Comparison of Machine Learning Performance for Earnings Forecasting
Asian Tax Journal Vol. 20 No. 6 (2019), pp. 9-34
Abstract
Many predict that the fourth industrial revolution will be triggered by the emergence of big data and machine learning(or artificial intelligence). In this context, big data is often likened to crude oil and machine learning is likened to the crude oil processing technology, and the concept of data science has also emerged, referring to scientific methodologies or processes that extract useful information for decision－making. In order to cope with such a trend, even in the field of accounting, researchers are struggling to introduce big data and artificial intelligence, but there are not many studies yet. In particular, it is difficult to find studies related to them in Korea. There have been enough studies that accounting earnings have the information contents that makes a difference in accounting information users’ decision making as a fundamental variable that determines the firm’s value. However, many studies have not been conducted that relate to the development of a predictive accounting earnings forecasting model. In this study, the accounting earnings forecasting model developed in the previous study was reinterpreted in the context of machine learning, and compared its performance with the predictive(machine learning) models known to represent further higher predictive performance to examine the possibility of introducing machine learning techniques in forecasting accounting earnings. To achieve the objective of this study, all 152 financial ratios for closing corporations in December were extracted and utilized from 2009 to 2018 among KOSPI companies belonging to the manufacturing sector provided by TS2000 of the Korea Listed Companies Association. This study reinterpret the findings of Ou and Penman(1989) that used the logistic model basically in the context of machine learning, and the predictability of these techniques was compared by adding the most commonly used models such as tree model, random forest model, and boosting method. As a result of this studies, the predictability difference between the models are statistically significant in overall, and the boosting technique with the highest predictive performance has about 10% higher predictive power than tree model that has lowest predictive power. These results are meaningful in that it is possible to provide more accurate accounting earnings forecast information using machine learning techniques, and it is expected this research to be the basis for the study using machine learning in various accounting research fields other than accounting earnings forecasting.
Keywords
- earnings forecasting
- machine learning
- predict performance
Related Articles
Intangible Asset Productivity and Analyst Forecast Error
25(1) 9-32
The Effect of Innovation Performance on Financial Analyst Forecasts in KOSDAQ
25(1) 107-128
Value－Relevance of Milestone Disclosures by High－tech Growth Firms
20(5) 129-163
Tax Avoidance and Corporate Transparency
20(2) 95-123
The Effect of the Information Content of Earnings after Mandatory IFRS Adoption on Financial Analysts’ Earnings Forecast Accuracy
19(3) 119-163