Sign predictability of intraday price returns to formulate appropriate trading strategies with optimum set of equities

Abstract

The prediction of stock market returns has been a focused research area for computational finance, time-series econometrics, and computer science researchers. Typically, market participants rely on technical and fundamental analyses to make predictions, determining their buy or sell strategies based on these assessments. However, scholars in finance and economics have remained skeptical about the predictability of stock returns, particularly in efficient markets. Besides the stock market earnings, a literature on capital martket's direciton predcition has been emerged. Studies in the field of computer science also supported this research area with the help of machine lerning, deep lerning and artifical neural network techniques. Empirical studies in applied economics focuses on efficient market hypotesis whereas computer science studies mostly compares prediction performances of diffirent techniques. In this thesis, closing price direction of 26 stocks in 5 miutes intervals predictited. Only the higly liquid stocks existed in the BIST 30 index for the entire year of 2018 are examined. The data includes 251 trading days and 84 data nodes for each day. Transaction costs are considered to be 2.5 bps and 15 bps is determined as the treshold for a direction signal. If 5 minute closing price of an equity went up by 0.15 percent, it is tagged as "positive" and the opposite as "negative" and the inbetween in considered "steady" to achieve economically significant results. A significant data source—Borsa Istanbul's 'data analytics' information distributed through data vendors—is utilized. There are 39 analytics providing order book statistics regarding for last 1 minute, 5 minute and intraday periods. Initially, a specific method for extracting valuable insights from the complex raw data of Borsa Istanbul is described and implemented. Subsequently, standard scaling and dimension reduction methods, such as Principal Component Analysis, are employed to enhance efficiency. Seven different machine learning algorithms, Logistic Regression, K-Nearest Neighborhood, Support Vector Machines with radial and sigmoid kernels, Naïve Bayes, Decision Tree, Random Forest, are compared. All the methods trained with former 95 percent and tested using the last 5 percent of the complete dataset of the year 2018. The performance levels of each method for twenty-six highly liquid stocks are assessed in terms of Macro Averaged F-Measures. Furthermore, in this thesis, the effectiveness and significance of machine learning algorithms are compared using confidence intervals calculated from the confusion matrices for Macro Averaged F-Measures—an innovative approach within the economics and finance literature. Additionally, the use of three classes for stock price directions 'positive ,' 'negative" and 'steady' rather than just two is a rare aspect in academic studies, aligning more closely with data analytics practitioners' methodologies. Concerning the 5-minute lagged data, a statistically significant predictability is found in nine of the equities. K-Nearest Neighbors, Decision Tree and Random Forest yielded significant predictions. Naïve Bayes and SVM-sigmoid achieved better for a few equites while Logistic Regression and SVM-rbf not for any. However, for the 10 and 60-minute lagged data, predictability remains only in four and none of the equities, respectively. Essentially, markets assimilated Borsa Istanbul's data over time for those equities. Moreover, economic gains for the nine equities are analyzed with algorithms not allowing short selling and allowing short selling, based on these predictions. 'Positive", 'negative' and 'steady' signs treated as 'buy', 'sell' and 'keep the position' signals in the first scenerio while 'steady' allowed as the 'short-sell' signal in the second scenerio. Those scnerios are compared with passive buy and hold strategy and it has been revealed that determined trading strategies depending on the machine lerning algorithms performs higher earnings or less losses for KOZAA, KOZAL and KRDMD stocks when short selling is not allowed. If short selling is allowed TKFEN is also added to this list. Test period performances of the determined trading strategies and the stocks' own price performances are graphically illustrated in the thesis. Since four equities are observed to generate higher economic gains through machine learning-supported trading strategies compared to their own price performances, these findings for those wquities indicate, within the framework of the 'Efficient Market Hypothesis,' a lack of 'Semi-strong-form efficiency.' The methods and techniques used in this study such as creating tree dimensional confusion matrices, determining standard deviation and confidence intervals for MA F-Measures, and benchmarking them with a theorical MA F-Measure and developing trading strategies with three signals will support further research