GPU Correlation Caching - Research Notes

Experiment Overview

Problem

The calculate_correlation_matrix() function in alpaca_trading/selection/portfolio/correlation.py was slow because:

1. Rolling correlation is O(n² × m): Nested Python loops at lines 138-153
2. CPU-bound pandas: Uses window_df.corr() repeatedly in loop
3. No caching: Recomputes from scratch every run, even with identical data

Symptom: Selection hangs at "Correlations: 1/2 (50%)" for 60-90+ seconds with 100 symbols, 4-6+ minutes with 200 symbols.