TeX 62.4%
Python 37.5%
1% ============================================================================2\section{Discussion}3\label{sec:discussion}4% ============================================================================56\subsection{Answers to the Research Questions}78\paragraph{Structure of connectedness.} Cross-asset shocks explain 54.8\% of the ten-day forecast-error variance of log realized volatility, an estimate that is stable across lag orders, horizons and the treatment of weekend crypto variance (53.7--55.5\%) and whose bootstrap interval is about four percentage points wide. The hierarchy has three robust elements: the two equity index futures are the only net transmitters whose intervals exclude zero in the full sample and in every robustness specification; the realized volatility of the VIX, crude oil and sterling are robust net receivers; and the remaining instruments---the euro, the yen, gold, bitcoin and ether---have net positions that cannot be distinguished from zero over the decade. This ordering is consistent with the primacy of U.S. equities found by \citet{DieboldYilmaz2012} for stocks, bonds, exchange rates and commodities, and it adds two observations: the realized volatility of an implied-volatility index behaves as a derivative of its underlying, and crude oil, often cast as a source of macroeconomic volatility, has been a receiver of equity-market volatility over 2016--2026.910\paragraph{Regime dependence.} The total index varies much less across regimes (59.9--71.7\%) than the directional composition. The COVID-19 crash is the only sub-sample in which an instrument other than the equity futures takes the lead, USD/JPY, with a point estimate that is large and a sign that is established but a magnitude that the data cannot pin down ($+7$ to $+135$ percentage points). The statement in the earlier version that equity futures became net receivers during the crash is not supported once sampling uncertainty is taken into account; nor is the statement that bitcoin became a net transmitter during the 2022 tightening cycle. What the regime analysis establishes with confidence is more modest: the hierarchy is dominated by equities in calm and macro-driven periods, exchange-rate volatility became the common factor during the dollar funding squeeze of March 2020, and in the most recent period bitcoin and ether are net receivers.1112\paragraph{Crypto integration and its direction.} Integration has increased, and in a well-defined sense. The share of bitcoin's forecast-error variance explained by traditional instruments rose from below 10\% in 2017 and 2019 to 20--34\% in 2020--2024 and 46\% in 2025, bitcoin's own share fell from 78\% before 2020 to 48\% after, and the pairwise shares with ES grew in both directions. The direction of net transmission, however, runs from equities to crypto: bitcoin's net spillover is negative in every calendar year since 2020 and significantly so in the spot-ETF period, its 500-day rolling net spillover has been negative since 2022, and the session-ordered regressions show that the U.S.-session volatility of ES, NQ and the VIX predicts bitcoin's next Asian-session volatility whereas bitcoin's overnight and weekend volatility does not predict U.S.-session equity volatility. The one exception, a small but significant predictive effect of bitcoin's earlier-session volatility on the U.S.-session volatility of the VIX, is consistent with option markets pricing overnight risk-asset news before the cash open rather than with crypto volatility originating shocks. These results align with the post-2020 increase in crypto--equity comovement documented by \citet{Iyer2022} and \citet{Elsayed2022} and with the finding of \citet{ConlonMcGee2020} that bitcoin did not act as a safe haven in March 2020, and they qualify the interpretation of growing integration as growing influence: the volatility of the two largest cryptocurrencies has become more exposed to the traditional system, not the reverse.1314\paragraph{Time variation.} Total connectedness fell to its minimum during the 2017 crypto boom, rose sharply during the COVID-19 crash and again during the 2022 tightening cycle, and reached the highest values of the sample in 2025, when the calendar-year estimate was 71.5\% against 51--55\% in 2016--2019. Whether the 2025 level reflects a lasting change in market structure---the integration of crypto through exchange-traded funds, the growth of volatility-targeting and option-selling strategies---or the particular shocks of that year cannot be settled with one year of data. We refrain from interpreting the rolling index as an early-warning indicator: in the trading-day panel it does not rise ahead of the March 2020 crash, and the February 2018 spike that motivated this interpretation in the earlier version was an artefact of forward-filled observations.1516\paragraph{Intraday distribution and sessions.