of the correlation swap which is determined at inception. The ratio of the intercept to this notional is the correlation swap rate for maturity T , which is also determined at inception. As we move through calendar time, the conditional expected value of the floating leg of a seasoned correlation swap is a martingale fluctuating in the interval [− 1, 1 ] . By multiplying Sby 2 Hand subtracting one, we obtain such a process. 2.3. Protection Leg of CDO The protection leg of a CDO has a nonnegative value that fluctuates in the interval [ 0, H] . The value before the maturity date T is in ( 0, H] , while the value at T is in [ 0, H] . The dynamics presented here fluctuate in the interval (0, H). However, we will show a way to change the domain to (0, H]. 3. Genesis Fix a probability space (Ω , F , Q) . We will refer to Q as the risk-neutral measure. Let X0∈R and for positive constants Hand h, consider the following1mean-repelling OU process: dXt=2πh2 H2Xtdt +hdWt,t≥0, (1) where Wis a Qstandard Brownian motion. 1 The scaling factor 2 π in the drift of X is entirely optional, but is inserted here so that the volatility of the bounded Brownian motion can later be expressed in terms of an un-normalized Gaussian function, rather than a normalized one. Risks 2017,5, 61 3 of 11 Let S(x):R7→ R be a C2 function which solves the linear second order ordinary differential equation (ODE): h2 2S00(x) + 2πh2 H2xS0(x) = 0, x∈R. (2) As this is a linear first order ODE in S0(x), the general solution is given by: S(x) = Zx aexp −Zy b 2π H2zdzdy, (3) where the lower integral limits a and b are arbitrary. Since S is increasing in x , it is generally referred to as a scale function of the diffusion process X. Let SH(x)denote the particular increasing function obtained when a=−∞and b=0: SH(x)≡Zx −∞exp −Zy 0 2π H2zdzdy. (4) Some easy calculus gives: SH(x) = Zx −∞exp −1 2 2πy2 H2dy =HZ√2πx H −∞ exp[−z2 2] √2πdz =HN √2πx H!,t≥0, (5) where N(d):R7→ ( 0, 1 ) is the standard normal cumulative distribution function (CDF). Since N maps Rto (0, 1), (5) implies that SHmaps Rto (0, H). Let {St ; t≥ 0 } be the stochastic process obtained by evaluating the scale function SH on the OU process X: St≡SH(Xt) = HN √2πXt H!, (6) from (5). It is well known that S is a time-homogeneous driftless diffusion. Since the range of the function SH(·) is bounded, it is clear from (6) that the process S takes values on ( 0, H) . Since S is bounded, it is clearly a martingale and not just a local martingale. 4. Valuing Perpetual Claims Without Knowing Volatility Consider perpetual claims written on the path of a single underlying asset. In this section, we show that we can value several such claims without knowing the volatility of the underlying asset. The price of the underlying asset will be a continuous martingale, but it need not be the particular continuous martingale S introduced in the last section. To distinguish general results applying to continuous martingales from particular results applying to the driftless bounded diffusion S defined by (6), we will denote the former process by M . We being with some observations about hitting probabilities of Sand then generalize to M. Let `and rbe two real-valued constants satisfying: −∞<`<X0<r<∞. (7) Risks 2017,5, 61 4 of 11 In words, ` and r are finite and bracket X0 . Let σ` and σr respectively denote the first passage times of Xto `and r. Let S0∈(0, H)be the initial value of the Sprocess: S0≡SH(X0) = HN √2πX0 H!, (8) from (6). Let Dand Ube defined by: D≡SH(`)U≡SH(r). (9) Since SH(·)is increasing, (7) implies that Dand Usatisfy: 0<D<S0<U<H. (10) In words, D and U are both in ( 0, H) and bracket S0 . Let τu and τd respectively denote the first passage times of S to U and D . Since S is a process in natural scale, the probability that S hits U before Dstarting from S0is given by: Q{τu<τd|S0=S}=S−D U−D,S∈(D,U). (11) In fact, (11) is a standard result that holds for any continuous martingale with unbounded quadratic variation. We will