The Statistical Significance of the Inclusion of Graph Neural Networks in the Financial Time Series Forecasting Problem

Battle Count: 96

The paper addresses stock price forecasting for S&P 100 constituents, which is directly relevant to quantitative trading. However, the improvements from GNN inclusion are modest for most models, and the statistical significance is metric-dependent. The paper focuses more on methodological rigor (statistical testing) than on trading strategy development or portfolio-level performance. The forecasting horizons (1, 5, 20 days) are relevant for short-to-medium term trading strategies.

The Time-Geometric model requires implementing: (1) visibility graph construction at each time step, (2) GCN layers with message passing, (3) LSTM processing of geometric patterns, (4) combination with temporal models, and (5) fully connected output layers. The dynamic graph computation adds computational overhead. Hyperparameter optimization via Optuna with 1000 trials per model is computationally expensive. The statistical testing framework (multiple tests, Friedman, Nemenyi) adds analytical complexity.

The paper provides detailed hyperparameter tables for all models, specifies the dataset (S&P 100 from Yahoo Finance), describes pre-processing steps, and uses well-known algorithms (visibility graph, GCN, Optuna). However, no code repository is explicitly mentioned, and the exact random seeds and full training pipeline details are not fully specified. The ts2vg Python package is referenced for visibility graph computation.