China Warrants The Price Of The Neural Network Prediction Model And Its Structural Optimization

As the development of the Chinese security market, especially the sell-short mechanism has just recently been introduced into our security market, which offers the necessary condition of introducing the derivatives. However, the traditional pricing models or methods that is based on the developed countries maturity security market with lots of assumptions could not describe the real market especially the derivatives traded in China. Hence we need to develop a new type of model to pricing or predicting price of the warrants which are trading in our Chinese market now.Using the artificial neural networks the nonlinear prediction models for the Chinese warrants has successfully been modeled. Under the construction of such nonlinear models, we develop an algorithm initiatively to solve the problem about how to screen the input indexes for the neural networks while facing with multiple indexes. Empirical tests also have been carried out for the nonlinear models, and the results showed that the hit ratio of ups and downs for the next five days was 80%The thesis starts with a brief literature to demonstrate that the financial markets may be predictable to some extents and in probabilityAnd then, we introduce the traditional models and methods and test them using a historical data from the Chinese warrant market. We show with evidence that those models and methods are unable to price the Chinese warrant.In the third section of this paper, we introduce a brief notion of the nonlinear prediction model and show the way how to construct the model using the Artificial Neutral Networks.In the following parts, a nonlinear prediction model is worked out by using of the Multilayer Feedforward Neural Networks (MFNN) with Back Propagation (BP) training algorithm and another model using the Radial Basis Function Neural Networks (RBFNN), and the two models are compared with the traditional ones on the same historical data. The empirical tests show that the model with the best performance is the one modeled using the MFNN with BP algorithm, the second best one is the model using the RBFNN, and the ones which perform the worst are the traditional models and methods.