CityU Institutional Repository >
CityU Electronic Theses and Dissertations >
ETD - Dept. of Management Sciences >
MS - Doctor of Philosophy >
Please use this identifier to cite or link to this item:
|Title: ||Computational intelligence based models for earnings per share forecasting|
|Other Titles: ||Ji yu ji suan zhi neng de mei gu shou yi yu ce mo xing|
|Authors: ||Dong, Gang ( 董綱)|
|Department: ||Department of Management Sciences|
|Degree: ||Doctor of Philosophy|
|Issue Date: ||2011|
|Publisher: ||City University of Hong Kong|
|Subjects: ||Earnings per share -- Forecasting.|
|Notes: ||CityU Call Number: HG4028.E27 D66 2011|
ix, 111 leaves 30 cm.
Thesis (Ph.D.)--City University of Hong Kong, 2011.
Includes bibliographical references (leaves 103-111)
|Abstract: ||Earnings per Share (EPS) is the amount of earnings per outstanding share of a company's stock, which is the basic measurement of financial performance of a company. Earnings analysis, especially earnings forecasting, plays a critical role in many investment and other financial decisions. EPS forecasting has long been of considerable interest to researchers and practitioners. Many investors select stocks on the basis of companies' earnings forecasts which are often used to determine earnings expectations when investigating firm valuation, cost of capital and the relationship between earnings and stock prices. Since the early 1980s, earnings forecasting research has become closely aligned with capital markets research. Capital markets research requires a proxy for (unobservable) market earnings expectations and earnings forecasting research has provided such proxy measures.
Traditionally, forecasts by analysts from big financial institutions are used as proxies for the unobservable 'market' expectations of a future earnings realization. However, these forecasts have two disadvantages: 1) researchers have found analysts' forecasts have systemic biases which may be a result of intentional manipulations by analysts; and 2) for individual investors it is not easy to get analysts' forecasts as investors need to pay for these forecasts. As an alternative to analysts' forecasts, forecasts generated by statistical models do not have the above disadvantages. Forecasts generated by statistical models are objective and free.
Autoregressive Integrated Moving Average (ARIMA) model was the first statistical model used for quarterly EPS forecasting. Due to the excellent ability of capturing properties of time-series, various ARIMA models have been developed by researchers and these models have been the core of the early EPS forecasting literature. However, none of ARIMA models has achieved satisfactory forecasting accuracy. Researchers have always been making efforts to search for methods that can improve the accuracy of statistical models. Recently, researchers began to try different models in two ways: 1) using nonlinear models to identify the nonlinearity of EPS time-series data; and 2) incorporating accounting indicators in financial reports in the forecasting models. For the first way, Neural Network (NN) model was the first computational intelligence model applied to capture the nonlinearity of EPS time-series data. For the second way, NN models are still used to build the relationship between multiple inputs (past EPS data and accounting indicators) and a single output (future EPS data). However, NN models have some disadvantages such as over fitting. Support Vector Machine model (SVM), a new computational intelligence method, has been introduced, and its performance has been proved to be better than NN model by empirical tests. In this research, Least Squares Support Vector Machine (LS-SVM), the novel version of SVM, is proposed to overcome the disadvantages and improve performances of NN models. Based on LS-SVM models, four new types of statistical models are proposed in this research for quarterly EPS forecasting: univariate LS-SVM model, hybrid models of ARIMA and LS-SVM models, multivariate LS-SVM model and combined model.
The proposed models are compared with models introduced by other authors in previous literature using the data sample of 232 listed American companies. The empirical results show that Median Value model (combined model) and multivariate LS-SVM model have the best performances and outperform the other models in terms of performance score, MAPE, large errors and average rank. Although these two models have the best performances in accuracy, they are not the best models when measured by association of abnormal returns and earnings surprises. In terms of association, traditional ARIMA models have the best performance among all models. Therefore, we conclude that Median Value model and multivariate LS-SVM models have the best accuracy and traditional ARIMA models have the strongest association.|
|Online Catalog Link: ||http://lib.cityu.edu.hk/record=b4085914|
|Appears in Collections:||MS - Doctor of Philosophy |
Items in CityU IR are protected by copyright, with all rights reserved, unless otherwise indicated.