BROWNIAN M OTION Standard Brownian motion Defining properties Terminology § µ is the drift coefficient § σ is the diffusion coefficient (volatility) Construction from a standard Brownian motion process § Independent increments ( Bt − Bs is independent of {Br , r ≤ s} whenever s < t ) § Wt = W0 + σ Bt + µt § Stationary increments (the distribution of Bt − Bs depends only on t − s ) Geometric Brownian motion § Continuous sample paths ( t → Bt ) § B0 = 0 § § Gaussian increments ( Bt − Bs : N ( 0, t − s ) ) Further properties § cov ( Bs , Bt ) = min ( s, t ) § § § {Bt , t ≥ 0} is a Markov process {Bt , t ≥ 0} is a martingale {Bt , t ≥ 0} returns infinitely often to any level 1 Bct also defines a standard Brownian motion process (scaling property) § B1 ( t ) = c § B2 (t ) = tB1 t also defines a standard Brownian motion process (time inversion property) Result § The probability that the sample path of a standard Brownian motion is differentiable at any given time t0 is zero. (Proof required) Brownian motion in general Defining properties § Independent, stationary, Gaussian increments § Wt − Ws : N ( µ ( t − s ) , σ 2 ( t − s ) ) St = eWt = exp (W0 + σ Bt + µt ) § log St : N (W0 + µt ,σ 2t ) MARTINGALES Martingales in discrete time Definition § The current value X m is the optimal estimator of the future value X n . Equivalently: § E X n < ∞ for all n § E X n X 0 , X 1 ,..., X m = X m for all m < n Conditional expectation framework § E E X Y = E [ X ] § E X Y is the optimal estimator of X based on Y in the sense that for every function h: E § E {( X − E X Y ) } ≤ E {( X − h (Y )) } . (Proof required) 2 2 {( X − E X Y ) f (Y )} = 0 (Proof required) Result § A martingale has a constant mean ( E [ X n ] = E [ X 0 ] for all n ). (Proof required) Martingales in continuous time Properties Definition t § ∫ Ys dBs , t ≥ 0 is a martingale 0 t § E ∫ Ys dBs = 0 0 2 t t t § E ∫ Ys dBs = E ∫ Ys 2ds = ∫ E Ys 2 ds 0 0 0 § E X t < ∞ for all t § E X t Fs = X s for all s < t Filtration framework § The filtration of a stochastic process X t is denoted Ft , and is the set of all events in the sample space that depend only on X s ,0 ≤ s ≤ t . § A random variable Y is Ft -measurable if the event {Y ≤ y} belongs to Ft for all values of y. § A stochastic process Yt , t ≥ 0 is adapted to the filtration Ft if Yt is Ft -measurable for all t. 0 Definition (general case) { } § If a sequence of step functions Ys( n) converges to the adapted integrand Ys , then the t sequence of integrals ∫ Ys ( n) dBs converges to the limit defined as ∫ Ys dBs . 0 0 t § This approach can be used for any Y satisfying E ∫ Ys 2 ds < ∞ . 0 § The properties of the simple case (above) are preserved. t Results § E t § The sample paths of ∫ Ys dBs are continuous {( X − E X F )Y } = 0 for all F -measurable bounded Y t t § E E X Ft = E [ X ] § If X is Ft -measurable, then E X Ft = X § If Y is Ft -measurable and bounded, then E XY Ft = Y .E X Ft § If X is independent of Ft , then E X Ft = E [ X ] STOCHASTIC CALCULUS Ito integrals Definition (simple case) Y 0 ≤ s < t1 § If Ys = 0 , where Y0 is F0 -measurable and Y1 is Ft1 -measurable, then: Y1 t1 ≤ s ≤ T t if t < t1 Y0 Bt § ∫ Ys dBs = 0 Y0 Bt1 + Y1 ( Bt − Bt1 ) if t ≥ t1 Ito processes Definition A time-homogeneous diffusion (Ito) process has the following defining properties: § It is a Markov process § It has continuous sample paths § There exist functions µ ( x ) and σ 2 ( x ) > 0 such that as h → 0+ E X t +h − X t X t = x = hµ ( x ) + o ( h) 2 E ( X t +h − X t ) X t = x = hσ 2 ( x ) + o ( h ) 3 E X t+h − X t X t = x = o (h ) § Similar to a general Brownian