以類神經網路研究證券成長風格與價值風格之分類、辨識

行政院國家科學委員會補助專題研究計畫成果報告

※※※※※※※※※※※※※※※※※※※※※※※※※

※

以類神經網路研究證券成長風格與價值風格之分類、辨識

※

※

※

※

※

※※※※※※※※※※※※※※※※※※※※※※※※※

計畫類別：□個別型計畫　　□整合型計畫

計畫編號：NSC 89-2416-H-004-036

執行期間：88 年 08 月 01 日至 89 年 07 月 31 日

計畫主持人：蔡瑞煌

共同主持人：

本成果報告包括以下應繳交之附件：

□赴國外出差或研習心得報告一份

□赴大陸地區出差或研習心得報告一份

□出席國際學術會議心得報告及發表之論文各一份

□國際合作研究計畫國外研究報告書一份

執行單位：國立政治大學資訊管理學系

中　華　民　國　89 年　8 月　16 日

行政院國家科學委員會專題研究計畫成果報告

Equity Style Classification, Identification and Investing Str ategy with Ar tificial Neur al Networ ks and Discr iminant Analysis

計畫編號：NSC 89-2416-H-004-036 執行期限：88 年 08 月 01 日至 89 年 07 月 31 日 主持人：蔡瑞煌 國立政治大學資訊管理學系 計畫參與人員：李如琪 國立政治大學資訊管理學系 一、中文摘要 本研究應用類神經網路在股票風格投 資方面之分類、辨識。類神經網路在樣本 內與樣本外的分類正確率皆優於區別分析 ﹐而且類神經網路在樣本內的訓練範例中 達成了百分之百的分類正確率。此外﹐我 們也解決了傳統方法無法展示股票風格動 態的問題。檢視各種風格投資策略在台灣 股票市場的績效表現之後﹐我們以類神經 網路為基礎﹐提出一個簡單而容易實行的 投資策略。由這個策略的表現可以說明﹐ 即使在考慮了風險因素之後﹐積極的風格 投資策略的確可以增加投資組合的績效表 現。 關鍵詞：類神經網路; 風格投資分析 Abstr act

This paper investigates whether or not the classifications of stock styles can enhance profit. A newly developed method, which is called RNBP (reasoning neural network with backward propagation), is employed in this paper to classify the Taiwan stocks into growth and value stocks. RNBP, though stemming from artificial neural network, has improved the weaknesses of conventional artificial neural networks substantially. To make a comparison, the conventional style classification method, the discriminant analy-sis (DA), is also examined. Classification accuracy is first compared, and then the style investment strategies are performed. Our re-sults show that RNBP outperforms DA in both in-sample and out-sample classification accuracies. The style investment strategies based on RNBP are also significantly superi-or to those of DA.

Keywor ds: Artificial Neural Networks, Style analysis, Growth stock, Value stock

二、緣由與目的

The classification of the style of stocks has recently received a great attention. Stocks with similar characteristics may tend to per-form as a group over several economic and business cycles, and constitute equity market segments (Berstein, 1995). Two typical styles, the value style and the growth style, are often discussed in the literature. Value style stocks are stocks with lower prices/earning (P/E) ratio, or price/book (P/B) ratio, whereas growth style stocks are those have above-average growth prospect. Sharpe (1992), who had analyzed almost 400 mutual equity funds, found that 90 to 95 percent of their perfor-mances could be attributed to the style allo-cation. Jacobs and Levy (1996) found that accurate style allocations yielded superior realized returns.

Although the choice of a portfolio style is considered as an important step in the in-vestment decision making process, the methods of classifying the styles remain vari-able. Gallo & Lockwood (1997) mentioned that there were three approaches to style clas-sification for actively managed portfolios. The first approach was to classify the style on the basis of interview with professionals. The second was the return-based approach, which computed the correlation between the return of the mutual funds and the returns of a num-ber of selected indexes. The third was the characteristic-based approach, where the manager assigns the style for stocks based on the security information including factors such as P/E, P/B and dividend payout ratios (Ramaswami, 1994).

Though the first approach had it own merits, it had been criticized to be subjective. This subjective assessment has made the classification more art than science. With

respect to the second approach, Christopher-son (1995) argued that the computed correla-tion might lead to misclassificacorrela-tion because historical correlations were noisy forecasts of the future correlation. Furthermore, the rigid classification, which separated the stocks into either growth or value stocks, ignores the intermediate cases. The third approach util-ized the discriminant analysis method (DA), which overcame the above defects by consid-ering the time-varying weight of each stock. For instance, a stock is a mixture of the two styles with a weight of 80% to the first style and a weight of 20% to the second style. Ac-cordingly, the third approach may be more appropriate than the first two in classifying the stocks.

While the characteristic approach ap-pears better than the other two, the use of DA in separating the stocks also has one weak-ness. DA is a linear classification method, which separates the stocks by a linear regres-sion line. While this linear separation is easy to apply, it serves only as an approximation to the complexity of the real world. For ex-ample, if stock styles indeed exist, but can be separated only on the basis of an S-shaped plane, then the linear separating line may result in misclassification. Hence, a nonlinear separation method, which contains the spirit of DA on one hand and can accommodate the nonlinear relations between variables on the other, should be the most appropriate.

