- Rolling correlation on short windows: a 20-day window jumps around, so the filter turns on and off for no good reason.
- Long windows: 120 days reacts too slowly to regime changes, for example a sterling-specific crisis.
- Correlation on M1 returns: noisy, and dominated by the dollar leg anyway.
- Counting exposure by currency instead of correlation: simpler, and often more robust.
It's easier to test, easier to explain and doesn't need parameter tuning.
Has anyone made correlation filters work in a live EA?