motion, but with variable drift and diffusion coefficients. General results Notation t t 0 0 § An Ito process can be defined in integral notation as: X t = X 0 + ∫ Ys dBs + ∫ Z s ds § A shorthand form is to use differential notation: dX t = Yt dBt + Zt dt t t § M t = exp ∫ f ( s ) dBs − 12 ∫ f 2 ( s ) ds is a martingale 0 0 t t 2 § ∫ f ( s ) dBs : N 0, ∫ f ( s ) ds 0 0 Ito’s Lemma Ornstein-Uhlenbeck process Lemma for a function of standard Brownian motion Stochastic differential equation § df ( Bt ) = f ′ ( Bt ) dBt + 12 f ′′ ( Bt ) dt dX t = −γ X t dt + σ dBt Lemma for a function of a diffusion process and time ∂f ∂f ∂f ∂2 f § df ( t , X t ) = Yt dBt + + Zt + 12 2 Yt 2 dt ∂x ∂x ∂t ∂x § Define U t = f ( t , X t ) = e γ t X t and use Ito’s lemma to calculate § Substitute for dX t and simplify § Express in integral form and then convert back to X t Derivation § Taylor’s formula gives: df ( t , X t ) = § Substitute dX t = Yt dBt + Zt dt ∂f ∂f ∂2 f 2 dt + dX t + 12 2 ( dX t ) ∂t ∂x ∂x § Use the fact that ( dt ) = dtdBt = dBt dt = 0 and ( dBt ) = dt 2 2 Geometric Brownian motion Stochastic differential equation § dSt = σ St dBt + α St dt Process to solve § Use Ito’s lemma to calculate d log St § Substitute for dSt and simplify § Express in integral form and then convert back from logs Solution § St = S0 exp (α − 12 σ 2 ) t + σ Bt § ( St : lognormal (α − 12 σ 2 ) t , σ 2t S0 ) Process to solve Solution t § X t = X 0e −γ t + σ ∫ e −γ ( t − s) dBs 0 Properties σ2 § X t : N X 0e −γ t , (1 − e −2γ t ) 2 γ § This process is the continuous equivalent of an AR(1) process. THE CONTINUOUS TIME LOGNORMAL M ODEL Definition § log Su − log St : N µ ( u − t ) , σ 2 ( u − t ) for u > t Where: § St is the security price at time t § µ is the drift § σ is the volatility Equivalently: § St = exp (W0 + σ Bt + µ t ) § log Su = log St + µ ( u − t ) + ε uσ u − t Momentum effects § Evidence of momentum effects is inconsistent with the independent increments assumption Non-normality of returns § Evidence suggests that there are more days with little or no movement in prices than would be expected if log returns were normally distributed (higher peak) § Evidence suggests that there are more days with very large price movements than would be expected if log returns were normally distributed (fatter tails) STOCHASTIC MODELS OF ASSET PRICES ( CONTINUED) Cross-sectional and longitudinal properties Properties Cross-sectional properties § Mean and variance of the log returns are proportional to u − t § Returns over non-overlapping intervals are assumed to be independent § E [ Su ] = St exp ( µ (u − t ) + 12 σ 2 ( u − t )) § Looks at the distribution over all simulations at a particular point in time § Dependent on the initial conditions § var [ Su ] = St2 exp ( 2µ ( u − t ) + σ 2 (u − t ) ) exp (σ 2 ( u − t ) ) − 1 Reasons why it may be inappropriate Volatility § A constant σ is inconsistent with historically observed market volatility § Implied volatility in option prices has fluctuated significantly over time Drift § It can be argued that µ should vary over time, particularly with interest rate changes Mean reversion § Evidence of mean-reverting behaviour is inconsistent with the independent increments assumption Longitudinal properties § Looks at a statistic sampled from a single simulation over a long period of time § Based on an average over varying market conditions, not those at a particular date Random walk § Because of the independent increments property, returns are independent over time § Cross-sectional and longitudinal simulations coincide (not true of other models)