TeX 62.4%
Python 37.5%
1% ============================================================================2\section{Methodology}3\label{sec:methodology}4% ============================================================================56\subsection{Realized Volatility}7\label{sec:rv}89For instrument $i$ on calendar day $t$, let $P_{i,t,j}$ denote the closing price of the $j$-th 5-minute bar, $j = 1, \ldots, M_{i,t}$, where $M_{i,t}$ is the number of bars available on that day. The 5-minute log return and the daily realized variance are10\begin{equation}11 r_{i,t,j} = \ln P_{i,t,j} - \ln P_{i,t,j-1},12 \qquad13 \RV_{i,t} = \sum_{j=1}^{M_{i,t}} r_{i,t,j}^2 .14 \label{eq:rv}15\end{equation}16If the log price is a continuous semimartingale with spot variance $\sigma_{i,t}^2(s)$ and the sampling interval shrinks, $\RV_{i,t}$ converges in probability to the integrated variance $\int_{t-1}^{t} \sigma_{i,t}^2(s)\,ds$ \citep{Andersen2003,BarndorffNielsen2002}; in the presence of jumps it converges to the integrated variance plus the sum of squared jumps \citep{Andersen2007}. At the 5-minute frequency the bias induced by microstructure noise is small for liquid instruments \citep{HansenLunde2006,Liu2015}. We report realized volatility in annualized form,17\begin{equation}18 \sigma_{i,t} = \sqrt{252 \cdot \RV_{i,t}},19 \label{eq:rv_annualized}20\end{equation}21and require $M_{i,t} \geq 20$ bars for a day to be retained. Because $\sigma_{i,t}$ is strongly right-skewed and heavy-tailed whereas its logarithm is close to Gaussian \citep{Andersen2001,Andersen2003}, the connectedness analysis is conducted on22\begin{equation}23 y_{i,t} = \ln \sigma_{i,t},24 \label{eq:logrv}25\end{equation}26which is equal to $\tfrac12 \ln \RV_{i,t}$ up to a constant. The transformation matters for the VAR because the generalized variance decomposition weights observations by their squared innovations: in levels, a handful of extreme days dominates the residual covariance matrix, and the decomposition then describes those days rather than the typical transmission mechanism. Results in levels are reported as a robustness check. Stationarity of $y_{i,t}$ is assessed with the augmented Dickey--Fuller test \citep{DickeyFuller1979} and the KPSS test \citep{Kwiatkowski1992}.2728\subsection{Vector Autoregression}29\label{sec:var}3031Let $\mathbf{y}_t = (y_{1,t}, \ldots, y_{N,t})'$ collect the $N = 10$ log realized volatilities on trading day $t$. The dynamics are approximated by a covariance-stationary VAR($p$),32\begin{equation}33 \mathbf{y}_t = \boldsymbol{\mu} + \sum_{\ell=1}^{p} \boldsymbol{\Phi}_\ell \mathbf{y}_{t-\ell} + \boldsymbol{\varepsilon}_t,34 \qquad \mathbb{E}[\boldsymbol{\varepsilon}_t \boldsymbol{\varepsilon}_t'] = \boldsymbol{\Sigma},35 \label{eq:var}36\end{equation}37estimated equation by equation by least squares \citep{Lutkepohl2005}. The lag order is chosen by comparing the Akaike, Schwarz and Hannan--Quinn criteria over $p \in \{1, \ldots, 10\}$ (\Cref{tab:lag_selection} in the appendix). The criteria select $p = 5$, $3$ and $4$ respectively; we adopt the Hannan--Quinn choice, $p = 4$, which coincides with the original specification of the paper, and report $p \in \{1, 2, 3, 6\}$ as robustness checks. The moving-average representation of \eqref{eq:var} is $\mathbf{y}_t = \boldsymbol{\mu}^{*} + \sum_{h=0}^{\infty} \boldsymbol{\Psi}_h \boldsymbol{\varepsilon}_{t-h}$ with38\begin{equation}39 \boldsymbol{\Psi}_0 = \mathbf{I}_N,40 \qquad41 \boldsymbol{\Psi}_h = \sum_{\ell=1}^{\min(h,p)} \boldsymbol{\Phi}_\ell \boldsymbol{\Psi}_{h-\ell}, \quad h \geq 1 .42 \label{eq:psi_recursion}43\end{equation}4445\subsection{Generalized Forecast-Error Variance Decomposition}46\label{sec:gfevd}4748Following \citet{DieboldYilmaz2012}, the $H$-step-ahead generalized forecast-error variance decomposition of \citet{Koop1996} and \citet{Pesaran1998} is49\begin{equation}50 \theta_{ij}(H) = \frac{\sigma_{jj}^{-1} \sum_{h=0}^{H-1} \left(\mathbf{e}_i' \boldsymbol{\Psi}_h \boldsymbol{\Sigma} \mathbf{e}_j\right)^2}{\sum_{h=0}^{H-1} \mathbf{e}_i' \boldsymbol{\Psi}_h \boldsymbol{\Sigma} \boldsymbol{\Psi}_h' \mathbf{e}_i},51 \label{eq:gfevd}52\end{equation}53where $\sigma_{jj}$ is the $j$-th diagonal element of $\boldsymbol{\Sigma}$ and $\mathbf{e}_i$ is the $i$-th