Correlations change in days. Rolling windows adapt in weeks. That lag costs real money.

This paper introduces a Transformer + Graph Neural Network that forecasts 10-day-ahead stock–stock correlations — and uses them directly for statistical arbitrage clustering.

• Predicts correlations in Fisher-z residual space (stable, realistic),

• Learns regime-aware market structure via graph attention,

• Preserves the shape of the correlation network (not just point accuracy).

The payoff?

• Sharpe 1.84 vs 0.65 for the S&P 500,

• −9% max drawdown vs −34%,

• Biggest gains during crises, when backward-looking clusters fail.

The insight is simple but powerful: Stop trading yesterday’s correlation network. Trade tomorrow’s.