soon discuss results for such processes, but, for now, we confine our explorations to the driftless bounded diffusion Sdefined by (6). Consider a claim that pays one dollar if the process S hits U before D and zero otherwise. If interest rates are zero, then (11) gives the value of such a claim as a function of the starting point S of the underlying. Notice that given a direct observation of the starting value S , this value is invariant to the volatility of the process S . Since σ`=τd and σr=τu , we can make the corresponding statements for hitting probabilities and claim values when Xis the underlying. In fact, the affine form of the hitting probabilities of S generalize to any continuous martingale M , whose quadratic variation becomes infinite as the horizon become infinite: lim T↑∞hMiT=∞. (12) This condition is needed to rule out continuous martingales which absorb at spatial boundaries placed between the starting value and the barrier of interest. Let (−∞ , ∞) be the state space of the continuous martingale. Suppose again that Dand Uare both in (−∞,∞)and bracket M0: −∞<D<M0<U<∞. (13) Consider the random payoff from a “perpetual” claim on M that pays Rd at time τd if M hits D first and Ru at time τu if M hits U first. Assuming zero interest rates and no dividends from the underlying asset before the first exit, the initial value of this claim is: V0=M0−D U−DRu+U−M0 U−DRd. (14) By setting Rd=0 and Ru=1, one obtains the following generalization of (11): Q{τu<τd|F0}=M0−D U−D, (15) Risks 2017,5, 61 5 of 11 where we assume that the initial filtration F0 contains M0 . By setting Rd=D and Ru=U , one can interpret Rd and Ru as liquidating dividends from the asset whose price is M . In this case, no assumption on interest rates is needed since the replicating strategy just holds one share. For the above claims, monitoring of the barriers begins at time t= 0. Consider instead a more general claim where monitoring of the exit barriers starts from some fixed time T≥ 0. We claim that static positions in European options maturing at T can be used to span this payoff. In particular, consider a butterfly spread with strikes D , K , and U where −∞<D<K<U<∞ . Suppose that the positions in European options are chosen so that the butterfly spread pays off one dollar if ST=K at T . Then we claim the value of this butterfly spread on M is the joint risk-neutral probability that M hits K after Tbefore it hits Dor Uafter T. Butterfly spreads can be synthesized using puts or calls. We will use puts and hence let P0(K , T) denote the initial price of a European put of strike K∈R and maturity T≥ 0. Consider the following butterfly spread payoff: BST=(U−MT)+−(K−MT)+ U−K−(K−MT)+−(D−MT)+ K−D. (16) The initial cost of forming this butterfly spread is: BS0=P0(U,T)−P0(K,T) U−K−P0(K,T)−P0(D,T) K−D. (17) Since the payoff in (16) is bounded between 0 and 1, no arbitrage forces the value in (17) to also be bounded between 0 and 1. Let τT B be the first time after T that the continuous martingale M touches a barrier B . If the martingale never touches B after T , then τT B=∞ . Let τT DU ≡τT D∧τT U be the first time after T that the martingale touches either Dor U. If the martingale never touches Dor Uafter T, then τT DU =∞. Theorem 1. No arbitrage and zero interest rates implies: BS0=Q{τT K<τT DU}. (18) Proof. We need to show that BS0 is the initial cost of a strategy that pays one dollar if τT K<τT DU and which pays zero otherwise. Consider the following trading strategy. At time 0, the investor buys the butterfly spread by: 1. buying 1 U−Kputs struck at U 2. selling 1 U−K+1 K−Dputs struck at K 3. buying 1 K−Dputs struck at L. The net cost is given in (17). The put portfolio is held static to T . If ST<D or ST>U , then the portfolio expires worthless. This matches the payoff of the desired claim since if ST<D , then we must have τT K≥τT D=τT DU and similarly, if ST>U, then we must have τT K≥τT U=τT DU. If ST∈(D,U), then use the payoff from the portfolio to finance the following positions: 1. if ST∈(D,K)buy 1 K−Dshares and borrow D K−Ddollars. 