Hull White Tech Note
Hull White Tech Note
Hull White Tech Note
1 A Continuous Process
The starting point is a stochastic process for a continuous state variable z :
zt+h
, zt = ,zth +
"t;t+h ;
1
where h is a presumably short time interval measured in annual units and f"t;t+h g NID0; h. Parameter 1 0 governs the rate of mean reversion, with large values indicating faster mean reversion. Later on, we let the short rate be z plus a time-dependent constant. Other examples de ne z as the log of the short rate or other monotonic function. For future reference, consider the salient properties of the process for z . Over the interval h, the rst two conditional moments of z are
which de nes M = ,h 0 and V = 2h. Generally both h and are fractions. In any case, we assume ,1 M 0.
Probabilities
Overview. The trick is to pick probabilities for the tree that reproduce the mean and variance of 1. Two issues arise. i We have three parameters two probabilities and to reproduce two moments the mean and variance of z . Where does the extra parameter come from? ii How do we know the probabilities are positive? This puts limits on i. Interior nodes. At each node, we choose the probabilities to match the mean and variance of changes in z . For a node with arbitrary j n doesn't matter here, since z is stationary, the mean and variance are E z = u i + m i0 + d i, = M i h i Var z = E z 2 , E z 2 = u i 2 + m i02 + d i, 2 , jM 2 = V These equations are linear in u ; d. At i = 0, u 0 = d 0 = V =2 2 = 0 say. For arbitrary i, the probabilities are 1 hiM + iM 2i u i = 0 + 2 m i = 1 , 20 , iM 2 2 h i 1 ,iM + iM 2 : d i = 0 + 2 This version illustrates one of the choices involved in this modeling e ort: we can choose the initial up and down probabilities 0, or we can choose the distance between nodes, but we can't choose them independently. The missing moment. Hull and White set 0 = 1=6; the question is why. Our answer perhaps theirs, too is that we are trying to approximate a normal process. In addition to mean and variance, we might consider central moments k k = E z , E z 2
for k 2. With k = 3 we are measuring skewness, with k = 4 kurtosis. Skewness is zero for both the continuous normal process and trinomial tree: both are symmetric. The choice 0 = 1=6 reproduces the kurtosis of the normal 4 = 3V 2 in state i = 0 the initial node. The fourth moment satis es = 0 4 + 1 , 2004 + 0 , 4 = 3V 2 : Since V = 0 2 2, we nd 0 = 1=6 and = V =201=2 = 3V 1=2, as suggested by Hull and White. In states i 6= 0, skewness is nonzero and kurtosis is di erent from the normal. What we care about, however, is the conditional distribution over several periods at the initial node, something we won't go into here. Roughly speaking, however, the central limit theorem makes average changes more normal the more periods we have. Comment: not quite true here because of mean reversion...
4
The edges. Hull and White bound the tree at jij = imax , a xed number whose value we'll go into later. For now, the question is how we set the probabilities at the edges of the tree. A picture is useful here: see Exhibit 2 of Hull and White 1994 or Exhibit 1 of Hull and White 1996. The idea is to point the branches back into the tree: at the upper edge, we have branches pointing same, down, more down. If we call these by the same names up, middle, down, the probabilities solve these equations, = Var z = The solution is
u i E z
V:
h i = 1 + 0 + 1 3iM + iM 2 2 m i = ,20 , 2iM , iM 2 1 hiM + iM 2i : d i = 0 + 2 At the bottom edge, we reverse u and d . Note, in general, that u i = d ,i, so it's su cient to specify the upper edge.
is initially decreasing in i, but for 0 1=8 it starts increasing again without going negative. m is decreasing throughout, hitting zero rst at iM = ,1 , 201=2. With 0 = 1=6, i is bounded by 2=31=2 : i 3 ,M Remember: M is negative and small. d is increasing in i so we can ignore it as a relevant bound on i.
u
Now consider the edge probabilities: initially declines with i, then increases. As long as 0 1=8 it's always positive. m initially increases, then decreases. For 0 1=8, it takes positive values in the interval 1 , 1 , 201=2 i 1 + 1 , 201=2 4 ,M ,M
u d
In both the interior and on the edge, then, the binding constraints on i come from the middle probability. The edge places both upper and lower bounds on imax, since for i small, the structure of the tree on the edge no true up move is inconsistent with mean reversion which is small for small i. The interior places an upper bound on imax , 1. The issue is how the two interact. De ne a = 1 , 201=2. Consider the upper bound in relation 4, imax = ,1 + a=M . With this value, there is no guarantee that imax , 1 satis es the interior condition 3. Alternatively, consider the lower bound in 4, imax = ,1 , a=M . In this case, imax , 1 automatically satis es 3 as long as 0 3=8 this is a su cient condition. We should also worry about i being an integer, but you get the idea. Perhaps for this reason, Hull and White 1994, p 12 suggest choosing 1 , 1 , 20 1=2 = ,0:18350 ; imax = ,M ,M the last equality stemming from the choice 0 = 1=6.
Summary
We take as given values of M and V summarizing, respectively, the mean reversion and volatility of the process for z . Given these values, we set 0 = 1=6 and = 3V 1=2. The number of states, imax , is set equal to the smallest integer greater than 1,201=2,1 =M = ,0:18350=M . Probabilities are then set by 2 in the interior jij imax and by 3 on the upper edge i = imax. On the lower edge, we use the relation u i = d ,i. 4
Given z , we set short rates as follows. De ne z i; n = z i; n + n , where = 0 ; 1; : : : is a vector of constants that depend on n but not i. Now de ne the short rate by one of the following: z i; n ri; n = exp z i; n
0 0
The former applies to a linear Ho and Lee" kind of tree, the latter to a log-linear BlackDerman-Toy" kind of structure. Finally and this is simply a collection of arbitrary conventions, de ne the one-period discount factor in each state by bi; n = e,ri; nh=100: The conventions are that short rates are continuously compounded and reported as annual percentages. Calibration follows the usual routine: we choose to reproduce observed spot rates. We'll do an example shortly.
