本篇論文主要探討股市投資問題中的因果關係做為本研究的實驗對象，著重於探討如何提昇投資績效，而欲提昇投資績效，則須瞭解影響績效之因素及績效觀察值間的因果關係，我們將利用資料挖掘技術之關聯規則方法，有利於尋找影響績效之技術指標及績效觀察值(ex.股價反轉點)間的因果關係之規則，本研究稱之為因果關聯規則（Causal Association Rule），我們可將這些規則，組合成證券交易策略。 過去有諸多學者提出了許多關聯規則方法，然而這些傳統資料挖掘之關聯規則方法，均會產生大量的高頻項目集（Large Itemset），以致產生的規則太多、不易評估有趣性且較沒有效率，因此本研究提出一個CFP演算法架構，其中主要是改良FP-Growth演算法，以減少產生不必要之高頻項目集，使能更有效率地產生有趣的因果關聯規則。 現今常見資料離散化的處理方法，分為等距劃分法及等量劃分法兩種。然而一般投資者在進行股票進出場買賣操作時，所參考的股市技術指標數值都是累計值。本研究提出等距累計劃分法及等量累計劃分法之資料離散化概念，所離散化的技術指標適合投資者股票進出場買賣操作，同時採用累計之概念，可挖掘跨階層(level-crossing)之因果關聯規則，以挖掘更多可能有趣的規則。經過實驗t檢定結果，本研究之演算法能有效率挖掘因果關聯規則，在效率上的確有不錯的表現，顯著優於傳統之FP-growth方法。本研究在實驗中亦發現本研究之演算法隨著資料量的增加，效率更加顯著，因此適合挖掘較大型資料庫。 將挖掘之影響投資績效之因果關聯規則，依影響績效之不同構面排序進行分析，藉以提供投資者進行投資策略之安排上的協助，並藉由發掘技術指標與特定投市投資問題之關聯規則，提供投資者避險之參考。 This thesis mainly probes into the causality among the investment problems of the stock market to do for the experimental subject of this research. We focus on discussing how about to promote the performance of investment. If we want to promote the performance of investment, we must understand the causality among the factor which influences the performance and performance observing value. we will utilize the method of association rule of data mining to help to look for association rules about causality among the technological indicators which influences the performance and performance observing value (ex. the reversal point of the stock price). We call these rules as Causal Association Rules. We can make these rules up into the tactics of securities trading. In the past, many scholars proposed a lot of methods of association rules, but these methods will produce a large number of large itemsets. So that there are too many rules and it is difficult to assess the interesting of rules and relatively inefficient. So we propose a CFP algorithm structure which mainly improve FP-Growth algorithm to reduce mining the unnecessary large itemsets and enable only producing the interesting causal association rules efficiently. The common data dispersed methods now have equal width interval and equal frequency interval. But when investors pass in and out stock market to buy or sell stocks, they usually reference the aggregate value of technological indicators. So we propose equal width aggregate interval and equal frequency aggregate interval. These two data dispersed methods can also support mining causal association rules with level crossing so that we can mine more interesting rules. As the result of t test, the performance of our algorithm is better than FP-growth algorithm apparently. We also find the CFP algorithm is suitable for mining large-scalar database. We arrange causal association rules in an order by different point of view to analysis so as to offer investors assistance in arrangements of investment tactics and the reference of to avoid the loss.