2. if ST∈(K,U)short 1 U−Kshares and lend U U−Kdollars. If τT K>τT DU , then at the first exit time τT DU of the corridor (D , U) , liquidate the stock bond portfolio for zero. Otherwise, if τT K<τT DU , then at the hitting time τT K , liquidate the stock bond portfolio for one dollar. Since the quadratic variation of M grows without bound, the risk-neutral probability that τT K and τT DU are both infinite is zero. This concludes the proof. Q.E.D. Risks 2017,5, 61 6 of 11 There is a second kind of butterfly spread with a probabilistic interpretation. We first consider the simpler spot starting case. Consider a perpetual claim written on the continuous martingale M satisfying both (12) and (13). Suppose again that the claim pays off at τDU . For K∈(D , U) , suppose that the payoff at time τDU is the Local Time of M at time τDU . Loosely speaking, the payoff at τDU accumulates over time twice the instantaneous variance experienced by the process M at K until τDU : LM τDU (K)≡2ZτDU 0 δ(Mt−K)dhMit. (19) From the Tanaka Meyer formula, the cost of creating this payoff is: ˆ G(M,K) = (2(M−D)(U−K) U−D,−∞<D<M<K<U<∞ 2(K−D)(U−M) U−D,−∞<D<K<M<U<∞.(20) This can be written much more succinctly as: ˆ G(M,K) = 2[(M∧K)−D]+[U−(K∨M)]+ U−D. (21) when graphed against either M or K , the function ˆ G is flat at zero outside (D , U) and triangular in between. When graphed against M , the kinks are at D , K , and U and the change in slope of ˆ G at K is two. Again, it is remarkable that the claim can be valued without knowledge of the volatility of M. To value the forward-start version of the above claim, i.e., the claim paying: 2ZτT DU T δ(Mt−K)dhMit(22) at τT DU, consider the following butterfly spread payoff: BST=2[(MT∧K)−D]+[U−(K∨MT)]+ U−D. (23) This payoff can be synthesized by: 1. buying 2 U−K U−Dputs struck at D 2. selling two puts struck at K 3. buying 2 K−D U−Dputs struck at U. The initial cost of forming this butterfly spread is: BS0=2U−K K−DP0(D,T)−2P0(K,T) + 2K−D U−KP0(U,T). (24) Suppose that this butterfly spread is held static to maturity. If MT<D or MT>U , then the butterfly spread expires worthless. This matches the payoff since we already know at T that local time at Kcannot increase from zero without Mfirst hitting Dor U. If MT∈(D , U) , then the payoff at T finances the initial position in the following trading strategy in stocks and bonds conducted over the period (T,τT DU). 1. if MT∈(D,K], be long 2 U−K U−Dshares and borrow 2 U−K U−DDdollars 2. if MT∈(K,U), be short 2 K−D U−Dshares and lend 2 K−D U−DUdollars. This strategy is self-financing except when the stock price is near the intermediate strike K . Using the Tanaka Meyer formula, one can show that this strategy generates the increase in the local time from Tto τT DU. Risks 2017,5, 61 7 of 11 5. Diffusion Coefficient So far we have been able to value various perpetual claims relative to either the price of the underlying asset or relative to European options written on that asset. For these claims, we have not needed to know the diffusion coefficient of the underlying asset. To value other kinds of claims (e.g., finite lived ones), it will be useful to examine the diffusion coefficient of S as we will show that it appears in the Jacobian when we develop the transition PDF of S. To obtain the diffusion coefficient of S, we use Itô’s formula on (6): dSt=√2πN0 √2πXt H!hdWt,t≥0, (25) as we already know that S is driftless. Since N0(·) and h are positive, so is the diffusion coefficient of S . Evaluating (5) at Xtand inverting implies: √2πXt H=N−1St H, (26) where N−1(p):( 0, 1 )7→ R is the inverse of the standard normal CDF. Substituting (26) in (25) gives the SDE followed by S: dSt=a(St)dWt,t≥0, (27) where the diffusion coefficient (normal volatility) of Sis given by: a(S)≡√2πN0N−1S Hh=e−[N−1(S H)]2 2h. (28) We note from (27) that S is indeed a time-homogeneous driftless diffusion. The diffusion