The standard BDT model is constructed algorithmically to be consistent with both the existing term structure of zerocoupon yields, and (optionally) the term structure of yield volatilities. That is, the main aim of the tree building procedure is to derive a binomial representation for the level of the shortterm interest rate such that zerocoupon bond prices computed from the tree are exactly equal to the set of zero coupon prices that are directly observable in the market. If desired, the model can also be constructed so that the implied interest rate distribution at each time step matches an observed interest rate volatility curve. 

A general method for constructing binomial trees for the short rate is described fully in Hull and White (1996), Rebonato (1998), and Jamshidian (1991). Clewlow and Strickland (1998) present a comprehensive summary of the Jamshidian construction method, and that reference is the main source of the discussion that follows. 

Consider constructing a binomial tree for the short rate that spans a period between time 0 and time T. The entire period can be split into an arbitrary number of i = 1 to N periods such that the length of each period is equal to Dt = T / N. With two discrete outcomes possible from every node, there are a total of i + 1 nodes at each of the i 'th timesteps. This general binomial representation is shown in the following figure, where each node is indexed as (i,j). 

The number of states at each timestep is indexed by j. At the initial branch in the tree (i = 0), there is only 1 state and that is indexed as j = 0. At the end of the first period, the short rate has either increased to state j = 1 or decreased to state j = 1. In general, the rate at any arbitrary node i, j (that is not at the extreme top or bottom of the tree) will either move to a node indexed as i + 1, j + 1 (after an upjump) or i + 1, j  1 (after an downjump). The extreme top and bottom nodes are indexed as N, N and N, N, respectively. 

As at time 0, we assume that both the term structure of zero coupon rates and rate volatilities are known. These curves are typically defined for a discrete number of points, and some interpolation scheme will be used to determine rates for maturity dates that are not explicitly included in the input data. Assuming that the zero rate for maturity i is defined as ZR(i ), then a discount function can be defined as: 



where, 
P(0,i ) = the time0 price of a zerocoupon bond with unit face value that matures at time i Dt = the length (in years) of the timestep assumed in the tree construction 

Note that this representation of the discount function assumes that the input zero rates are defined with a compounding frequency that matches the chosen length of the timesteps used in the construction of the tree. 

The variable of interest at each node in the tree is the shortterm interest rate that will apply over the next time interval, r _{i, j}. This rate is known at the beginning of each period and defines the current one period discount rate: 



where, 
DF_{i,j } = the time i discount factor for state j Dt = the length (in years) of the timestep assumed in the tree construction 

With this background, we can now consider the general intuition behind the tree building process set out by Jamshidian (1991). The technique is referred to as a forward induction process, and involves working through each successive branch in the tree searching for the collection of shortterm interest rates which will return a zerocoupon bond price that is consistent with the observed bond price. Further, if the model is also fitted to observed yield volatilities then the chosen set of rates must be consistent with both yields and volatilities. 

Jamshidian shows that the level of the short rate at the i'th step in the tree is characterised by: 



where, 
U(i ) = the median of the lognormal distribution for r at step i s(i ) = the shortrate volatility at step i z(i ) = the value of a standard Wiener process 

Clewlow and Strickland (1998) show that for each step i in the binomial tree, the expression in the above equation can be operationalized to give an expression for the short rate at each node in the tree. That is, 



The remainder of the tree building process is then related to determining the appropriate values for both U(i ) and s(i ) such that the resulting tree is arbitragefree. That is achieved in part using the concept of state prices (also referred to as ArrowDebreu prices), which are defined as the time zero value of a pure security that provides a payoff of $1 if node (i, j) is reached, and zero otherwise: 



where, 
the time zero state price for node (i, j) PV = present value operator 

Zerocoupon bond prices can also be defined in terms of the entire vector of state prices at step i : 



State prices can therefore be interpreted as the discounted probability of reaching any node in the tree, and they represent the basic building blocks that can be used to value any financial claim through time. Further, given that the state price at node (0, 0) is equal to 1 by definition and by assuming that the riskneutral probabilities of an upjump and downjump are both equal to 1/2, then we can construct the state prices at all successive nodes 'within' the tree using the following relation: 



The state prices for the top and bottom nodes in the tree at each step i are given by: 



Consider first just fitting a BDT tree to a set of known zerobond prices. By assuming a constant volatility (s(i ) = s), the expression in equation (from above) for the short rate is only a timevarying function of U(i ). At each step i in the tree, we have computed all j state prices and are attempting to determine the noarbitrage short rate at all j nodes. We do so by searching for the rates that satisfy the requirement that: 



That is, given that the current value of a zerocoupon bond maturing at time (i + 1) . Dt is known, we can use a technique such as NewtonRaphson to solve for the only unknown in the above equation, U(i ). 

Although the BDT model was initially derived specifically using the binomial lattice framework, several authors have subsequently shown that the model implies the following stochastic differential equations for the evolution of the short rate through time. These processes can be directly compared to the equivalent expressions for the HullWhite and BlackKarasinski models discussed in the following sections. The original specification of the model, fit to both the term structure of zero rates and the term structure of volatilities is: 



where, 
r = the instantaneous short rate at time t q(t) = value of the underlying asset at option expiry s(t) = time dependent short rate volatility 

When the BDT model is restricted to fit the initial term structure only, the expression in the above equation reduces to: 



where, 
s = the constant short rate volatility 


Copyright 2013 Hedgebook Ltd.