DCAGAT: A Graph Attention-Based Model with Reconstruction Regularization for Dollar-Cost Averaging Investment Prediction

Abstract

The financial market is characterized by high volatility and noisy data, making it a formidable challenge to forecast trends and design robust investment strategies. In this paper, we propose an innovative prediction model that integrates multi-feature fusion with graph attention mechanisms to address these challenges, specifically tailored for dollar-cost averaging (DCA) strategies. Our model, termed DCAGAT, leverages Graph Attention Networks (GATs) to dynamically assess the interdependencies among various financial assets. By incorporating multiple market features—historical price fluctuations and trading volumes—the model constructs a comprehensive representation of market dynamics. A key innovation is the inclusion of an autoencoder-inspired reconstruction verification mechanism, which mitigates overfitting by ensuring that the model focuses on persistent market trends rather than transient noise. We validate the effectiveness of DCAGAT using historical data from Yahoo Finance and ETF from 2012 to 2022. We benchmark DCAGAT against four neural baselines—DNN, Conv1D, GCN and ST-GCN—on three DCA-oriented metrics: one-, seven- and fourteen-day directional accuracy (Adir) and Top-5 hit-rate (Aselect). While DCAGAT matches the best baseline on Adir, it consistently improves Aselect across the evaluated scenarios, with every gain passing the paired statistical-significance test., underscoring its superior stock-selection capability. Overall, our research provides a robust framework for financial market forecasting by combining advanced graph-based learning techniques with feature-rich data integration, offering valuable insights for investors seeking to optimize multi-day investment decisions in unpredictable market environments.