A3T-GCN for FTSE100 Components Price Forecasting

Battle Count: 108

The paper directly addresses stock price forecasting for FTSE100 constituents, a core quantitative trading task. The A3T-GCN architecture demonstrates practical applicability with R²=0.9936 and MRE of 3.46% for next-day predictions. The findings on ALR features reducing computational requirements while maintaining accuracy are practically relevant for trading systems. However, the 3.46% MRE may limit profitability for blue-chip stocks, and the model's inability to predict event-driven movements constrains real-world trading utility. The paper acknowledges that similar strategies (LSTM) have been arbitraged away, suggesting limited alpha persistence.

Implementation requires: (1) Graph construction from sector classifications and correlation matrices using NetworkX/NumPy, (2) Feature engineering of 8 node features including rolling window annualized log returns, (3) Stacked TGCN cells with 2-layer GCNConv + GRU + attention mechanism in PyTorch, (4) Handling 54.5GB processed dataset, (5) Multiple hyperparameter configurations (8 versions tested). The architecture is moderately complex with multiple interacting components, though the PyTorch string representation suggests a defined implementation. Computational requirements are significant (54.5GB data, multiple training runs).

The paper provides detailed methodology including hyperparameters (learning rate 0.005, weight decay 0.00001, batch size 32, 10 epochs), data sources (Yahoo Finance, Wikipedia), feature engineering steps, and graph composition procedures. However, no code repository is mentioned, and the stochastic nature of the model introduces variability. The 90/10 train-test split and specific date ranges are provided. The 54.5GB dataset size and computational constraints are noted.