Model Optimization for Stock Market Prediction using Multiple Labelling Techniques

Abstract

In the field of stock market prediction, various labelling techniques have been developed, aiming to improve the accuracy of stock prediction. However, few comparisons and evaluations of labelling techniques have been made in this field. To address this problem, the effectiveness of three labelling methods Raw Return (RR), Fixed Time Horizon (FTH), and Triple Barrier (TB) methods have been studied and compared on Nasdaq 100 Index (NDX) with Multivariate Long Short-term Memory (LSTM) Fully Convolutional Network (MLSTM-FCN) deep learning model. The results are then compared using the confusion matrix and classification report. Experiment results demonstrate that the TB method achieves the highest F1 score on buying signal due to TB being an advanced method that adds two horizontal barriers defined by stop-loss and take-profit. Additionally, the model utilizing the FTH method has the highest overall accuracy, and the model using the RR method generates more accurate predictions of selling signals. The result, therefore, demonstrates that TB method can utilize its additional two barriers to improve price prediction accuracy.