Artificial Neural Networks (ANN) is an ideal tool for this nonlinear separation re-quirement. The merit of ANN is that it is data driven and model free1. In addition, it has

1

In the context of traditional statistical methods, ANN can be considered as a multivariate nonlinear non-parametric inference technique that is data driven and model free. Multivariate implies that the ANN inputs comprise many different variables whose inter-dependencies and causative influences are exploited. Nonparametric, model free, means there are no pre-sumptions regarding the relation between input and

powerful pattern classification capabilities, surpassing other techniques in many applica-tions (Altman, 1994; Brockett, 1994; Trippi & Turban, 1996). Furthermore, the ANN model is not subject to DA's constraining assumptions, such as linear separability and independence of the classifying variables.

One of the most popular ANN is the layered feedforward network with the back propagation learning algorithm (BP) (Rumel-hart, Hinton and Williams, 1986). However, BP contains some undesirable predicaments, including the proper number of hidden nodes being unknown, the relatively optimal learn-ing results and the sluggish learnlearn-ing process (Tsaih, 1993). To remedy these shortcomings, various modifications are proposed; however, a generalized solution has not been found. Tsaih (1997, 1998) has recently developed a reasoning neutral network (RN), which adopts a learning procedure that ensures an optimal solution. The merit of RN is that it can deal with not only conventional binary output patterns but also non-binary output ones. However, computing time of RN is nontrivial owing to the increasing “nodes” it creates. Combining the merits of BP and RN to overcome their weaknesses, Tsaih, Chen and Lin (1998) presented reasoning neutral networks with back propagation learning algorithm (hereafter, RNBP). From the evaluation of the training and testing samples, RNBP outperforms BP. Accordingly, using RNBP to classify the stock styles seems the ideal step to pursue.

The purpose of this paper is to use arti-ficial neural network to re-assess the equity style investment strategies. To the best of our knowledge, this is the first study of employ-ing artificial neutral network to classify and identify the equity styles. We examine whether or not the RNBP can outperform the DA method, which serves as our benchmark. Taiwan stock market data from 1987:Q1 to 1997:Q3 are used for illustration. Since the

output variables. Data driven implies that the weights of ANN are estimated from the (given) training data.

use of RNBP and DA requires the sample to be split into training samples and evaluation samples, which are close to the conventional in-samples and out-samples, respectively, we first separate the stock data into these two types of samples using date of 1996:Q1. The training samples are the data used to build up the “relation pattern” between the styles and input variables2. Once this relation pattern is built up (learned), it is used to clas-sify the evaluation samples. The classifica-tion performance of both training and evaluation samples are compared.

三、結果與討論

Our results are fruitful. First, with re-spect to the classification accuracy, the RNBP achieves the 100% classification accu-racy for the in-sample evaluation and 97% accuracy in out-sample data. Both ratios are significantly higher than those of DA. It seems that if stock styles indeed exist, it shall be separated on the basis of a non-linear separation surface. The proposed style in-vestment strategy is then implemented by constructing portfolios based on a style dy-namics, which is the style changing pattern. Portfolio returns are computed. Results show that the style strategies based on RNBP en-hance profit regardless of risk adjustment, whereas results do not lend support to style investment strategies based on DA. Although this study uses data from Taiwan, the pro-posed model can be applied to other coun-tries immediately. 四、計畫成果自評 就目前所知﹐應用類神經網路在股票 風格投資方面的研究不多。本研究是少數 之一。研究內容與原計畫相符程度很高﹐ 而研究成果的學術價值亦有。不過﹐離預 期目標仍有一段距離。會嘗試在學術期刊 發表。 2

The input variables are P/E, P/B, P/S, and SGR.

五、參考文獻

Altman, E., "Corporate Distress Diagnosis: Comparisons Using Linear Discriminant Analysis and Neural Networks (the Italian experience)." Journal of Banking & Fi-nance, 1994 May, pp. 505-529.

Bauman, W. & Miller, R., "Investor Expecta-tions and the Performance of Value Stocks versus Growth Stocks." Journal of Portfolio Management, 1997 spring, pp.57-68.

Balch, Hardy, Scheinman & Winston, "Asset Allocation Is Not the Most Important Deci-sion You Will Ever Make." Seventh Asset Allocation Conference for the Institute for International Research, February 22, 1993. Bernstein, R., Style Investing-Unique Insight

into Equity Management. John Wiley & Sons, Inc, 1995.

Brockett, P., "A Neural Network Method for Obtaining An Early Warning of Insurer In-solvency." Journal of Risk & Insurance, Sept. 1994, pp. 402-424.

Case & Cusimano, "Historical Tendencies of Equity Style Returns and the Prospects for Tactical Style Allocation." Equity Style Management, IRWIN, 1995.

