Expert Systems, Journal Year: 2025, Volume and Issue: 42(3)

Published: Jan. 29, 2025

ABSTRACT Technical analysis, which includes technical indicators and charts derived from specific rules, has proven effective widely used for stock movement prediction. However, chart evaluation is often limited by subjectivity, arising sparse types substantial information loss due to rigid rules. While pattern recognition algorithms have been developed address this issue, they still rely on manual labelling primarily focus closing prices, leaving much of the chart's broader untapped. To overcome these limitations, we propose a novel framework called ChartNet, designed extract general reduce subjectivity in analysis. ChartNet employs unified representation across financial series with varying simplification levels leverages triplet function unsupervised training, eliminating need labelled data. Compared several state‐of‐the‐art baselines, our reached best prediction accuracy CSI‐300, SZ‐50 components Dow Jones Index 2022: 65.91%, 63.70% 64.96% respectively. In backtesting using actual data, achieves highest average return 1.12 1.15. Furthermore, highlight interpretability through two case studies, some important failure cases, illustrating its capability uncover meaningful insights charts. This research contributes advancing objective promoting more comprehensive understanding chart‐based performance.

Language: Английский