Abstract:
This thesis consists of two parts focusing on predicting the SET50 index return. The first part uses a support vector machine model to predict the daily directions of SET50 index returns while the second part uses the index future returns from the Hangseng, the Dow Jones and their combinations as inputs of a two-stage model which is composed of a support vector machine model for the directional prediction in the first stage and a neural network model for the value prediction in the second stage to predict five-minute SET50 index returns. Firstly, in this thesis, the combined data preprocessing and classification technique for the stock return direction is proposed. The support vector machine is selected due to its popularity for the stock return direction prediction. To give additional information for the support vector machine, the higher order differences and the higher order lags are fed as the additional inputs. Our experiments show that the predictive accuracy with respect to the number of difference orders and the number of lags are statistically insignificant. Nonetheless, the best setting of the support vector machine model shows the accuracy improvement over the other models such as the neural network model by Fernandez-Rodriguez et al. (2000). Secondly, the two-stage model for the five-minute stock index returns is proposed. The inputs are the Hangseng and the Dow Jones index future returns and their combinations. Returns with small magnitude are filtered out at different thresholds ranging from 0 percentile to 95 percentile. Our two-stage model uses a support vector machine model to generate the direction of the five-minute stock index return and attaches it as the input of the second stage which is a neural network to predict the value of the five-minute stock index. Our two-stage model outperforms a single neural network model in terms of accuracy. The magnitude of the stock returns used in the model affects the predictive performance. Dropping low percentile ranks of the stock index returns improves the predictive accuracy.
