整合模糊應用與圖形識別方法於台灣股價趨勢行為之發現

Abstract

台灣股市成立至今，累積了大量的時間性股市資料，在龐大資料庫中隱含著許多成功投資者的投資模式。股市交易是買賣雙方投資行為的結果，其受影響之因素有很多，例如政治、總體經濟、國際情勢、社會消息、成交量等因素均會影響投資行為。股價是一切因素的結果，因此，股價走勢圖可視為最具代表性之投資者行為模式。 獨立成份分析(ICA)是一混合訊號分離的技術，可以由混合訊號中找出獨立元素或是潛在行為。而股價就如同混合訊號，是許多因素的綜合結果，因此本研究先利用獨立成份分析方法，發掘隱藏在股價之內的獨立成份，雖然無法了解每個成份本身代表何種因素，但是利用此方法，去除解釋股價能力較低之元素，使得股價的干擾減少，將有利於發現有意義有價值之股價走勢。 圖形識別(Pattern Recognition)是近年來熱門的研究領域之一，其根據物體或是圖片的顏色、形狀等屬性、特徵來做辨認，通常是由已知的圖樣去找尋最相似的圖形，此方法已大量的應用在指紋比對、人臉辨識、語音辨識、簽名認證、數位浮水印等，本論文則企圖將圖形比對之方法應用於股市時間序列。 本實驗由證券基金會(SFI)之資料庫收集自1996年11月26日至2004年4月28日台灣股市個股收盤股價日資料，利用圖形識別方法，針對當時個股股價的走勢，比對歷史股價中重覆出現的圖樣。本研究試圖提出一個新的比對方法，以減少比對的時間：先針對一段待測股價時間序列圖進行特徵抽取(Feature Extraction)，找出人類視覺上最重要的特徵點(Feature Points)，透過特徵點間的相關資訊進行比對工作，包括了圖形的角度以及相對長度。 實驗結果找出了歷史中重覆出現的圖樣，也就是相似的走勢，證明了股價走勢有歷史重現的現象，據此可針對相似走勢的未來趨勢來作預測分析。實驗結果證明所提圖形比對方法之平均獲利較買進持有策略、最小平方法求得的趨勢線來的好，可作為投資人投資時的決策參考。

There have been a large amount of stock data and valuable trading strategies stored in Taiwan stock market database. The stock price, impacted by various factors, is the result of buyer-seller investment decisions. The influencing elements consist of politics, economics, international statuses, and news information all around the world. Since the stock price is capable of reflecting the value-relevant information in the market, the stock patterns can be regarded as the behaviors of investors in the environment. Independent component analysis (ICA) is a technique for separating the mixture signals to find out the independent components or hidden factors. In this thesis, since stock price is a mixture signal, ICA can be used to extract the unknown independent components. In this way, by eliminating the component with lower effect is able to reduce the unnecessary noise and discover the significant trend patterns. Pattern recognition is one of the most popular research areas in recent years. The approach is used to identify the objects or images by their forms, outlines, colors or other attributes. The general purpose of pattern recognition is to find out the most similar patterns. The major applications of pattern recognition contain signature certification, digital watermarking, fingerprint, face and speech recognition, etc. In this research, pattern recognition is applied to finance time series mining for discovering the concealed knowledge in the stock market. In the thesis, the experimental data is collected from Securities & Futures Institute (SFI) on daily closing price of Taiwan Stock Exchange Weighted Stock Index (TAIEX). The simulation time period is from November 26, 1996 to April 28, 2004 for 1,942 trading days. The pattern recognition concept is used to match the current stock price trend with the historical repeatedly appearing patterns. Accordingly, a new method is developed in the thesis to extract the features from the stock chart for discovering the most critical feature points. The matching procedure is processed through the corresponding information of the feature points, which is capable of reducing the experimental time. After the simulation, the historical repeatedly appearing patterns, namely the similar trend, will be discovered. The results show that the stock trends are continually occurring. Consequently, the future trend of matching patterns will be analyzed to offer investors the information for making strategies. Besides, the average profit of proposed model is higher than buy-and-hold strategy and the least square line.