Chen Y., Thomas, D. & Nixson, M., "Gener-ating-Shrinking Algorithm for Learning Arbitrary Classification." Neural Networks, 1994, 7, pp. 1477-89.

Chiou, Y., Liu, S. & Tsaih,R., "Applying Reasoning Neural Networks to the Analysis and Forecast of Taiwan’s Stock Index Variation." Taipei Economic Inquiry, 1996, Vol. 34, No. 2, pp. 171-200.

Christopherson, J., "Equity Style Classifica-tions." Journal of Portfolio Management, 1995 spring, pp.33-43.

Damodaran, A., Damodaran on valuation : security analysis for investment and corpo-rate finance, New York : Wiley, 1994. David, M. & Ramesh, S., "Analyzing

Mathematical Models with Inductive Learning Networks." European Journal of Operation Research, 1995, pp. 387-401. Fahlman, S. & Lebiere, C., "The

To-uretzky, D. (Eds.), Advances in Neural In-formation Processing Systems II (Denver, 1989), San Mateo: Morgan Kaufmann. Fama, E. & French, K., "The Cross-Section

of Expected Stock Returns." The Journal of Finance, Vol. 47, No. 2, June, 1992, pp.427-465

Frean, M., "The Upstart Algorithm: A Method for Constructing and Training Feedforward Neural Networks." Neural Computation 2, 1990, pp. 198-209.

Gallo, J. & Lockwood, L., "Benefit of Proper Style Classification of Equity Portfolio Managers." Journal of Portfolio Manage-ment, 1997 spring, pp. 47-55

Jacobs, B. & Levy, K., "High-Definition Style Rotation." The Journal of Investing, 1996 fall, pp.14-23.

Karels, G. & Prakash, A., "Multivariate Normality and Forecasting of Business Bankruptcy." Journal of Business Finance and Accounting, 1987 winter, pp. 573-93. Lin, H., Tsaih, R., & Jee, R., "Exploring the

Relative Abilities of Neural Networks and VAR Models in Forecasting Taiwan Bond Prices." Review of Securities and Futures Markets, 1997, Vol. 9, NO. 1, pp. 63-113, Taipei.

Mezard, M & Nadal, J., "Learning in Feed-forward Layered Networks: The Tiling Al-gorithm." Journal of Physics A 22, 1989, pp. 2191-2204.

Markham, I. and Ragsdale, C,. "Combining Neural Networks and Statistical Prediction Problem in Discriminant Analysis." Deci-sion Sciences, Vol 26, No.2, 1995 Mar/Apr, pp. 229-42.

Naimimohasses, Barnett, Green, & Smith, "Sensor optimization using neural network sensitivity measures." Measurement Sci-ence & Technology, 1995, Sep, pp. 1291-1300.

Ragsdale C. & Stam, A., "Introducing Dis-criminant Analysis to the Business Statis-tics Curriculam, Decision Sciences, Febru-ary 1992, pp. 724-45.

Ramaswami, M., "Return Enhancement through Size and Style Management." Blending Quantitative and Traditional

Eq-uity Analysis, 1994, pp. 109-17.

Sarkar, D., “Methods to speed up error back-propagation learning algorithm,” ACM Computer Surveys, 1995, Vol. 27, No. 4, Dec., pp. 519-542.

Sharpe, W., "Asset Allocation: Management Style and Performance Measurement." Journal of Portfolio Management, 1992, winter, pp.7-19

Trippi, R. & Turban, E., Neural Networks in Finance and Investing, IRWIN, 1996. Tsaih, R., "The Softening Learning

Proce-dure." Mathematical and Computer Model-ling, 1993, Vol. 18, No. 8, pp. 61-64.

Tsaih, R., "The Softening Learning Procedure for The Networks With Multiple Output Nodes." MIS Review, 1994, vol. 4, pp. 89-93, Taipei.

Tsaih, R., "Learning Procedure that Guaran-tees Obtaining the Desired Solution of the 2 Classes Categorization Learning Prob-lem." The 1st Asia-Pacific Conference on Simulated Evolution and Learning, Korea, 1996, pp. 446-53.

Tsaih, R., "The Reasoning Neural Networks." In Ellacott S., J. Mason & I. Anderson (Eds.), Mathematics of Neural Networks: Models, Algorithms and Applications, 1997, pp. 366-371.

Tsaih, R., “An Explanation of Reasoning Network Networks.” Mathematical and Computer Modelling, 1998, Vol. 28, No. 2, pp. 37-44.

Tsaih, R., Chen, W. & Lin, Y., "Application of Reasoning Neural Networks to Financial Swaps." Journal of Computational Intelli-gence in Finance, Vol. 6, No. 3, 1998, pp. 27-37.

Tsaih, R., Hsu, Y. & Lai, C., "Forecasting S&P 500 Stock Index Futures with the Hy-brid AI system." Decision Support Systems, 1998, Vol. 23, No. 2, pp. 161-174.

Yoon, Y., Swales, G., & Margavio, T., "A Comparison of Discriminant Analysis ver-sus Artificial Neural Networks." Journal of the Operation Research Society, 1993, 44(1), pp. 51-60.