} Equity-futures volatility is concentrated in the U.S. cash session, which generates 53--54\% of daily variance, with the familiar peaks at the open and before the close; the realized volatility of the VIX is even more concentrated (63\%). Bitcoin's volatility is nearly uniform across the 24 hours, with a peak-to-trough ratio of 1.2 and a mild hump in U.S. hours, and weekends carry 22\% of weekly variance. Round-the-clock trading does not translate into round-the-clock information for equity markets: once Friday's volatility is accounted for, the volatility of bitcoin over the weekend carries no incremental information about the volatility of ES on Monday.1718\subsection{Comparison with the Literature and Implications}1920The connectedness level of 55\% is within the range reported for daily volatility systems of comparable size \citep{DieboldYilmaz2012,DieboldYilmaz2014}, and the dominance of equity futures matches most multi-asset applications. Our crypto results sit between the early finding of isolation \citep{Corbet2018,Baur2018} and the post-pandemic finding of integration \citep{Iyer2022}: the gross linkages have grown, the net direction is from equities to crypto, and the within-crypto block remains the dominant source of crypto variance, as in \citet{Ji2019a}. The intraday results confirm, with ten years of 5-minute data, the U-shaped equity pattern of \citet{Wood1985} and \citet{Harris1986} and the U.S.-hours hump in bitcoin volatility of \citet{Eross2019}.2122For practitioners, three implications follow. The realized volatility of VIX-linked instruments responds to equity-futures volatility rather than anticipating it, so that positions in VIX products are not an independent source of volatility information at the daily frequency. The diversification benefit of cryptocurrencies has eroded in the volatility dimension since 2020 and is smallest in the most recent period, precisely when equity volatility is high. And because equity-futures variance is concentrated in the U.S. cash session while bitcoin's is spread evenly, risk limits that are calibrated on daily volatility understate the intraday concentration of equity risk and overstate that of crypto risk.2324\subsection{Methodological Lessons}2526Two methodological conclusions apply beyond this paper. First, in panels that mix 24/7 and exchange-traded instruments, the construction of the trading calendar is a modelling decision with first-order consequences: forward-filling the exchange-traded series over weekends changes the total index by five percentage points, the net spillover of ES by fifteen, and the apparent isolation of the crypto block from a cross-asset share of 6\% to one below 1\% (\Cref{sec:appendix_audit}). Second, connectedness estimates from short windows are imprecise: a ten-variable VAR(4) has 41 coefficients per equation, and in windows of 85--300 observations the 90\% intervals of net spillovers are 30--130 percentage points wide. Regime narratives built on the sign of such estimates, including some in the earlier version of this paper, do not survive the addition of confidence intervals. Time-varying-parameter VARs \citep{Antonakakis2020} or shrinkage estimators would be natural ways to recover time variation with fewer effective parameters.2728\subsection{Limitations}29\label{sec:limitations}3031Several limitations remain. The 5-minute realized variance includes jump variation and, for the composite cryptocurrency feeds of 2016--2017, variation that reflects price discrepancies across exchanges rather than fundamental volatility; the results for 2018--2026 are qualitatively unchanged, but a jump-robust estimator \citep{Andersen2007} and a single-exchange price series would be preferable. The filter that removes 5-minute returns beyond twenty standard deviations, applied before realized volatility was computed, may have removed genuine extreme moves on crash days and so understated realized volatility on the days that matter most for connectedness; its effect could not be quantified without re-processing the raw bars. The continuous futures series splice contracts at roll dates and roll returns were not removed. The VAR is linear and symmetric and does not distinguish spillovers from good and bad volatility \citep{Barunik2016,Barunik2017} or across frequencies \citep{Barunik2018}. The residual bootstrap ignores conditional heteroskedasticity in the innovations of log realized volatility and the intervals are therefore indicative. Finally, with ten instruments the system omits sovereign bonds, emerging-market currencies, agricultural commodities and smaller cryptocurrencies, and the ``U.S. equity'' factor identified here may partly stand in for a global risk factor that a broader system would attribute differently.32