Interesting, but I think there are a few econometric checks missing that could materially affect the conclusions. Most importantly: if any of this is based on price levels or smoothed price levels (e.g. 90-day moving averages), both series are almost certainly non-stationary. Correlations and cross-correlations between non-stationary series can look highly significant even when the relationship is spurious, unless you explicitly test for stationarity or cointegration. Related to that: - Using rolling averages introduces strong autocorrelation and overlapping observations, which inflates t-stats and p-values if not corrected. - The peak at a −49 day lag looks like it was selected after scanning many lags, that’s effectively multiple testing, so the reported p-value likely overstates significance. - Min–max normalization makes unrelated trending series visually align, which can be misleading. To really support a predictive claim, do: 1. stationarity tests (ADF/KPSS) and returns-based analysis, 2. lag selection done out-of-sample or with multiple-testing correction, 3. HAC / bootstrap inference, and 4. a clean out-of-sample forecast comparison vs a simple benchmark. Without those, it’s hard to rule out a constructed lead-lag relationship rather than a genuine signal.

---