1\begin{table}[tp]2\centering3\footnotesize4\begin{threeparttable}5\caption{Forecasting realized factor variance: HAR versus GARCH(1,1) and rolling-variance benchmarks, out of sample 2016--2026 (H5)}6\label{tab:har}7\setlength{\tabcolsep}{3pt}8\begin{tabular}{lrrrrrrrrrr@{}lrr}9\toprule10Factor & $h$ & \multicolumn{3}{c}{QLIKE} & \multicolumn{2}{c}{MSE ratio to HAR} & \multicolumn{3}{c}{$R^2$ (logs)} & \multicolumn{3}{c}{DM, QLIKE} & $n$ \\ \cmidrule(lr){3-5}\cmidrule(lr){6-7}\cmidrule(lr){8-10}\cmidrule(lr){11-13} & & HAR & GARCH & Roll & GARCH & Roll & HAR & GARCH & Roll & \multicolumn{2}{l}{vs GARCH} & vs Roll & \\11\midrule12\multicolumn{14}{l}{\textit{Panel A: Out-of-sample forecast accuracy by horizon $h$ (yield days)}} \\13Level & 1 & 4.256 & 4.317 & 4.367 & 1.59 & 1.59 & 0.64 & 0.44 & 0.33 & $-$2.89 & $^{**}$ & $-$5.33 & 2,710 \\14 & 5 & 5.899 & 5.957 & 6.021 & 1.38 & 1.48 & 0.65 & 0.47 & 0.28 & $-$2.19 & $^{*}$ & $-$3.96 & 2,706 \\15 & 22 & 7.462 & 7.502 & 7.593 & 1.10 & 1.42 & 0.48 & 0.36 & 0.03 & $-$1.74 & $^{*}$ & $-$3.38 & 2,689 \\16Slope & 1 & 4.380 & 4.462 & 4.496 & 1.74 & 2.00 & 0.62 & 0.22 & 0.16 & $-$3.65 & $^{***}$ & $-$5.85 & 2,710 \\17 & 5 & 6.021 & 6.092 & 6.134 & 1.56 & 1.96 & 0.63 & 0.25 & 0.10 & $-$3.13 & $^{**}$ & $-$5.50 & 2,706 \\18 & 22 & 7.574 & 7.620 & 7.690 & 1.23 & 1.93 & 0.46 & 0.16 & $-$0.17 & $-$3.03 & $^{**}$ & $-$4.60 & 2,689 \\19Curvature & 1 & 6.220 & 6.284 & 6.400 & 1.33 & 1.65 & 0.70 & 0.58 & 0.40 & $-$3.14 & $^{**}$ & $-$5.55 & 2,710 \\20 & 5 & 7.889 & 7.946 & 8.071 & 1.42 & 1.91 & 0.74 & 0.65 & 0.43 & $-$3.40 & $^{***}$ & $-$5.56 & 2,706 \\21 & 22 & 9.473 & 9.538 & 9.718 & 1.65 & 2.74 & 0.70 & 0.60 & 0.31 & $-$3.11 & $^{**}$ & $-$3.72 & 2,689 \\22\midrule \multicolumn{14}{l}{\textit{Panel B: HAR coefficients, one-day horizon, full-sample in-sample OLS (log realized variance, 5-minute sampling)}} \\23Factor & & Constant & $\beta_{\mathrm{day}}$ & $\beta_{\mathrm{week}}$ & $\beta_{\mathrm{month}}$ & $R^2$ & & & & & & & \\24Level & & 0.234 & 0.342 & 0.425 & 0.162 & 0.64 & & & & & & & \\25Slope & & 0.257 & 0.330 & 0.427 & 0.166 & 0.61 & & & & & & & \\26Curvature & & 0.214 & 0.319 & 0.345 & 0.294 & 0.69 & & & & & & & \\27\bottomrule28\end{tabular}29\begin{tablenotes}[flushleft]\footnotesize30\item \textit{Notes:} Target: realized variance of the factor (5-minute sampling, bp$^2$) cumulated over the next $h$ yield days. HAR: regression of the log target on the daily, weekly (5-day) and monthly (22-day) averages of log realized variance, with a log-normal correction; GARCH: GARCH(1,1) with constant mean on daily CMT factor changes, $h$-step cumulative variance forecast; Roll: 22-day rolling variance of daily factor changes times $h$. All models are re-estimated monthly on an expanding window; evaluation 2016-01-04 to 2026-11-04. MSE ratio is the benchmark's mean squared error (in bp$^4$) divided by that of HAR. QLIKE $=\log f + y/f$; $R^2$ (logs) is one minus the ratio of squared log forecast errors to the variance of the log target. DM is the Diebold--Mariano statistic for the QLIKE loss differential (HAR minus benchmark; negative favours HAR) with a Bartlett HAC variance ($1.3\,T^{1/2}$ lags, at least $h-1$). Stars (own column) on the HAR-vs-GARCH statistics mark Holm-adjusted two-sided $p$-values across the nine tests of the family: $^{*}$ $p<0.10$, $^{**}$ $p<0.05$, $^{***}$ $p<0.01$; 7 of the nine are below 0.05 and 9 below 0.10. Statistics versus the rolling variance are not starred; all nine are below $-3.38$. $n$ is the number of out-of-sample forecast days. Panel B reports in-sample OLS coefficients on the daily, weekly (5-day) and monthly (22-day) averages of log realized variance with the full-sample $R^2$ (which differs from the out-of-sample $R^2$ (logs) of Panel A); no standard errors are shown.31\end{tablenotes}32\end{threeparttable}33\end{table}34