Applying artificial neural networks to portfolio selection: Empirical study in Taiwan stock market

The problem. The purpose of this study is (a) to investigate the application of neural networks on stock predictions which were performed by using financial ratios and behavioral finance proxies as indicators, (b) to examine whether the predicted returns closely reflected the actual returns in different industries and market states of Taiwan's security market from 1999 through 2008, and (c) to identify which indicators were most effective in predicting returns from investing strategies created through the neural networks technique. Thus, this study will extract useful information for individual and institutional investors in the Taiwan stock market.;Method. In this study, the research methodology was twofold. First, a nonlinear model (neural networks) was used to predict stock returns from the set of all stocks listed on the Taiwan Stock Exchange (TSE). This model also evaluated the performance of these predictions relative to the market returns measured by the Sharpe Ratio for risk-adjusted returns (Sharpe, 1966). Second, two statistical calculations were conducted using SPSS software, namely the Pearson correlation test and the dependent samples (paired samples) t-test. All statistics were performed at the .05 level of significance (alpha set to .05).;The Pearson correlation tested the relationship between stock returns and each of the nine financial variables: market capitalization, dividend yield, P/S ratios, P/B ratios, price-to-cash flow ratios, short-term rate of return, long-term rate of return, turnover rate and earning to price ratios. The dependent samples (paired samples) t-test investigated the differences between predicted stock returns (created through the neural networks by using financial ratios and behavioral finance proxies) and actual stock returns, and compared the mean of monthly predicted stock returns with the mean of monthly actual returns within different industries (Electronics/Traditional) and market states (Bull/Bear). In addition, the descriptive statistics of variables were used to compute means, variances, and standard deviations.;Data analysis proceeded as follows: First, the author presented the descriptive statistics. Second, the author applied a Pearson correlation to measure the relationships among stock returns, financial ratios, and proxies of behavioral finance. Third, the author performed a dependent samples (paired samples) t-test to investigate whether the monthly predicted stock returns created through the neural networks were different than actual returns on the market states (Bull/Bear) and on the industries (Electronics/Traditional).;Result. The results showed that all nine factors except the price-to-cash flow ratio related significantly with stock returns and helped explain average stock returns in the Taiwan stock market during the 10 year testing period (1999--2008). Financial ratios (market capitalization, dividend yield, P/S ratio, and P/B ratio) and behavioral finance proxies (short-term rate of return, long-term rate of return, turnover rate, and E/P ratio) proved to be important determinants of stock returns. The paired samples t-test results indicated that the predicted stock returns based on fundamental analysis approximated actual returns in the traditional industry, and that predicted stock returns based on behavioral finance analysis approximated real returns in the electronics industry. Therefore, it is recommended that investors use fundamental analysis to predict stock returns of the traditional industry, and behavioral finance analysis to predict stock returns of the electronics industry.