unit vector. The numerator is the contribution of a one-standard-deviation shock to variable $j$---allowing for its contemporaneous correlation with the other shocks---to the $H$-step forecast-error variance of variable $i$, and the denominator is that forecast-error variance. Because the generalized shocks are not orthogonal, the rows of $[\theta_{ij}(H)]$ do not sum to one, and we normalize54\begin{equation}55 \tilde{\theta}_{ij}(H) = \frac{\theta_{ij}(H)}{\sum_{k=1}^{N} \theta_{ik}(H)},56 \qquad \sum_{j=1}^{N} \tilde{\theta}_{ij}(H) = 1,57 \quad \sum_{i,j=1}^{N} \tilde{\theta}_{ij}(H) = N .58 \label{eq:gfevd_normalized}59\end{equation}60We use $H = 10$ trading days, as in \citet{DieboldYilmaz2012}, and report $H \in \{5, 20\}$ as robustness checks. Since the decomposition converges as $H$ grows, the choice of horizon affects the level of the measures more than the ranking of the instruments.6162\subsection{Connectedness Measures}63\label{sec:measures}6465All measures are expressed in percent of forecast-error variance. The total spillover index is the share of forecast-error variance in the system that is due to cross-variable shocks,66\begin{equation}67 \TSI(H) = \frac{100}{N} \sum_{\substack{i,j=1 \\ i \neq j}}^{N} \tilde{\theta}_{ij}(H) .68 \label{eq:tsi}69\end{equation}70The directional spillover received by $i$ from all other variables and the spillover transmitted by $j$ to all other variables are the off-diagonal row and column sums,71\begin{equation}72 \mathrm{FROM}_i(H) = 100 \sum_{\substack{j=1 \\ j \neq i}}^{N} \tilde{\theta}_{ij}(H),73 \qquad74 \mathrm{TO}_j(H) = 100 \sum_{\substack{i=1 \\ i \neq j}}^{N} \tilde{\theta}_{ij}(H),75 \label{eq:directional}76\end{equation}77and the net directional spillover of $i$ is78\begin{equation}79 \mathrm{NET}_i(H) = \mathrm{TO}_i(H) - \mathrm{FROM}_i(H) .80 \label{eq:net_spillover}81\end{equation}82$\mathrm{FROM}_i$ lies between 0 and 100; $\mathrm{TO}_j$ is unbounded above and can exceed 100 when a variable explains a large share of many others; and $\mathrm{NET}_i$ is positive for a net transmitter and negative for a net receiver, with $\sum_i \mathrm{NET}_i = 0$. We express $\mathrm{NET}_i$ in percentage points. Finally, the net pairwise directional spillover from $i$ to $j$,83\begin{equation}84 \mathrm{NET}_{i \to j}(H) = 100\,\big[\tilde{\theta}_{ji}(H) - \tilde{\theta}_{ij}(H)\big],85 \label{eq:net_pairwise}86\end{equation}87is the signed edge of the connectedness network of \citet{DieboldYilmaz2014}.8889\subsection{Inference}90\label{sec:inference}9192The connectedness measures are smooth functions of $(\boldsymbol{\Phi}_1, \ldots, \boldsymbol{\Phi}_p, \boldsymbol{\Sigma})$, and we approximate their sampling distribution with a residual bootstrap \citep{Kilian1998,KilianLutkepohl2017}. Let $\hat{\boldsymbol{\varepsilon}}_t$ denote the centred least-squares residuals of \eqref{eq:var}. For each of $B = 500$ replications, residuals are resampled with replacement, an artificial sample is generated recursively from the estimated coefficients and the first $p$ observed values, the VAR is re-estimated and the measures are recomputed. We report 90\% percentile intervals. The bootstrap treats the residuals as independent and identically distributed and therefore does not capture conditional heteroskedasticity in the innovations of log realized volatility; the intervals should be read as indicative of estimation uncertainty rather than as exact. For the calendar-year sub-samples we use $B = 300$.9394\subsection{Rolling Windows and Sub-Samples}95\label{sec:rolling}9697Time variation is examined in three complementary ways. First, the VAR and the decomposition are re-estimated over rolling windows of $w$ trading days advancing by one day,98\begin{equation}99 \TSI_t(H; w) = \TSI\big(H; \mathbf{y}_{t-w+1}, \ldots, \mathbf{y}_t\big),100 \label{eq:rolling_tsi}101\end{equation}102and similarly for $\mathrm{NET}_{i,t}(H; w)$. A ten-variable VAR(4) with intercepts has 41 coefficients per equation, so that a 200-day window leaves fewer than five observations per coefficient; the resulting over-fitting biases the total index upwards and makes it noisy. We therefore report $w \in \{200, 250, 500\}$ and treat the 500-day window as the more reliable description of