تکنیک های شبکه عصبی برای پیش بینی عملکرد مالی: یکپارچه سازی تجزیه و تحلیل بنیادی و فنی

This research project investigates the ability of neural networks, specifically, the backpropagation algorithm, to integrate fundamental and technical analysis for financial performance prediction. The predictor attributes include 16 financial statement variables and 11 macroeconomic variables. The rate of return on common shareholders' equity is used as the to-be-predicted variable. Financial data of 364 S&P companies are extracted from the CompuStat database, and macroeconomic variables are extracted from the Citibase database for the study period of 1985–1995. Used as predictors in Experiments 1, 2, and 3 are the 1 year's, the 2 years', and the 3 years' financial data, respectively. Experiment 4 has 3 years' financial data and macroeconomic data as predictors. Moreover, in order to compensate for data noise and parameter misspecification as well as to reveal prediction logic and procedure, we apply a rule extraction technique to convert the connection weights from trained neural networks to symbolic classification rules. The performance of neural networks is compared with the average return from the top one-third returns in the market (maximum benchmark) that approximates the return from perfect information as well as with the overall market average return (minimum benchmark) that approximates the return from highly diversified portfolios. Paired t tests are carried out to calculate the statistical significance of mean differences. Experimental results indicate that neural networks using 1 year's or multiple years' financial data consistently and significantly outperform the minimum benchmark, but not the maximum benchmark. As for neural networks with both financial and macroeconomic predictors, they do not outperform the minimum or maximum benchmark in this study. The experimental results also show that the average return of 0.25398 from extracted rules is the only compatible result to the maximum benchmark of 0.2786. Consequentially, we demonstrate rule extraction as a postprocessing technique for improving prediction accuracy and for explaining the prediction logic to financial decision makers.

