AC-Seq2Net: A Math-Guided Hybrid Deep Learning Framework for Explainable Geometric Pattern Recognition in Financial Time Series

Abstract

Financial market data is usually rough and volatile. This noise makes it difficult to detect clear price structures from raw intraday records. Traders still rely on geometric patterns such as symmetrical triangles to describe consolidation and potential breakouts. Deep learning models are able to acknowledge market internal conditions with high scores though they tend to acquire concealed relations. They do not capture the triangle geometry in an explicit form. This introduces a gap of representation and lowering interpretability. In order to overcome this problem, we proposed AC-Seq2Net. It is a math-guided hybrid symmetrical triangle recognition model in intraday price data. The pipeline segments the price stream into rolling windows and applies local min-max normalization to keep the scale consistent across different price levels. Ground truth labels are produced through geometric boundary extraction using ordinary least squares trendlines, upon which slope convergence and volatility contraction are derived as explicit engineered features. Unlike purely black-box sequence models, AC-Seq2Net integrates these geometric descriptors alongside hybrid temporal representation learning. A two-stage 1D-CNN extracts local structural patterns, a Bidirectional LSTM captures temporal dependencies, and an Attention mechanism highlights critical convergence boundaries. This design encodes symmetrical triangle geometry directly rather than approximating it through latent feature learning alone, bridging deep learning performance with trader-interpretable geometric reasoning validated through SHAP-based explainability. The triangle probability is produced using a sigmoid layer. Experiments on longitudinal S&P 500 and NASDAQ 100 data report 97.70% accuracy and 0.9947 AUC, confirming that geometric signals rather than random fluctuations drive model decisions.