#64 — FX stat-arb (Avellaneda-Lee eigen-residual)

Hypothesis — what it tests

Whether FX pair returns, after hedging out the common currency factors via PCA, leave a mean-reverting residual that can be traded with a fixed fade-the-s-score rule.

Description

This tested the one 'phonon'/eigen-mode idea from a physics-finance analogy the project had never run: PCA the 12-pair (8-currency) return panel into risk-weighted eigen-portfolios (the currency factors), regress each pair on the top-k factors, and trade the Ornstein-Uhlenbeck s-score of the hedged residual, following Avellaneda & Lee (2010). It was built fully causal with a fail-fast IC test before any trading simulation. The residual autocorrelation's sign turned out to track the macro regime — strong momentum in 2022 (IC +0.19, t=+10, the Fed-hike/JPY-collapse year) flipping to strong reversion in 2024-25 (IC −0.13, t=−7) — so a fixed always-fade bet fails the IS-first gate outright. Even where the residual does revert, the signal lives in a 12-leg hedged basket whose spread would eat it, and the only retail-deployable naked single-pair form is ~3x weaker and not significant OOS (t=−1.24). Closed as non-stationary and uncapturable; PCA eigenmodes of FX majors turn out to be roughly the currency factors the project already traded via CSI.

Key result

Indicators

Algorithms

Code