Tail Dependence Without Linear Correlation: Cross-Asset Extreme Co-Movement at Sub-Second Timescales

Abstract

The Epps effect — the decay of measured cross-asset correlation as the sampling interval shrinks — is among the most robust findings in market microstructure. We reproduce it at tick resolution on five US large-cap technology equities (AAPL, MSFT, NVDA, TSLA, AMZN) and then ask a question the correlation framework cannot answer: if these tickers stop co-moving linearly at sub-second scales, do they also stop co-moving in the tails? Scope, stated up front: all five names are drawn from a single sector, so every result is a statement about cross-ticker dependence within US large-cap technology, not about asset classes in general. A common technology factor is a live alternative explanation this design cannot exclude, and a cross-sector replication is the first test nominated for falsifying or narrowing the finding. Using 100 ms to 5 min volume-weighted returns over ten trading days (2024-06-24 to 2024-06-28, quarter-end; 2024-10-14 to 2024-10-18, control), the leading eigenvalue share of the cross-asset correlation matrix sits at 0.205 at 100 ms against an independence floor of 0.200 for five assets, with mean pairwise correlation of 0.0008 — linear structure is essentially absent. Over the same bins, intervals in which three or more of the five assets simultaneously produce a 95th-percentile absolute return occur 5.9 to 13.7 times more often than independence predicts, across the three statistically powered timescales (100 ms, 1 s, 5 s). This excess is not an artifact of shared trading activity. Restricting the test to intervals in which all five assets genuinely traded — holding activity constant by construction — leaves the excess intact and, at 1 s and 5 s, larger (7.1x and 13.7x respectively, both with all ten days statistically powered). The result is further robust to volatility clustering: measured lag-1 autocorrelation of the burst indicator is 0.20 to 0.29 per asset, violating the i.i.d. assumption of the parametric null, so significance is re-established against a circular-shift null that preserves each asset's own temporal clustering exactly while destroying cross-asset alignment. Across 20 day-timescale tests, not one of 20,000 shuffles reached the observed count, and the observed value exceeds its own day's null maximum by margins of 14 to 165 counts. We conclude that linear and tail dependence decouple across timescales: correlation-based risk aggregation, which by construction sees only the former, will understate joint extreme risk at sub-second horizons by roughly an order of magnitude within a concentrated book. Limitations stated in the paper: five same-sector tickers; ten days, one venue tape, one year; mechanism not identified (cross-venue latency arbitrage, shared liquidity provision, index-level flow and a common technology factor are all consistent and are not distinguished); the single-asset control exponents do not exactly reproduce the companion Paper A because cross-asset work requires an absolute epoch-aligned grid rather than a per-file one; and no systematic related-work search was performed, so the contribution is claimed relative to Paper A and not as priority over the literature. Two timescales (10 s, 30 s) show the largest raw ratios but are statistically underpowered and are explicitly not claimed. All analysis code, result JSONs, figure generators, and a re-runnable claims verifier are included.