Neural networks have become a popular tool for financial decision making [2], [5], [16], [17], [18], [25], [33], [35], [36], [41] and [43]. There are mixed research results concerning the ability of neural networks to predict financial performance. Due to a variety of research design and evaluation criteria, it is difficult to compare the results of different studies [9] and [27]. Past studies in this area are subject to several problems. First, time horizons for experiments are short. When time horizons are short, experimental results may be tampered by situational effect and economic fluctuations. Second, sample sizes are small. When sample sizes are small, experimental results may be biased and cannot be generalized to the future. Third, many studies do not investigate the statistical significance of performance differences. Because variance is a significant factor in investment, ignoring performance variance in the forecasting process is undesirable at least [3] and [6]. Fourth, the selection of predictor attributes in past studies is based on either fundamental or technical analysis. Fundamental analysts believe that an investment instrument has its intrinsic value that can be derived from the behavior and performance of its company [7], [8], [14], [22], [28], [29], [34], [40] and [42]. The fundamental approach utilizes quantitative tools, mainly the financial ratios compiled from financial statements as well as qualitative indicators, such as management policy, marketing strategy, and product innovation, to determine the value of an investment instrument. Technical analysts, on the other hand, believe that the trends and patterns of an investment instrument's price, volume, breadth, and trading activities reflect most of the relevant market information a decision maker can utilize to determine its value [20], [24] and [37]. Instead of analyzing fundamental information about companies, the technical approach tries to identify turning points, momentum, levels, and directions of an investment instrument, using tools such as charting, relative strength index, moving averages, on balance volume, momentum and rate of change, breadth advance decline indicator, directional movement indicator, and detrended price oscillator. There are divergent opinions about what other trends in the macroeconomic, political, monetary, and societal sentiment spheres should be incorporated in technical analysis [4]. It is the purpose of this research project to address the above four problems for neural network as a data mining tool for financial forecasting. This project applies neural networks to a sample of 364 S&P companies for the period of 1985–1995. We attempt to present a formal study on the complex phenomenon of financial performance using company financial and macroeconomic data as predictor variables, neural networks as the data mining tool, and rate of return on common shareholders' equity as the to-be-predicted variable. Paired t tests are adopted to verify the statistical significance of performance differences between neural networks and the market's top performers as well as overall averages. We believe that neural networks are an excellent tool for forecasting financial performance for the following reasons. First, neural networks are numeric in nature, which is especially suitable for processing numeric data such as financial information and economic indicators. The numeric nature of neural networks is in contrast to symbolic manipulation techniques such as ID3 [26] and AQ [21], which were designed for processing nominal variables. Because numeric data must be converted into nominal values before they can be used as input to symbolic manipulation techniques, there are the problems of losing information, inappropriate data intervals, and different conversion methods leading to different mining results. Neural networks, on the other hand, can accept numeric data directly as input for mining purposes. Second, neural networks do not require any data distribution assumptions for input data. This feature allows neural networks be applicable to a wider collection of problems than statistical techniques such as regression or discriminant analysis. Third, neural networks are an incremental mining technique that permits new data be submitted to a trained neural network in order to update the previous training result. In contrast, many symbolic manipulation and statistical techniques are batch-oriented, which must have both new and old data submitted as a single batch to the model, in order to generate new mining results. For financial applications with new data being available from time to time, neural networks can accommodate new information without reprocessing old information. Fourth, neural networks are model-free estimators. This feature allows interaction effect among variables be captured without explicit model formulations from users [13] and [45]. Basically, the more hidden layers in a neural network, the more complicated the interaction effect can be modeled. Although neural networks as a data mining tool have the above merits, they have their fair share of problems. One common difficulty for neural network applications involves the determination of the optimal combination of training parameters including the network architecture (the number of hidden layers and the number of hidden nodes), the learning rate, the momentum rate, the order of submitting training examples to the network, and the number of training epochs. There are various heuristic rules and common practices for selecting the parameters [44], but the selection process remains as an art rather than a science, and varies from problem to problem. Another common problem in financial applications is noisy data. Because data are collected empirically from different sources, they are subject to corruptions during the retrieval, encoding, transfer, and decoding process. Financial frauds are also potential sources for noisy data in the corporate world. In this research, we take a different approach in addressing the parameter selection and noisy data problem for neural network applications. Instead of addressing the problems before the data mining process, we address the problems after the data mining process. Data preprocessing is a common step for data mining, which can be used to reduce data noise and other irregularities in data sets. As for the parameter selection problem, the traditional approach is to try different parameter combinations on a subset of the available data, with the objective of identifying a satisfactory set of parameters for the entire mining process. Because an exhaustive test of all parameter combinations is impractical, there is no guarantee for an optimal solution. Instead of trying to identify a satisfactory set of parameters before the mining process, we adopt the technique of extracting rules from trained networks after the mining process. Rule extraction, as a postprocessing technique, has the potential ability to generate a more precise and accurate mining result by reducing redundant, conflicting, and erroneous information due to noisy data, input selection problems, and parameter misspecification. The remaining of this paper is organized as follows. Section 2 reviews the relevant research in the literature. Section 3 presents the experimental design for this study. The experimental results are described in Section 4 and are discussed in Section 5. The last section concludes the paper by summarizing the findings and suggesting some research directions in this area.

In this study, we designed a series of experiments to test the prediction power of neural networks on financial performance. The experimental results support the ability of neural network to significantly outperform the minimum benchmark based on a highly diversified investment strategy. In addition, the technique of incorporating previous years' financial data in the input vector for neural network training can significantly increase the return level, which demonstrates the benefits of integrating fundamental analysis with technical analysis via neural network training. Moreover, the integration works especially well when the economy is in recession. As for the prediction efficacy of macroeconomic variables for financial performance, we did not find any confirmative evidence in this problem setting. The experimental results show that financial statement variables and macroeconomic variables together cannot generate significantly higher returns than the average index. We also extracted rules from trained neural networks and found that the extracted rules can predict significantly better than neural networks per se, can perform as accurately as the maximum benchmark, and can reveal the prediction logic to financial decision makers. Among all the research studies on the abilities of neural networks to select high-return companies, this research provides some strong evidence on the benefits of neural networks. Will the promise hold up in the coming 10 years? That remains to be confirmed by another research study. Another research direction in this area is to construct or manipulate the neural network to emphasize the relative importance of learning the positive cases correctly. One method can be changing the error function or the learning procedure to provide positive reinforcement for classifying the positive cases correctly. This is basically the procedure of recognizing differential costs for making different mistakes. As the focus of the problem is on selecting high-return companies, the relative cost of misclassifying a low-return company as high return can be more damaging than committing the reverse error.