運用改良式深度學習方法建構套利策略模型於高頻配對交易

Abstract

配對交易(Pairs Trading)是一種統計套利策略，找出一組表現相似的資產，一旦 發現價格偏離理性均衡(Rational Equilibrium)時，投資人將買進弱勢標的資產並同時 賣出強勢標的資產，賺取其中的價差(Spread)，其收益穩定的特性讓它逐漸成為國際金 融市場常用的交易策略。 本研究利用近日火紅的深度學習技術，建立智慧型套利交易系統，嘗試從大量的台 灣指數期貨與小型台指 Tick 資料中捕捉套利訊號，我們提出了改良的時間序列視覺化 方法，將不同時間框架的歷史波動度轉為圖片，對於捕捉套利機會有更佳的效果，除此 之外，我們也改良了卷積神經網路模型，將套利的知識規則帶入深度學習模型之中，利 用 Filterbank Learning 卷積神經網路，將套利知識規則 Filter 取代隨機產生的 Filter，讓模型可以依據知識規則捕捉出更高品質的特徵並且獲得最好的準確度以及獲 利能力。 經實驗結果，改良後的波動度視覺化方法最高有 66.80%的準確度，與價格偏離交易 策略最高有 16.10%的改善率，而 Filterbank Learning CNN 結合波動度視覺化方法有 高達 67.03%的準確度，與價格偏離交易策略有高達 16.16%的改善率，平均一天有高達 577 點的獲利能力，驗證資訊技術結合財金知識可以創造更好的套利交易模型。

Pairs trading is a statistical arbitrage and market-neutral strategy, which long and short highly correlated financial assets simultaneously, and then profit from the variations of asset prices exceeding rational equilibrium. This investment tool try to catch the price spreads between different trading targets. The attributes of return stability and low risk also make the pairs trading much more popular in internal financial markets. This paper proposed a novel intelligent pairs trading system in high-frequency Taiwan Stock Exchange Capitalization Weighted Stock Index Future (TX) and its light-weighted version MTX data based on deep learning technique. A time-series data visualization method is applied to represent the correlation between historical volatilities under different timeframes. Moreover, this research also improved convolutional neural networks (CNN) model with a specially designed filter banks. Compared to the traditional random-generating filters, our approach can model some investment know-hows and increase the training accuracy and efficiency dramatically. The experimental results show that our improved time-series visualization method can achieve near 66.80% testing accuracy, which is 16.10% better than original rule-based trading rules. The filterbank learning CNN can further improve this learning system to 67.03% accuracy. In average, the extended trading strategy based on our proposed selection model can earn 577 points per day by buy-and-hold trading method. These experiments prove that the combination of advanced artificial intelligence techniques and financial knowledge can create much better arbitrage-trading model.