The study has two objectives. The first, to develop an earnings movement prediction model to help investors in their decision process, the second, to explore the potential of Recurrent Neural Networks (RNN) in financial statement analysis and present a detailed model for its application. RNNs' two major advantages are: they do not make assumptions regarding the data and allow users to search whatever functional form best describes the underlying relationship between financial data and changes in earnings; they dynamically account for time – series behavior, earnings of a certain time period are not independent of earnings in previous time periods. The paper utilizes the newly mandated XBRL data, whose benefits are that it is freely available, easily accessible and is more timely than traditional data bases. The use of RNNs is validated in the results by providing a higher accuracy prediction than neural networks and logistic regression.
Keywords: Artificial intelligence; deep learning; earnings prediction; recurrent neural networks (RNN); XBRL; financial statement analysis.
Improvement of financial statement analysis modeling, by creating more advanced and comprehensive models, is conditional for developing the capabilities of capital market research (Vasarhelyi et al. 2015). There is a growing and promising accounting research applying advanced Artificial Intelligence (AI) tools in accounting estimates encouraging the development of new methods (Ding et al. 2020). This study attempts to add to this literature by presenting a comprehensive model for predicting earnings directional movement utilizing an AI technique, Recurrent Neural Networks (RNN), demonstrating that such models can be used successfully for financial statement analysis.
Financial statement analysis is the process of identifying aspects of the financial statements which can be used by decision makers to estimate future value of the firm. While stock prices represent firms' value they may deviate in the short run from this value. Fundamental analysis, of published financial statements, may discover values that are not reflected in the stock price and identify over and underpriced stock (Ou and Penman 1989). Fundamental analysis allows the user to test directly the validity of the economic intuition that underlies the original construction of the financial statements (Abarbanell and Bushee 1997).
Ou and Penman (1989) were the first to use linear regression to analyze financial ratios and indicate the directional movement of future earnings. Regression models' ability to forecast earnings depends on a variety of factors such as functional form, choice of predictors, choice of estimator and the behavior of error terms. Monahan (2017) therefore argues that regression models are ill-suited for generating forecasts of earnings mainly because during any reasonable length of time the likelihood that there will be a significant disruption at the firm-level, industry-level, or macro-level of any of these factors, is high. AI methods, on the other hand, are more flexible since they do not rely on restrictive statistical and economic assumptions (for example the linearity of the financial data) and they are not affected by reasoning preconceptions, instead, they exploit patterns and trends in the historical data to make predictions (Qi 1999). AI techniques are considered to be faster and more effective for intricate and computationally demanding decision processes (Danėnas 2013), and have also been found to be efficient in accounting processes (Granlund 2011).
Recently there have been attempts at using AI techniques in different areas of accounting: to predict the quality of accounting numbers in bankruptcy predictions (Barboza et al. 2017); in detection of misstatements (Bertomeu et al. 2021); in detection of accounting fraud (Bao et al. 2020) and in accounting estimates (Ding et al. 2020). These studies used different techniques such as: NN, gradient boosting machine, support vector machines, bagging, boosting, random forest, and more.
Studies of AI applications in accounting found NN to be the technique most widely used, however only a minor percentage of these studies was used for the study of financial data forecasting (Amani and Fadlalla 2017; Emerson et al. 2019). There have been some more current attempts in this area using other AI methods, more specifically: decision trees (Chen et al. 2021; Hunt et al. 2019; Anand et al. 2019) and Support Vector Machines (SVM) (Baranes and Palas 2019). Decision trees allow an interpretation of model parameters, NN discriminating power is often considered superior and may be due to the fact that they do not allow assumptions and interpretations of model parameters and interactions (Dreiseitl and Ohno-Machado 2002). SVM, like NN, does not rely on assumption regarding the model parameters interactions and usually achieves higher generalization performance than NN (Cao and Tay 2003). However, NN and SVM have two major limitations, the first is that they do not account for dependent outcomes over time. The second limitation is that they are hard to train if they become too complex and require a large amount of computation time when solving large-size problems (Batres-Estrada 2015; Cao and Tay 2003). The current study presents a deep learning NN model, RNN algorithm, which attempts to resolve these issues.
