EPSO-GHSOM股票巨量資料選擇交易策略

Abstract

巨量資料分析（Big Data Analytics）是以資料觀點來進行分析研究、探討問題，找出未發現的知識與態樣。巨量資料分析包括三個部份，分别為資料存取運算、資料隱私與領域知識、巨量資料探勘。資料存取運算部份為處理巨量資料與分析的分散式平台與技術（technology），資料隱私及領域知識部份為解決問題的專業領域知識（domain）；巨量資料探勘部份則為分析巨量資料所引用的方法（technique）等。巨量資料分析的特點，在平台架構部份，使用分散式運算處理儲存架構，鬆綁了電腦資源的限制；而在分析部份，巨量的歷史執行資料常常蘊含著大量有價值的潛存資訊和知識，讓資料說話之新思維的解決問題方法，能夠忠實客觀呈現問題或事實真象，對於問題獲得解決及知識發現將帶來重要的助益。\n 因此，本研究提出一個基於巨量資料分析的觀點為主軸之創新的EPSO-GHSOM股票選擇交易策略。在資料存取運算部份，本研究採Hadoop分散式運算架構、HBase分散儲存資料庫、Elastic Search技術，以及撰寫資料分析應用程式，建構股票選擇交易策略分散式運算平台；在資料隱私及領域知識部份，則以基本分析之價值投資理論、變動天數移動平均線技術指標與其黃金死亡交叉決策準則，作為鑑別股票優劣與買賣交易點的領域知識探討核心；分析資料來源部份，以股票交易資料、公司財務報告重要資訊、網頁財經新聞訊息等各類來源資料作為分析標的；在巨量資料探勘部份，本研究提出改良PSO演算法之EPSO（Elite Particle Swarm Optimization）最佳化演算法，以及資料驅動點概念，並運用增長層級式自我組織映射圖（Growing Hierarchical Self-Organizing Map，簡稱GHSOM）演算法及網路探勘暨情緒分析（Web Mining with Sentiment Analysis）等方法，處理結構化與非結構化資料，作為資料探勘與知識發掘的分析核心，建構由來源資料自動探索並決定的股票選擇交易策略模型，從中進行知識挖掘，透過資料的角度發現股票選擇交易策略態樣、準則存在，提供以資料觀點的新方法給予投資人進行股票選擇交易決策建議。更明確地說，整體EPSO-GHSOM股票選擇交易策略，先經由股票選擇策略後挑選出優質的股票，再依股票交易策略決定最佳的買賣點提供給投資人進行決策。\n 本研究依所提方法進行實證結果發現，（1）在投資報酬表現部份，績效優於長期持有交易策略、MMPSO策略、KennedyPSO策略。（2）在股票選擇策略部份發現，屬量分析以稅後淨利、股東權益報酬率、每股盈餘等指標鑑別公司經營獲利能力最強，而且從5年財務指標趨勢發現有優質成長明星股（straight up）、經營不善之地雷股（straight down）及混合型股（U- or W-shaped）等三種型態，由於混合型股仍包含獲利能力好及表現不佳之公司，針對混合型股分群結果特徵分析後建立filter過濾股票規則，有效提升股票鑑別力。而在屬質之網路財經新聞資料情緒分析部份，發現實驗來源媒體報導多以每股盈餘、稅後淨利等財務指標作為評價依據，與屬量分析的分析結果相依性高，並考慮網路財經新聞資料情緒積分高低，篩選出情緒積分表現高之股票，更強化股票選擇策略能力。（3）在股票交易策略部份發現，多數的股票不符合移動平均線SMA技術指標的黃金死亡交叉決策準則，而且實驗所得最佳決策之SMA天數型態多屬中長期天數。另外，ROI績效受買賣高低價的影響，當價差愈大，ROI獲利或損失則愈大。（4）本研究與其他研究比較，測試期間ROI獲利表現較於其他方法相對偏高，整體策略考量涵蓋層面較其他方法廣泛。\n 由於Big Data Analytics屬於近年來新興發展學科，實際研究案例尚少，本研究所提出之整體研究方法、系統架構與建置步驟，除能作為股票選擇交易決策的參考外，並可套用至其他巨量資料分析研究案例進行建構，及以本研究為基礎發展理財智慧代理人等輔助決策模型。

