基於自適應機器學習與平行運算之即時交易決策系統

Abstract

本篇論文以台灣指數期貨之分鐘資料作為研究標的，且有別於以往將一段時間的所有資料都輸入至模型，本研究則根據每分鐘資料之多週期的投資報酬率，尋找歷史資料中與其相似之資料再進行建模，用以預測未來30分鐘後價格之漲跌方向，而在方法學上，則針對傳統Adaboost演算法的弱分類器，加以強化改良為Joint-Adaboost演算法，利用其成對特徵學習之概念，使準確率從53.8%得以提升至61.68%，再輔以OpenCL之框架，讓GPU建模的速度倍增至單純使用CPU的83.02倍，進而使原本因複雜計算而可能產生的延遲決策等待時間，能縮短至與即時無異而不損失其應有之獲利，最後，綜合良好的準確率以及OpenCL的運算加速之驚人表現，顯然能支持本研究提出之尋找相似資料即時建模有其效用性存在，並使投資人能有信心依此而建構一套完整的即時交易決策輔助系統。

In This paper, we propose a novel idea to build a financial module by looking for similar data with return on investment (ROI), which was calculated by Taiwan stock price index futures minutes data. And our target is to predict the direction of the price change after thirty minutes. It will be different from the usual method which only input all the data from a period of time into the module. In Methodology, we focus on the weak classifier of the traditional Adaboost algorithm, to enhance to become Joint-Adaboost algorithm, based on the concept of paired feature learning, to exert the potential of the data classifier, therefore, the final classifier can be able to get the accuracy from 53.8% into 61.68%. However, in the experiments, with OpenCL to parallelized Joint-Adaboost algorithm, using the high performance graphic card will be 83.02 times faster than only using CPU, which minimized the calculation time, which might cause the delay by the complex calculation algorithms, to achieve our real-time goal. In the result, the algorithm have the abilities to create over 60% more accuracy and also accelerate the calculation speed. With the support on this study we can prove the algorithm module to be working properly and looking for similar data to build the module is meaningful, so the investors will have faith to build a whole real-time trading decision support system with us.