/

Derivative Pricing – Semi-Analytic Lattice Integrator Approaches

1. Introduction

[this page | pdf | references | back links]

Return to Abstract and Contents

Next page

 

1.1          The price of a (European-style) derivative can be calculated as the discounted expected value of its payout function, where the expectation is carried out using the risk-neutral probability distribution. If the payout function and the risk-neutral probability distribution are both simple functions then the price of the derivative may be derivable analytically. For example, the Black-Scholes formulae can be derived analytically (see e.g. Black-Scholes derivation as the limit of a binomial tree or Black-Scholes derivation using stochastic calculus). The hedging parameters, i.e. greeks, applicable to the derivative may also be analytically tractable (see e.g. hedging parameters applicable to vanilla and binary puts and calls in a Black-Scholes world). However, for more complicated pay-offs or more complicated risk neutral probability distributions it is usually not possible to derive equivalent analytical formulae.

 

1.2          The most common approach used to circumvent this problem is to use numerical approaches such as binomial or trinomial trees that converge to the correct price or other hedging parameter as the tree becomes more and more finely grained. Unfortunately, these algorithms are not usually very good at handling singularities in derivative payout functions. These singularities can arise directly in the payout function, e.g. payout functions applicable to digital options have discontinuities at the strike price. They can also arise via discontinuous first partial derivatives with respect to the underlying (price) process. These are more common. For example, the first derivative of the payout function of a vanilla call option is discontinuous at the strike price because the payoff function has a kink there.

 

1.3          Some of the problems these discontinuities create can be mitigated by judicious choice of where to position the nodes of the relevant tree. However, an arguably better approach, if it is practical to implement, is to approximate the payoff function as the sum of components that are analytically tractable. In particular, it is often possible to find payoff functions with specified domain limits (i.e. ranges over which they are non-zero) that are analytically tractable. This leads to the semi-analytic lattice integrator (‘SALI’) approach, see e.g.  Hu, Kerkhof, McCloud and Wackertapp (2006).

 

1.4          As with other derivative pricing approaches, the SALI approach notes that the value, , of a (non-dividend paying) derivative at time , relative to the chosen numeraire asset, , satisfies:

 

 

where  is the information set (or filtration) generated by the underlying processes and  is the relevant risk-neutral expectation operator. Usually, we will restrict ourselves to Markovian models, in which case the above formula can also be written as:

 

 

 

1.5          The observation underlying SALI is that if the payout can be written as a function of the underlying Markov process then it can be decomposed into the sum of a finite number of smooth subcomponents i.e. as:

 

 

where  denotes a (typically) low-dimensional underlying Markov process.

 

Here the  are indicator functions, i.e.  if ,  otherwise. The decomposition might use one smooth function between consecutive discontinuities, or it might use several that are pasted together, e.g. cublic spline functions.

 

1.6          The pricing problem can thus be re-expressed as:

 

 

where  where .  can in practice be truncated to be within some ‘envelope of support’ that includes essentially all of the probability density applicable to the pricing problem. For example, for a Weiner process, one might use an outer envelope spreading out to, say, four standard deviations, since virtually none of the probability density is outside this spread. However, care is needed in such a truncation if the payoff function becomes sufficiently large sufficiently rapidly at the edge of the distribution, see e.g. the Cost of Capital pricing model.

 


NAVIGATION LINKS
Contents | Prev | Next


Desktop view | Switch to Mobile