LSTM-based Stock Price Prediction by Utilizing an Adaptive Labelling Strategy to Optimize the Feature Selection Duration

Abstract

Abstract: The growing popularity of predicting stock prices using advanced machine learning techniques, particularly Long Short-Term Memory (LSTM) models, has been driven by their ability to uncover complex patterns that are challenging for humans to identify. This trend has been supported by the widespread availability of data. In this study, we employed an adaptive labelling strategy to maximize intraday returns and examined the impact of the quantity of training data points on the accuracy of LSTM-based predictions of intraday returns. We evaluated the average returns and volatility across different training spans, verifying the results over an extended period. By comparing various partition sizes, we determined that a partition size of 240 yielded high Sharpe ratios (>2) and improved mean intraday returns. Our research demonstrates that an adaptive labeling strategy combined with LSTM models can be instrumental in achieving maximum intraday returns, with a partition size of 240 being the optimal choice for predicting stock prices.