The Predictive Power for the Quarterly Earnings Per Share based on Time Series and Artificial Intelligence Model

• Main Authors: Ya-Jung Shih, 施雅蓉
• Other Authors: Syou-Ching Lai
• Format: Others
• Language: en_US
• Published: 2004
• Online Access: http://ndltd.ncl.edu.tw/handle/29179758900041335723

碩士 === 國立成功大學 === 財務金融研究所 === 92 === 　　The purpose of this study is to compare the forecasting ability among the ARIMA model, the Transfer Function model, the Artificial Neural Network model and the Genetic Algorithm model. To evaluate the forecasting accuracy, there are two dimensions taken into consideration: 1) the deviation between the actual quarterly EPS value and the forecasted quarterly EPS value, and 2) the changing direction from quarter to quarter between the actual quarterly EPS value and the forecasted quarterly EPS value. 　　In the aspect of the deviation between the actual quarterly EPS value and the forecasted quarterly EPS value, the empirical results show that the Transfer Function model outperforms the ARIMA model. Therefore, the settings of time lags of the Transfer Function model are adopted to the other two models. The empirical results reveals that the Genetic Algorithm model shows the best forecasting accuracy in both dimensions while the Artificial Neural Network model shows the worst forecasting accuracy in both dimensions. 　　In addition, both of the quarterly basic EPS data and the quarterly diluted EPS data were applied in forecasting future quarterly basic EPS. There is not enough evidence to support that using the diluted EPS data would yield higher accuracy than using the basic EPS data in the aspect of deviation. However, the empirical result shows that using the basic EPS data outperforms using the diluted EPS to forecast future basic EPS in the aspect of predicting the directions.