利用時間序列模型建構市場中性交易策略

Abstract

本研究討論如何構一個市場中性策略，消除市場趨勢的變動而賺取與市場無關之個別公司報酬來穩定獲利，我們利用時間序列模型，包括單根檢定，共整合理論，主成分分析法來找出合適的交易配對群以及配對群中個股投資的資金比例，並以無母數方法尋找開倉門檻極大化期望報酬率，也設定不同的停損門檻方法分析損益，也設計股數調整方法符合交易標的的不可無限切割的限制。不同的共整合方法篩選與不同的篩選配對機制而使得有不同的策略，因此我們建構迴歸模型，並設計虛擬變數(Dummy variables)，來判斷不同策略方法對於損益值的影響是否顯著，找出整體最佳策略。 在資料取美股S&P500成分股在2009年到2013年資料，並設定用250日觀察期資料決定交易策略後，用之後60日交易的結果，我們主要的結論為:兩檔構成價差序列的結果中，我們發現兩檔中對於單筆損益的最佳策略為策略三，以相關係數搭配Johansen共整合檢測法為篩選依據，但會有因對於單筆損益與總損益不同目標而有不同最佳策略結果，若以總損益當作目標，則發現最佳總損益會是在策略一，以PCA/HCA做篩選配對群，再以Johansen的共整合檢測法做篩選，並以定態性質來判定停損門檻，會得到最佳總損益。以定態性質做為新停損門檻的設定(New stop)，相較於其他較嚴格的停損門檻，對於極大化每一次交易損益值下是最佳策略。而對於當握有有限資金時，我們可選擇較嚴格的交易篩選機制，期望會有最高的單筆損益，在迴歸結果中是個股以相關係數>0.9與SNR >40可賺取最高的單筆損益值。在三檔的結果中，發現應剔除當有兩組共整合係數時的樣本，此樣本做交易會因在不同共整合係數上跳動的問題產生虧損，而三檔的樣本損益結果均為負值，猜測可能因為分類分群方法使將原本能賺取報酬的股票群剔除了，而使損益結果為負。

The thesis discusses how to construct a market neutral strategy to gain stable profits in investing S&P500 stocks, and we want to eliminate the systematic risk of the portfolio by longing an asset(s) and shorting another one(s) and earn the return which is irrelevant to the market returns. Time series methods including co- integration theory, ADF test, principal component analysis are used to find proper investing proportions for a selected group of stocks group to form a market neutral portfolio,The mean-reverting property of such a portfolio allows us to earn stable profits by buying at low and sell at high. Besides, the co-integration vector or the ratio of investment amount is properly adjusted to meet the restriction that the stock can’t be infinitely divided. To maximize the investment performance, we examine the impact of different setups and statistical methods on profits by constructing the regression model with dummy variables for different methods. We test our strategies with S&P500 component stocks during the period 2009~2013.The price information from the 250-day investigated periods used to determine the trading strategy, and the strategy is then used to trade in the following 60-day trading period. Sixteen consequent trading periods are tested during these five years and our main findings are presented as follows: First, the best strategy for maximizing each single trading for the pair trading strategy is strategy three: filtering an examination all possible trading pairs by correlation and Johansen test. However, to maximize the lump sum profit over a trading period rather than a profit of a single trading, the best strategy is strategy one, using PCA/HCA and Johansen test. Second, we create a stop loss strategy called “New stop”, which unwinds the trading when the spread series of the portfolio is not stationary. We find it to be the best stop strategy to earn in single profit. If we have limit money to invest, we can only choose the filter which tends to earn the most in each trading. We find the filters like correlation>0.9 and SNR>40 are the best filter in our regression results. Considering the spread of a market neutral portfolio composed of three stocks, we can delete the spread which have two co-integration vectors, since the spreads may jump repeatedly over two co-integration vector. That would cause our lost since our strategy may no longer be market neutral. Finally, the results in three stocks strategies all have negative profit in average we guess the reason is that we use inadequate filtering methods that removes profitable stock triples.