coefficient given in (28) is proportional to a standard Gaussian function of N−1S H . As a consequence, movements in Sbecome more certain when Sis near 0 or H: lim S↓0a(S) = 0 lim S↑Ha(S) = 0. (29) Besides its starting value S0∈( 0, H) , the process S has two free parameters H and h . We have already seen that H defines the right end point of the domain of S . As S fluctuates through ( 0, H) , its volatility √2πN0N−1St Hh evolves as a positive stochastic process. We now show that h defines the right end point of the latter process’ domain. The standard normal density function N0(·) achieves its maximum value when its argument vanishes. The inverse normal CDF N−1(p) vanishes when its argument is 1 2 . Hence, we conclude that volatility is maximized when St=H 2 , i.e., at the midpoint of its domain ( 0, H) . The maximum value of volatility achieved is: √2πN0N−11 2h=h. (30) Now the standard normal probability density function (PDF) is even about zero which implies that its integral, N(·) is the sum of 1/2 and a function which is odd about zero. It follows that N−1(·) is odd about 1 / 2. Since diffusion coefficient of S is proportional to the composition of the even function N0(·) with the function N−1(M H) of S which is odd about H 2 , the diffusion coefficient of S is symmetric about H 2. Risks 2017,5, 61 8 of 11 The speed density of a diffusion with scale density s(x) and variance rate σ2(x) is 1 σ2(x)s(x) . Since S is a process in natural scale, s(x) = 1 and its speed density is simply the reciprocal of its variance rate. Hence, (27) implies that the speed density of Sis given by: m(S)≡1 2πhN0N−1S Hi2h2 =e[N−1(S H)]2 h2,S∈(0, H). (31) Let τDU be the first time that S exits the interval [D , U] . Suppose we consider how the mean of the random variable τDU behaves in the limit as we set D=S−e , U=S+e and let e> 0 shrink down to zero. Clearly, the mean exit time approaches zero, but the question is at what speed. For a process in natural scale such as S ,Karlin and Taylor (1981) show on page 197 that the mean exit time E[τ(S−e,S+e)|S0=S] approaches zero like O(e2) , where the coefficient is given by the speed function m(S), i.e., m(S) = lim e↓0 E[τ(S−e,S+e)|S0=S] e2. (32) This is the likely origin of the term “speed density”. We note that the higher the diffusion coefficient of a process in natural scale, the lower its speed density. This observation prompted Rogers and Williams (1994) to jestingly suggest that m(·) alternatively be called a “sloth density”. As indicated in (31), the speed density of the S process is inversely proportional to a Gaussian function of N−1S H . As a consequence, S exits intervals much more slowly on average when it is near zero or Hthan when it is near H/2. Recall the rough interpretation of the local time of a continuous martingale as the amount of quadratic variation occurring at a point. More precisely, local time captures the stochastic rate at which the quadratic variation experienced below some point in space increases as we increase the point. Loosely speaking, the speed density of a time homogeneous one dimensional diffusion can be interpreted as the expected calendar time spent at a point, until the first time that the diffusion exits an interval. This rough interpretation is meant to be contrasted with the expected quadratic variation spent at a point until the first exit. Thus the speed density is used to convert the expected local time of a stochastic process into twice the spatial density of the occupation time. To illustrate these points, let M now denote a time homogeneous one dimensional diffusion martingale: dMt=a(Mt)dWt,t≥0. (33) From (19), the local time at Kevaluated at the random time τDU is: LM τDU (K)≡2ZτDU 0 δ(Mt−K)a2(K)dt. (34) In contrast, twice the density of the occupation time at Kuntil τDU is: 2ZτDU 0 δ(Mt−K)dt. (35) From the Tanaka Meyer formula, the cost of creating the payoff in (35) is: G(M,K) = (2(M−D)(U−K) U−Dm(K),−∞<D<M<K<U<∞ 2(K−D)(U−M) U−Dm(K),−∞<D<K<M<U<∞.(36)