Cascading Logistic Regression Onto Gradient Boosted Decision Trees to Predict Stock Market Changes Using Technical Analysis

Abstract

In the data mining and machine learning fields, forecasting the direction of price change can be generally formulated as a supervised classfii cation. This paper attempts to predict the direction of daily changes of the Nasdaq Composite Index (NCI) and of the Standard & Poor's 500 Composite Stock Price Index (S&P 500) covering the period from January 3, 2012 to December 23, 2016, and of the Shanghai Stock Exchange Composite Index (SSEC) from January 4, 2010 to December 31, 2014. Due to the complexity of stock index data, we carefully combine raw price data and eleven technical indicators with a cascaded learning technique to improve the performance of the classifi cation. The proposed learning architecture LR2GBDT is obtained by cascading the logistic regression (LR) model onto the gradient boosted decision trees (GBDT) model. Given the same test conditions, the experimental results show that the LR2GBDT model performs better than the baseline LR and GBDT models for these stock indices, according to the performance metrics Hit ratio, Precision, Recall and F-measure. Furthermore, we use these models to develop simple trading strategies and assess their performance in terms of their Average Annual Return, Maximum Drawdown, Sharpe Ratio and Average Annualized Return/Maximum Drawdown. When transaction costs and buy-sell thresholds are taken into account, the best trading strategy derived from LR2GBDT model still reaches the highest Sharpe Ratio and clearly beats the buy-and-hold strategy. The performances are found to be both statistically and economically signi ficant.