Mohammad Rasool Chopani; Farzaneh Nassirzadeh; Mahdi saehi
Abstract
Earnings per share is one of the most important financial statistics, mostly used in the evaluation of profitability, the risk associated with earning, and the stock price. In many countries, the importance of this measure is to the extent that it is considered as one of the principal scales in determining ...
Read More
Earnings per share is one of the most important financial statistics, mostly used in the evaluation of profitability, the risk associated with earning, and the stock price. In many countries, the importance of this measure is to the extent that it is considered as one of the principal scales in determining the stock price and is widely used in stock evaluation models. Thus, to predict the earnings per share, different algorithms have been used, some of which make use of the statistical models and some smart models.
The studies recently conducted on the precision of smart models show that in comparison to the statistical models, the smart models have performed better in classification and finding of an efficient solution. Thus, those investors using these models will find the investment opportunities much more efficiently. Therefore, regarding this necessity, this study has compared the errors of such models as Support Vector Estimator, Minimum Degree Estimator, and Fuzzy Neural Network (which are among the smart models having the lowest error rates) in predicting the earnings per share for the firms enlisted in Tehran Stock Exchange during the years of 2005 to 2012.
Research Methodology
In this research, nineteen different independent variables have been used in the three financial, fundamental, and macro groups. The relationships between the first group of variables and the earnings per share in the paper by Zhang et al. (2004), and the relationship between the second group of variables and the earnings per share in the papers by Lexian et al. (1390, 2011) and Brid (2001) have been confirmed. The dependent variable in this research is the annual earnings per share. Therefore, the current research tries to find a model, which has the highest precision in predicting the earnings per share, using the independent variables, either individually or in groups.
The selected sample in this research include 171 firms in 27 active industries, during the years of 2005 to2012, through random sampling and using cluster sampling from among the active firms in Tehran Stock Exchange.
After being collected and standardized, the data was classified into training and experimental data, using the K-Fold Cross-Validation method. The percentage of training data to the experimental data is assumed as 30-70 or 20-80; in this research, the 20-80 composition has been employed. In this research, the amount of K has been determined as 10.
Then, the main process of modeling is conducted in a way that the prevalent patterns and relations between the data (independent and dependent variables) are extracted, using the techniques of Support Vector Estimator, Minimum Degree Estimator, and Fuzzy Neural Network. In this stage, the training data are used for modeling. After extracting the data patterns, the precision of the proposed model is estimated, using the experimental data, and finally, to explore the models’ precision, such error measures as mean square error (MSE), Median Absolute Deviation (MAD), and determination coefficient have been used.
Research Findings:
The results show that when all the fundamental and financial variables are used simultaneously, the precision of the estimator model is highest. When the fundamental variables are used in LARS, the MSE and MAD are 3.505 and 306.301 respectively. When the financial variables are used in LARS, the MSE and MAD are 0.921 and 206.669 respectively, and finally, when all the variables are used in LARS, the MSE and MAD are 3.414 and 392.081 respectively. The obtained results have been presented in sum in the following table.
Conclusion
In this research, the models have been evaluated annually; the models have been conducted on each year, and the results have been compared with each other. Finally, the average annual errors have been considered as the basis of determining a more precise model in every state. Exploring the models’ final error the Minimum Degree Estimator model predict the earnings better than the Fuzzy Neural Network and Support Vector Estimator. Also, exploring the average errors for the Minimum Degree Estimator model shows that using the financial variables has resulted in the increase in the predicting capabilities of this model.
Keyword: Tehran stock exchange, Earning per share, Support vector regression, Least angel regression, Adaptive neuro Fuzzy Inference system.
