一個基於增強式學習的動態資料驅動系統: 以台灣股票市場為例

Abstract

傳統上,時間序列關聯強度的分析大多屬於靜態,通常在某特定時間區 段中分析時間序列的關聯,此方法簡單直覺,然而卻沒考慮資料間會彼此互相影響 的特性,以及沒考慮到關聯強度會隨著時間的改變而有所變動,造成所產生的關聯 強度數值與實際數值上有偏差,也同時影響到時間序列預測的結果。本論文提出一 個基於增強式學習的動態資料驅動系統,採用增強式學習方法找尋時間序列資料之 間的動態關聯強度,並使用動態資料驅動的架構,來因應時間序列資料彼此互相影 響的特性。我們選擇具有時間序列特性的股票市場為研究對象,時間序列的資料從 台灣股票市場中,選擇 50 家平均交易量 600-1000 的上市公司,資料時間為 2015 年 1 月 1 號到 2016 年 1 月 1 號。實驗結果顯示比較現有方法,預測準確度從 45.3%提 昇至 60.2%。本論文所提出的方法可作為規劃股票投資組合時,股票之間的連動關係 參考。

Most analysis on time series data are based on a static approach, such as selecting specific time period in order to analyze relation strength. Although they are simple and intuitive, they lack of considerations on data interaction between themselves and on the fact that the data correlation strength will change as time goes by. Consequently, their predictions usually are not accurate. This thesis proposes a dynamic-data-driven application system based on reinforcement learning. This proposed system uses reinforcement learning to find correlation strength between time series data and uses dynamic-data-driven application system to consider the characteristic of data interaction between themselves. The time series data are from 50 Taiwan IPO stock market(Jun. 1, 2015~Jun. 1, 2016), on which its volume is 600-1000 per day. The experimental results show that the prediction accuracy is raised from 45.3% to 60.2%. The proposed system can be used in portfolio making reference to the correlation between stocks.