類神經網路於美國SPDRs價格與趨勢預測模式之應用

The Application of Neural Networks in SPDRs Price and Trend Forecasting

王秋萍(Chiu-Ping Wang)

指導教授 ： 張淳智 ； 顏憶茹

國立臺中科技大學/商管學群/事業經營研究所/碩士(2008年)

摘要

目前，全球ETF市場發展速度相當驚人，美國不論是掛牌數或資產規模均大於其他各國，且不論是資產規模、發行量與成交量均呈逐年成長的趨勢，其中，第一檔ETF－SPDRs自1993年推出後，發展時間已長達14年，資產規模最大且流動性高。故本研究採用類神經網路結合技術面與經濟面因素，架構ETF價格與趨勢預測模式，並藉以預測SPDRs隔日收盤指數與長期趨勢，最後，亦透過結合濾嘴法則之交易策略運用預測結果進行模擬且計算其獲利程度。
實證顯示，以倒傳遞網路架構之ETF價格預測模式，上漲∕下跌的預測準確率為85.56%，而ETF趨勢預測模式中之多頭、空頭與振盪趨勢的預測準確率則為89.46%。此外，本研究也結合ETF價格預測模式與趨勢預測模式進行以濾嘴法則為基礎之交易策略模擬，而最佳之交易策略乃是結合趨勢預測模式之交易策略，其不僅優於買進持有策略與無風險利率之報酬，亦優於技術性分析之買賣策略投資績效，顯示出運用趨勢預測模式進行之長期持有策略將優於短期買賣之操作策略。

關鍵字

指數股票型基金 ； 類神經網路 ； 股價預測

並列摘要

The global Exchange traded funds (ETF) market development at a surprising speed now, no matter is the numbers or assets of ETF, the United States is greater than other countries, its assets, numbers and trading volumes is growing up year by year. As the United States first ETF - SPDRs public offering in 1993, development time has been for 14 years, the assets are largest and fluidity is high, so the study use neural networks combine technical indexes and economical indexes to build the ETF price forecasting model and trend forecasting model, and use the models to forecast SPDRs closed price and long-term trend of next day. The study is also using the predict results and using the trading strategies which based on filter rules to simulate and calculate its profit.
The results show that ETF price forecasting model which use back-propagation network for the up/down forecast accuracy is 85.56%, and the ETF trend forecasting model which use back-propagation network for the bullish/bear/oscillatory trend forecast accuracy is 89.46%. Additionally, the study also combine the ETF price forecasting model and ETF trend forecasting model to develop filter rules based trading strategies. The highest return on investment is the trading strategy that using the trend forecasting model, it’s not only greater than the return of buy and hold strategy and interest rate, but also greater than the strategy of traditional technical analysis. The results show that investment performance of the long-term hold strategy which by ETF trend forecasting model is greater than the operation of short-term trading strategy.

並列關鍵字

Exchange traded funds ； Neural networks ； Stock prediction

參考文獻

黃麒元（民94），全球主要指數股票型基金之外溢效果與績效之研究，中原大學企業管理研究所碩士論文。

戴淑瑩（民95），臺灣50指數ETF整合型預測之研究，成功大學統計學系碩士論文。

陳執中（民95），台股加權指數隔月收盤價預測之研究，成功大學統計學系碩士論文。

蔡晨瑩（民94），臺灣50指數ETF整合型分類預測之研究，成功大學統計學系碩士論文。

陳國玄（民93），人工?經網?與統計方法應用於台灣上市電子?股價指?預測與分?之研究，成功大學統計學研究所碩士?文。