Big data analytics is the process of analyzing data, examining problems, and identifying unknown correlations and patterns. It can be categorized into three parts: data accessing and computing, data privacy and domain knowledge, and big data mining algorithms. Data accessing and computing refer to decentralized platforms and technologies that handle big data and analysis and can be further sub-categorized into real-time and batch-processing platform/technology frameworks. Data privacy and domain knowledge refer to the specialized domain knowledge required to resolve problems. Big data mining algorithms refer to techniques used in analyzing big data. In terms of a platform framework, the application of decentralized processing and storage platforms in big data analytics alleviates restrictions on computer resources. In terms of analysis, historical big data often contain large amounts of valuable hidden information and knowledge. The novel problem-solving method of data narration provides a realistic and objective overview of situations and problems, which facilitates the resolution of problems and discovery of knowledge.\n Hence, this study proposed an innovative elite particle swarm optimization （EPSO）–growing hierarchical self-organizing map （GHSOM） stock selection and trading strategy that is based on big data analytics. In terms of data accessing and computing technology, this study used the Hadoop decentralized computing framework, Hbase decentralized storage database, Elasticsearch technology, and data analysis software to construct a decentralized computing platform for the stock selection and trading strategy. In terms of data privacy and domain knowledge, value investment theory, variable length moving averages （VLMA）, and golden cross and death cross decision rules were adopted as a basis for analysis to investigate the core domain knowledge of distinguishing between the pros and cons of shares and share trading. Data from the stock market, company financial reports, financial news from websites, and other data were collected as sources for the analysis. In terms of big data mining algorithms, this study proposed an improved particle swarming optimization （PSO）method called EPSO and used the concept of data-driven points, GHSOM, and web data mining with sentiment analysis to process structured and non-structured data, form core data mining and knowledge discovery data for analysis, and establish a model that automatically explores source information exploration and selects decision-making strategies. The model uncovers knowledge from data and identifies stock selection and trading strategies, patterns, and rules from a data perspective. This provides investors with a novel data analysis method and facilitates them in making decisions regarding the trading of stocks. Specifically, the EPSO–GHSOM stock selection and trading strategy first selects quality stocks using the stock selection strategy, and then, the stock trading strategy decides the optimal buy and sell points, providing investors with information for making decisions. \n Using our proposed method, we found that（1）in terms of investment return performance, the results are superior to the investment return rates of buy-and-hold, MMPSO, and Kennedy PSO strategies.（2）In terms of stock selection strategy of quantitative analysis, we found the net profit, ROE, and EPS demonstrated the strongest ability to distinguish among company profitability indicators. The financial index trend over five years also presented straight up, straight down, and U- or W-shaped patterns. Because U- or W-shaped trends contain companies of both good and suboptimal profitability indicators, the application of a filter for U- or W-shaped clustering characteristic analysis effectively increases the ability to distinguish among stocks. Using sentiment analysis on the qualitative online financial news information, we found that the media reports mainly used ROE, EPS, and other similar financial indicators as tools for evaluation, similar to the quantitative analysis results.（3）In terms of the stock trading strategy, most stocks failed to follow the golden cross and death cross decision rules of the SMA. Optimal SMA parameter days for decision making were largely mid-term to long term. Furthermore, ROI results were affected by high and low trading prices, increased price margins, and increased profit–loss margins.（4）The ROI performance was relatively better when compared with previous studies. In terms of domain knowledge analysis methods, considerations for the strategy cover a wider range of areas than other methods. \n In recent years, big data analytics has been an emerging science, and thus, practical study cases are scarce. The study method, system framework, and procedures proposed in this study can be applied to other areas of big data analysis in addition to being a reference for stock selection and trading strategy. The results obtained in this study can also form the basis for decision assistance models, such as an intelligent financial management agent.