GNN Architecture

The temporal GNN predicts spectral properties from sequences of graph snapshots.

[GCNConv -> BatchNorm -> ReLU -> Dropout] x L layers
    |
global_mean_pool -> graph embedding (per snapshot)
    |
Stack T embeddings -> LSTM (M layers)
    |
FC -> [lambda_2, spectral_gap, spectral_radius]

Node features (per bank): volatility 30d, mean return 30d, log price, beta proxy, momentum 20d.

Edge weights: Pearson correlation of daily returns (threshold 0.3).

Training: MSE loss, Adam + cosine annealing, gradient clipping, chronological 80/20 split.