NN models are inspired by studies of the information processing abilities of the human brain. Key attributes of the brain's information network include a non-linear, parallel structure and dense connection between information nodes (Haykin 1998). NN models represent the information processing of the brain by linking layers of input and output variables through processing units called hidden nodes. In a NN an input is processed through a number of layers and an output is produced, with an assumption that two successive inputs are independent of each other (Hill et al. 1994). This assumption is not true in the prediction of earnings, the prediction of earnings at a given time is dependent on earnings in previous periods (Mikolov et al. 2015). RNN can be thought of as a series of networks linked together, they are called recurrent because they perform the same task for every element of a sequence using the same trained RNN cell, with the output being dependent on the previous computations. RNNs have two types of "memory" which captures the information in a specific sequence, decides if it should be updated with new information, and uses it for the next sequence(Mikolov et al. 2010).
Formally, a simple RNN has three layers which are input, recurrent hidden, and output. The input layer is a sequence of vectors through time. The hidden layer has hidden units that are connected to each other through time with recurrent connections. However, for any standard RNN architecture, the influence of a given input on the hidden layers and eventually on the NN output would either decay or blow up exponentially when cycling around recurrent connections. To tackle this problem, Long Short-Term Memory (LSTM) has been revolutionarily designed by changing the structure of the hidden neurons in traditional RNN (Hochreiter and Schmidhuber 1997). The hidden layers contain the LSTM network, which consists of a memory cell with a forget gate, an input gate, and an output gate. The input gate defines what information is added to the cell, the forget gate defines what information is removed from the cell, and the output gate specifies which information is to be used as output (Jang et al. 2021). Today, research and applications of LSTM for time series prediction are proliferating (see examples in Alahi et al. 2016; Wang et al. 2017).
Further explanation of the use of LSTM and its application in the current study is provided in part four of the paper.
Although there is little research on the use of financial data in RNN models, there are some promising attempts in the areas of bankruptcy prediction (Jang et al. 2021); in stock price prediction (Selvin et al. 2017; Nikou et al. 2019); and even in fundamental analysis using two financial ratios (Alberg and Lipton 2017).
The aim of this paper is twofold, on the one hand to create an advanced and comprehensive model for financial statement analysis and on the other to present and evaluate the efficiency of RNN models for decision makers. The study fills in the gap in the accounting literature, as suggested by Amani and Fadlalla (2017), by presenting a detailed description of the structure for the prediction model and its underlying technique demonstrating how such a model can be implemented; utilizing XBRL financial reporting data to its fullest (via imputation); and basing the model on a wide-ranging accounting data.
RNN is a deep learnings AI model, the most important advantage of RNN in terms of earnings prediction is the fact that one period's information remains in memory, in other words they are able to remember the past. XBRL data, recently mandated by the SEC has several important advantages: it is freely available, easily accessible, and timelier than traditional datasets such as Compustat. XBRL data is specifically suited for AI models since it provides means to convert the information from human-readable formats (e.g., paper, PDF, HTML) to a machine readable format (Chen et al. 2021). It is expected to increase in quantity (as more statements are filed) and in quality (Yen and Wang 2015; Scherr and Ditter 2017) over time thus refining the results of AI models. The model is comprehensive and, unlike many previous models, using AI techniques in financial statement analysis, is based on a large amount of relevant accounting financial ratios, representing information from all the financial statements.
Such a model enhances the current literature and enables a better understanding of the use of advanced techniques in financial statement analysis in general, and earnings prediction in particular.
The paper is organized as follows: the second section reviews academic research; the third section outlines the data used; the fourth section provides a comprehensive explanation for the RNN model and its implementation, the fifth section provides the results of the analysis and compares them to other methods (including traditional statistical methods); and the last section concludes the paper.