以相關性為基礎之配對交易最佳化技術

Correlation-based Pair Trading Optimization Techniques

黃俊傑(Chun-Chieh Huang)

指導教授 ： 陳俊豪

淡江大學/工學院/資訊工程學系碩士班/碩士(2020年)

https://doi.org/10.6846/TKU.2020.00471

摘要

金融市場中充斥著眾多影響投資獲利程度的因素，故許多交易策略與組合最佳化技術不斷被提出。其中，配對交易策略是廣泛應用的交易策略之一，因其符合市場中性且容易使用。傳統配對交易策略是利用兩個相關性較高的股票或其它證券，一旦兩者之間出現了背離的走勢且此背離在未來是會得到糾正的，則可產生套利的機會。然而，負相關性較高的股票亦可形成配對交易，故本論文利用此特性，為使交易策略獲利，我們提出三個演算法來達成此目標，分別為：(1)相關係數為基礎之配對交易演算法（Correlation-coefficient based pairs trading algorithm, CPT)；(2)布林通道為基礎的配對交易演算法(Bollinger-band based correlation-coefficient based pairs trading Algorithm, BBCPT)；(3)遺傳為基礎的的配對交易演算法(Genetic bollinger-band based correlation-coefficient based pairs trading algorithm, GBBCPT)
在CPT方法中，首先計算標的中任兩公司的相關係數，如其相關性為負值且小於預設之門檻值，則形成一交易配對。接著，利用公司開盤與收盤股價漲(跌)幅度判定進場訊號後，做多預期上漲並放空預期下跌標的，達到停利或停損條件時則結束該組交易配對。此方法中之進出場訊號判定方式易導致獲利波動大，故提出方法二改善此問題。
在BBCPT方法中，其結合布林通道(Bollinger band)強化所提之配對交易策略，即利用歷史股價計算移動平均值、股價壓力值、股價支撐值等數值作為進出場訊號。雖然方法二強化了CPT的進出場訊號，但相關係數門檻值、進場通道寬度與出場通道寬度的設定會影響此方法的獲利，因此，我們進一步利用遺傳演算法進行參數的最佳化。
在GBBCPT方法中，每個染色體表示一組可能的相關係數門檻值、進場通道寬度與出場通道寬度，且使用染色體之累計獲利為適合度評估函數衡量其優劣，之後透過演化程序找出近似最佳的參數設定。
最後，實驗使用了台灣50成分股中共44家公司歷史股價資料，透過不同的實驗分析來驗證本文提出的三個方法的有效性，包含：(1)不同限制與參數設定對CPT的影響；(2) 不同限制與參數設定對BBCPT的影響與(3)不同訓練區間與參數設定對GBBCPT影響。

關鍵字

資料探勘 ； 相關性 ； 配對交易策略 ； 布林通道 ； 交易策略最佳化

並列摘要

The financial market is full of many factors that affect the return on investment, so many trading strategies and portfolio optimization techniques are constantly being proposed. Among them, the pairs trading strategy is one of the widely used trading strategies because it conforms to market neutrality and easy to use. In the traditional pairs trading strategy, the pairs are formed from two stocks or other securities with high correlation. Once a divergence trend occurs between the stocks in the pair, it can be known that the deviation will be corrected in the future, and an opportunity for arbitrage will appear. However, stocks with higher negative correlation can also be utilized to form trading pairs. Based on that property, to increase the returns of the pair trading strategies, this thesis presents three approaches to reach the goal, including: (1) Correlation-coefficient based pairs trading algorithm (CPT), (2) Bollinger-band based correlation-coefficient based pairs trading algorithm (BBCPT), and (3) Genetic bollinger-band based correlation-coefficient based pairs trading algorithm (GBBCPT).
In the CPT, the correlation coefficient of any two companies in the pair is calculated firstly. If the correlation is negative and less than a given threshold, a trading pair is formed. Then, using the difference of opening and closing stock prices to determine the entry signals, long the uptrend target and short the downtrend target. When reaching the take-profit or stop-loss conditions, the trading pair is closed. In the CPT, because the way to determine the trading signals may cause large fluctuation in profit, the second approach is proposed to solve this problem.
In the BBCPT, it combines the Bollinger band to strengthen the proposed pairing trading strategy. In other words, it utilizes the historical stock prices to derive moving averages, stock pressure values, and stock price support values for finding buying and selling signals. Although the BBCPT provides suitable trading signals for trading, the parameters those are the correlation coefficient threshold, entry channel width, and out channel width will affect the profitability, we thus further design an algorithm for parameter optimization using the genetic algorithms.
In the GBBCPT, a chromosome represents a possible correlation coefficient threshold, entry channel width, and out channel width. The cumulative profit is employed as fitness function to measure the quality of a chromosome. Then, the evolutionary process will be repeated to find a near optimal parameter setting for the pair trading strategy.
Finally, the empirical experiments were conducted on the historical stock price data that collected from 44 companies in Taiwan 50 ETF to verify the effectiveness of the proposed methods, including: (1) the influence of different restrictions and parameter settings on the CPT; (2) the influence of different restrictions and parameter settings on the BBCPT; and (3) the influence of different training intervals and parameter settings on the GBBCPT.

並列關鍵字

Data mining ； correlation ； Pairs trading strategies ； Trading strategy optimization

參考文獻

相關文獻

[1] A.M. Farley and S. Jones, “Using a genetic algorithm to determine an index of leading economic indicators, ”Computational Economics,vol.7, no.3, pp.163–173,1994.

[2] A. Mochon, D. Quintana, Y. Saez, and P. Isasi, “Soft computing techniques applied to finance, ” Applied Intelligence, vol. 29, no. 2, pp.111–115,2008.

[3] K.-J. Kim and I. Han, “Genetic algorithms approach to feature discretization in artificial neural networks for the prediction of stock price index,” Expert Systems with Applicatio- ns, vol.19, no. 2, pp.125–132,2000.

[4] Y. Becker, P. Fei, and A. Lester, “Stock selection – an innovative application of genetic programming methodology,” Genetic Programming Theory and Practice IV, vol. 5, ch- apter12, pp.315–334, 2006.