Optimal Data Construction in Supervised Machine Learning for Financial Prediction

Abstract

Despite the widespread adoption of machine learning and deep learning in financial market forecasting, achieving satisfactory results remains a persistent challenge due to the inherent complexities, nonlinearity, and uncertainties in financial data. This study addresses this challenge by introducing two innovative methods in supervised machine learning for feature and label construction. Features are derived from the graphical representation of price data to capture inherent patterns, and the target variable definition is grounded in data momentum, enabling predictions beyond the confines of the training data. These methodological advancements not only enhance regression accuracy, but also expedite model training by facilitating predictions of unobserved values during training. Empirical analysis, employing various financial market datasets and neural network models, demonstrates a substantial improvement in prediction accuracy and efficiency.