slow-moving trends. Second, four regime sub-samples are estimated separately with bootstrap intervals: the pre-COVID period (January 2019--February 2020), the COVID crash (March--June 2020), the 2022 rate-hiking cycle (January--December 2022) and the spot-ETF era (11 January 2024--March 2026). For the 85-day COVID window the lag order is reduced to $p = 1$; the VAR(4) estimates on the same window are reported in the appendix for comparison. Third, calendar-year estimates with $p = 2$ track the integration of the cryptocurrencies through the shares of bitcoin's forecast-error variance explained by the traditional instruments and the net pairwise spillovers between bitcoin and ES.103104\subsection{Granger Causality}105\label{sec:granger}106107For a pair of log realized volatilities $(y_{i,t}, y_{j,t})$ we test whether $p$ lags of $y_{j}$ help predict $y_{i}$ beyond $p$ own lags \citep{Granger1969}, with the regression108\begin{equation}109 y_{i,t} = \alpha + \sum_{\ell=1}^{p} \beta_\ell\, y_{i,t-\ell} + \sum_{\ell=1}^{p} \gamma_\ell\, y_{j,t-\ell} + u_t,110 \qquad H_0 : \gamma_1 = \cdots = \gamma_p = 0 .111 \label{eq:granger}112\end{equation}113We report the conventional $F$-statistic, $F = [(\mathrm{RSS}_R - \mathrm{RSS}_U)/p] / [\mathrm{RSS}_U/(T - 2p - 1)]$, and a Wald statistic based on a heteroskedasticity- and autocorrelation-consistent covariance matrix with ten lags \citep{NeweyWest1987}, which is more appropriate for volatility series. Granger causality is a statement about incremental one-day-ahead predictability, not about the economic size of the effect; we therefore also report the sum $\sum_\ell \hat{\gamma}_\ell$, which measures the long-run response of $y_i$ to a unit change in $y_j$ in this bivariate regression.114115\subsection{Intraday Periodicity and Session-Ordered Predictability}116\label{sec:sessions_method}117118The intraday analysis uses the 5-minute returns directly. Exchange time stamps are converted to Coordinated Universal Time (UTC) with the civil daylight-saving rules of the exchange's time zone, and three sessions are defined in UTC: Asian (00:00--08:00), European (08:00--14:30) and U.S. (14:30--21:00). For each instrument the hourly profile of volatility is the mean absolute 5-minute return by hour of the day, and the session realized variance on day $t$ is119\begin{equation}120 \RV^{s}_{i,t} = \sum_{j \in s} r_{i,t,j}^2, \qquad s \in \{\text{Asian}, \text{European}, \text{U.S.}\},121 \label{eq:session_rv}122\end{equation}123retained when the session contains at least twelve bars. The share of daily variance generated in session $s$ is the average over days of $\RV^{s}_{i,t} / \sum_{s'} \RV^{s'}_{i,t}$.124125The chronological ordering of sessions within a UTC day makes it possible to test whether volatility realized in an earlier session helps predict volatility realized in a later one. We estimate126\begin{equation}127 \ln \RV^{\mathrm{US}}_{i,t} = \alpha + \sum_{\ell=1}^{4} \beta_\ell \ln \RV^{\mathrm{US}}_{i,t-\ell} + \sum_{\ell=0}^{3} \gamma_\ell \ln \RV^{s}_{j,t-\ell} + u_t,128 \label{eq:session_predictive}129\end{equation}130where $i$ is ES, NQ or the VIX, $j$ is bitcoin and $s$ is the Asian or the European session of the same day, and test $\gamma_0 = \cdots = \gamma_3 = 0$ with a HAC Wald statistic. The reverse direction replaces the dependent variable by bitcoin's Asian-session variance on day $t$ and the predictor by the U.S.-session variance of ES, NQ or the VIX on the previous trading day. A final regression asks whether bitcoin's weekend variance (Saturday plus Sunday) predicts the Monday U.S.-session variance of ES once Friday's ES and bitcoin variances are controlled for. These regressions complement the daily VAR by exploiting an ordering of information that daily data cannot provide.131132\subsection{Network Representation}133\label{sec:network_method}134135The full-sample decomposition is represented as a directed graph whose nodes are the ten instruments and whose edges are the net pairwise directional spillovers of \eqref{eq:net_pairwise}, drawn from the net transmitter to the net receiver when $|\mathrm{NET}_{i \to j}|$ exceeds the 60th percentile of all pairwise values. Node size is proportional to $\mathrm{TO}_j$ and nodes are coloured by asset class.136