STGV-Similarity between trend generating vectors: A new sample weighting scheme for stock trend prediction using financial features of companies

Abstract

The problem of effective stock trend prediction has aroused much attention these years for its profitability. The development of algorithmic trading drives explosive growth in fast and effective techniques for trend predictions. However, little attention has been paid to sample weighting schemes in this field. In this article, we propose a new sample weighting scheme for stock trend predictions based on financial features of companies. Specifically, stock trends are supposed to be determined by hidden market states represented by trend generating vectors. These vectors can be generated from financial features of companies. The scheme considers similarities between trend generating vectors (STGVs) when assigning weights to samples from different periods to differentiate their prediction capabilities. Similarity scores calculated with a proper metric are adopted to measure similarity. We train models with STGV to predict stock trends. Extensive experiments are conducted to figure out the most suitable similarity metric used in STGV and demonstrate its superiority over other sample weighting schemes together with its generalizability.