GAT-GS: A Genetic Algorithm and Graph Convolutional Network-Based Model for Financial Time Series Forecasting

Abstract

We propose GAT-GS, a graph-attention forecaster for financial time series that fuses dynamic market graphs with sequence modeling and GA-based hyperparameter/feature search. Experiments use three datasets with rolling-origin evaluation and train/test leakage controls: (D1) Global—500 large-cap equities (2010–2020); (D2) Single-market—U.S., EU, and CN subsets (2015–2020); (D3) Industry—technology/finance/energy sectors (2015–2020). Baselines include ARIMA, LSTM, XGBoost, Prophet, CNN-LSTM, and a GCN variant. On the Global dataset (standardized returns), GAT-GS achieves MSE = 0.028 ± 0.003 (95% CI [0.024, 0.034]), MAE = 0.090 ± 0.006, and R2=0.92±0.02, outperforming CNN-LSTM (MSE 0.030, MAE 0.096, R2R^2R2 0.91) and LSTM (MSE 0.032, MAE 0.102, R2R^2R2 0.90). Gains are consistent on Single-market (MSE 0.029 vs. 0.031 best baseline) and Industry (MSE 0.030 vs. 0.032). Ablations show removing attention or community regularization increases MSE by +21% and +14%, respectively. Backtests (10 bps/side) yield 12–13% annualized return, 7–9% max drawdown, and Sharpe 1.3–1.7. Implementation: 60-day rolling correlations for graph edges (∣ρ∣≥0.5), Louvain communities, 2-layer LSTM (128), 2-layer GAT (4 heads), GA search over graph/sequence/fusion hyperparameters; Adam 1e−3, early stopping, RTX 3090/64 GB RAM.