TeX 62.4%
Python 37.5%
1\section{Audit of the Panel Construction}2\label{sec:appendix_audit}34\Cref{tab:audit_construction} reproduces the full-sample estimates of the May 2026 version of this paper and traces the effect of the two corrections made in the present version. The first row replicates the original computation exactly from the archived daily realized-volatility file: a panel of all 3{,}651 calendar days between 11 March 2016 and 11 March 2026, with missing values forward-filled for up to five days and realized volatility in levels. The second row removes the forward filling and the weekend rows, leaving the 2{,}514 trading days on which all ten instruments have an own realized volatility. The third row applies the log transformation to that panel and is the baseline of \Cref{sec:results}.56\begin{table}[htbp]7\centering8\caption{Effect of the panel construction on full-sample connectedness, VAR(4), $H = 10$.}9\label{tab:audit_construction}10\adjustbox{max width=\textwidth}{%11\begin{threeparttable}12\input{tables/tab_audit_construction}13\begin{tablenotes}[flushleft]14\footnotesize15\item \textit{Notes:} TSI in percent; the other columns are net directional spillovers in percentage points. The first row reproduces the published figures of the May 2026 version (TSI 55.44\%, ES $+26.80$, NQ $+19.72$, VIX $-23.37$, BTC $+1.23$). Of the 3{,}651 rows of the original panel, 1{,}044 were Saturdays or Sundays and 1{,}137 contained at least one forward-filled value. Source: \path{results/rev_audit_construction.csv}.16\end{tablenotes}17\end{threeparttable}}18\end{table}1920The forward filling lowers the total index by about five percentage points relative to the trading-day panel in levels (55.4\% against 50.0\%) while compressing the net spillover of ES from $+41.6$ to $+26.8$ and of NQ from $+29.8$ to $+19.7$; the near-coincidence of the original index (55.4\%) with the revised baseline (54.8\%) is the result of two offsetting changes, forward filling and the log transformation, and not of robustness. The forward-filled panel also drove the own shares of bitcoin and ether above 77\% and their cross-asset shares below 0.5\%, because on weekend rows the innovations of the eight traditional series are identically zero while those of the cryptocurrencies are not. In the rolling analysis, the original 200-observation window covered 200 calendar days, or about 143 trading days, and the index reached 81.5\% on 6 February 2018 against 56.4\% on the trading-day panel (\Cref{tab:rolling_summary}).2122\begin{table}[htbp]23\centering24\caption{Summary of rolling total spillover indices.}25\label{tab:rolling_summary}26\adjustbox{max width=\textwidth}{%27\begin{threeparttable}28\input{tables/tab_rolling_summary}29\begin{tablenotes}[flushleft]30\footnotesize31\item \textit{Notes:} Mean, minimum and maximum of the rolling total spillover index (\%) with the dates at which the extremes occur. The first row is the series published in the May 2026 version; the others are VAR(4), $H = 10$ estimates on log realized volatility over the trading-day panel. Source: \path{results/rev_rolling_summary.csv}.32\end{tablenotes}33\end{threeparttable}}34\end{table}3536\section{Lag-Order Selection}37\label{sec:appendix_lags}3839\Cref{tab:lag_selection} reports the information criteria for VAR($p$) models of log realized volatility on the full trading-day panel. The Schwarz criterion selects $p = 3$, the Hannan--Quinn criterion $p = 4$ and the Akaike criterion $p = 5$; the baseline uses $p = 4$.4041\begin{table}[htbp]42\centering43\caption{Information criteria for the lag order of the VAR, log realized volatility, 2016--2026.}44\label{tab:lag_selection}45\adjustbox{max width=\textwidth}{%46\begin{threeparttable}47\input{tables/tab_lag_selection}48\begin{tablenotes}[flushleft]49\footnotesize50\item \textit{Notes:} Akaike (AIC), Schwarz (BIC) and Hannan--Quinn (HQ) criteria for VAR($p$) with a constant, $p = 1, \ldots, 10$, computed on a common estimation sample; the minimum of each criterion is in bold. Source: \path{results/rev_table_lag_selection.csv}.51\end{tablenotes}52\end{threeparttable}}53\end{table}5455\section{Full Connectedness Tables by Regime}56\label{sec:appendix_regimes}5758\Cref{tab:regime_precovid,tab:regime_covid,tab:regime_ratehikes,tab:regime_etf} report the complete variance-decomposition tables for the four regimes summarized in \Cref{tab:regimes}. The COVID window uses $p = 1$; the others use $p = 4$.5960\begin{table}[htbp]61\centering62\caption{Connectedness table, pre-COVID regime (2019-01-01 to 2020-02-28), $T = 292$, VAR(4).}63\label{tab:regime_precovid}64\adjustbox{max width=\textwidth}{%65\input{tables/tab_regime_precovid}66}67\end{table}6869\begin{table}[htbp]70\centering71\caption{Connectedness table, COVID crash regime (2020-03-01 to 2020-06-30), $T = 85$, VAR(1).}72\label{tab:regime_covid}73\adjustbox{max width=\textwidth}{%74\input{tables/tab_regime_covid_crash}75}76\end{table}7778\begin{table}[htbp]79\centering80\caption{Connectedness table, rate-hiking regime (2022-01-01 to 2022-12-31), $T = 254$, VAR(4).}81\label{tab:regime_ratehikes}82\adjustbox{max width=\textwidth}{%83\input{tables/tab_regime_rate_hikes}84}85\end{table}8687\begin{table}[htbp]88\centering89\caption{Connectedness table, spot-ETF era (2024-01-11 to 2026-03-11), $T = 544$, VAR(4).}90\label{tab:regime_etf}91\adjustbox{max width=\textwidth}{%92\input{tables/tab_regime_spotetf_era}93}94\end{table}9596\section{Replication}97\label{sec:appendix_replication}9899The daily realized volatilities are produced by \path{scripts/01_analysis.py} from the vendor's 5-minute bars (about ten minutes and 8~GB of memory); all connectedness estimates, bootstrap intervals, robustness checks, rolling windows, regime and calendar-year tables and \Cref{fig:connectedness_heatmap,fig:rolling_spillover,fig:net_spillovers_regimes,fig:net_directional_btc_es,fig:network_graph,fig:yearly_integration} are produced by \path{scripts/02_connectedness_revision.py} from \path{results/daily_rv_all.csv} in about one minute; the intraday analysis of \Cref{sec:sessions} is produced by \path{scripts/03_intraday_sessions.py} from one-minute bars of the same vendor in about half a minute. Every table in the paper is generated by these scripts as a LaTeX fragment in \path{paper/tables/}.100