An empirical comparison of alternative option pricing models: Parametric vs. nonparametric approac

This dissertation consists of two parts. The first part compares alternative parametric option pricing models. In particular, we consider three recent models: (1) the constant elasticity variance model (CEV), (2) the stochastic volatility model (SV), and (3) the stochastic volatility and jump model (SVJ). The second part concentrates on the application of modern nonparametric neural network approaches on option pricing. The multilayer perceptron (MLP) network and the radial basis function (RBF) network are used in the construction of option pricing models.;In the original Black-Scholes option pricing model, the stock price distribution is lognormal and the volatility is constant. These restrictive assumptions have caused empirical biases in moneyness and time-to-maturity. Therefore, we study the empirical performance of a richer class of models. The CEV model allows a skewed distribution (in particular, noncentral chi-square distribution) for the stock price and hence introduces less bias in the moneyness dimension. The SV and the SVJ models allow the volatility to move randomly over time. This reduces the biases in both the moneyness and the time-to-maturity dimensions.;The neural network model, on the other hand, is not subject to parametric distributions. It "learns" from the data. This allows the option pricing not to be limited to the complete market assumption.;In this dissertation, the comparative study is conducted for various models. The empirical investigation employs the daily S&P 500 option prices. On the parametric front, our empirical results show that all models (CEV, SV, and SVJ) present a higher pricing accuracy over the original Black-Scholes model. Among the three models, SVJ performs better than SV, which in turn performs better than CEV. On the nonparametric front, the empirical fitting of the neural network methodology (MLP) outperforms the original Black-Scholes model but not the CEV, the SV, or the SVJ models.