4 Valuation
Consider a claim to state-contingent cash ows cj; n. There are two standard approaches to valuation. One is recursive: one period at a time, starting with the last cash ows. The other is to apply state prices Qi; n: X p = Qi; nci; n 5
i;n
We describe each in turn. We follow with no particular justi cation standard practice in equating true probabilities above and risk-neutral probabilities. With both of these methods, it's convenient to imbed the model in the more general setting of a Markov chain . In a Markov chain, a state variable like our z assumes a nite number of values, indexed by i. In the Hull and White model, i is an integer between ,imax and imax. Transition probabilities i; j describe the likelihood of moving from a given state i this period to another state j the next: = Probzn+1 = zj jzn = zi : The collection of such probabilities can be put into a transition matrix" . In our case, the transition matrix for imax = 3 is
i; j
Current State i 3 2 1 0 1 2 3
Future State j 1 0 1 u ,3 0 0 u ,2 0 0 m ,1 u ,1 0 d 0 m 0 u 0 0 d 1 m 1 0 0 d 2 0 0 d 3
3 0 0 0 0 0 u 2 u 3
In this way, the probabilities described earlier de ne the elements i; j of the matrix . This more general structure saves us work later on in allowing us to ignore things like the fact that state ,1 next period can be reached from 4 di erent states the previous period, but state 0 can only be reached from 3 read down the appropriate columns.
Recursive valuation
One approach to valuation is to start at the end, and value the cash ows at each node by discounting the cash ows of the nodes stemming from it. In the nal period nmax , the value of an asset is the value of the cash ow in that state:
pi; nmax
= ci; nmax:
In each node i; n for n nmax , the value is the current cash ow ci; n plus the sum of the values of the nodes j leading out of it: X pi; n = ci; n + bi; n i; j pj; n + 1: 6
j
This is where we assume that the 's are the risk-neutral probabilities. For a typical interior node, we could also write this as
pi; n
The Markov chain notation saves us from having to treat nodes di erently depending on how many nodes lead to it: the zeros in take care of this automatically.
State Prices
A second approach to valuation is to apply state prices directly to cash ows. De ne the state price Qi; n as the price in state 0,0 of one dollar payable in state i; n. State prices are readily computed by X Qj; n + 1 = i; j bi; j Qi; n 7
i
We start at the beginning, with Q0; 0 = 1 and Qi; 0 = 0 for i 6= 0, and use 7 to compute Q one period at a time. The sum over i here means over all states i at date n that can lead to j at n + 1. State prices can be applied directly to cash ows to value assets using 5. They can also be used to compute discount factors and spot rates when the model is calibrated. Discount factors and continuously compounded spot rates at date n = 0 are X bn = Qj; n j X bn+1 = Qj; nbj; n
yn
= ,100=nh log bn :
For small h these expressions are pretty much the same as those given earlier, but we stick with theirs to keep the numbers the same. Interest rate grid. The distance between interest rates in the tree is = 3V 1=2 = 0:0164895. The maximum state is the largest integer greater than 0:18350=0:095162 = 1:93, namely 2. That gives us 5 states, ranging from ,2 to 2. Probabilities. We apply 2,3 to get the transition probabilities 0.8993 0.0111 0.0896 0 0 0.1236 0.6576 0.2188 0 0 0 0.1667 0.6667 0.1667 0 0 0 0.2188 0.6576 0.1236 0 0 0.0896 0.0111 0.8993 Short rate tree. The z tree is
n
=0 0
=1
1.6490 0 1.6490
Hull and White give us = 5:09275; 6:50257; 7:33932; 8:05381, which gives us a short rate tree of
n=2 10.6372 8.1515 8.9883 5.0928 6.5026 7.3393 4.8536 5.6904 4.0414 n n
=0
=1
=0
=1
1.00000
n=3 n=4 0.03662 0.18772 0.36842 0.19466 0.03925 0.82767 0.76388 6.3046 6.7335
Calibration. This is the answer. In general, we would observe spot rates in the data and use a numerical procedure to adjust until the spot rates in the model matched those in the data.
6 Final Thoughts
Hull and White provide a useful discrete-state approximation to the Vasicek model and a framework that can be applied to other models as well. Their method is relatively simple to apply and apparently works well in practice, but it's not the only way to approximate a continuous model. Nelson and Ramaswamy describe how binomial models might be designed to do similar things. Tauchen and Hussey 1991 go the other direction, and consider Markov chain approximations to continuous processes. And the math literature is lled with numerical approaches to solving PDEs" see Judd 1998. The key issues are the speed and simplicity of the calculations. What we'd like to see is a systematic assessment of various approaches with these criteria.
References
John Hull and Alan White, 1994 Numerical procedures for implementing term structure models I: Single-factor models," Journal of Derivatives , Fall, 7-16. John Hull and Alan White, 1996, Using Hull-White interest rate trees," Journal of Derivatives , Spring, 26-36. Kenneth Judd, 1998, Numerical Methods in Economics , MIT Press. Daniel Nelson and Krishna Ramaswamy, 1990, Simple binomial processes as di usion approximations in nancial models," Review of Financial Studies 3, 393-430.