spb/wp47_uqo

1% ============================================================================2\section{Robustness and limitations}3\label{sec:robustness}4% ============================================================================56\paragraph{Targets.} The daily target was pre-specified as the CMT fit with eleven maturities. Table~\ref{tab:fidelityalt} shows7the fidelity for three alternatives. With the GSW zero curve (1--30 years) all three fidelities rise, most for the curvature,8because the factors no longer have to fit the bills. With the CMT maturities of one year and above the slope fidelity falls9rather than rises, because the one-year yield then anchors the short end alone. With maturities of two years and above the10Nelson--Siegel slope and curvature are poorly identified from seven points at a fixed $\lambda$ (daily-change standard11deviations of 15.5 and 28 basis points, a correlation of $-0.75$), the cash 2-, 5-, 10- and 30-year yields reproduce those12factors with $R^2$ above 0.95 while the futures reach 0.70 (slope) and 0.76 (curvature), and we do not use that target: for a target that lives13entirely in the 2--30-year segment the binding gap is the futures--cash basis, whereas for the primary target it is the bill14segment. The honest summary is that the futures track the 2--30-year shape of the curve, that a Nelson--Siegel fit extrapolates15a short end from that shape, and that the extrapolation is what the slope and curvature fidelities below one are made of.1617\paragraph{Instruments and weights.} A contract enters the rolling regression once it has at least 100 minute returns on 8518percent of the window's days; an earlier version of the rule (300 minutes, 95 percent), applied to the vendor's continuous19series, excluded the 2-year note and the ultra contracts because that series has sparse weeks around rolls, and reduced the20instrument set to three contracts. The weights in Appendix Table~\ref{tab:weights} vary across quarters, most for the curvature21portfolio, which is a long--short position in contracts whose daily returns are correlated at \corrDailyZnUs{} (10-year note and22bond) to \corrDailyUsUb{} (bond and ultra bond). Expanding-window weights give out-of-sample fidelities of \robExpLevel{},23\robExpSlope{} and \robExpCurv{}, within \robExpMaxDiff{} of the rolling three-year values.2425\paragraph{Sampling and noise.} The signature results of Section~\ref{sec:results-cov} are the reason the covariation,26connectedness and forecasting results use five-minute sampling. A tick-size benchmark is a lower bound on the noise---bid--ask27bounce, stale prices in Asian hours and the staleness of the 15-minute zero-fill rule all add to it---and we have not implemented28noise-robust covariance estimators. The event-window results are unaffected because thirty-minute sums average the noise out and29the placebo comparison is subject to the same noise.3031\paragraph{Timing.} The yield day ends at 15:30~ET because the H.15 yields are read then; the futures settle at 15:00~ET, and the32GSW curve is also an end-of-day object. Moving the boundary to 15:00~ET lowers the level fidelity from \robBLevelBase{} to33\robBLevelFifteen{} and moving it to 16:00~ET lowers it to \robBLevelSixteen{}, with similar losses for slope34(\robBSlopeFifteen{}, \robBSlopeSixteen{} against \robBSlopeBase{}) and curvature (\robBCurvFifteen{}, \robBCurvSixteen{}35against \robBCurvBase{}). A thirty-minute misalignment costs several points of $R^2$, which confirms that the H.15 yields are36read near 15:30 and that the futures-based factors are sharp enough to detect the timing. The FOMC release-time rule (14:15,3712:30 on press-conference days in 2011--2012, 14:00 from 2013) was checked against the minute of largest aggregate absolute return38of the 5-, 10-year and bond contracts between 12:00 and 14:45~ET on each statement day: the two coincide within two minutes on39\fomcTimeWithinTwoPct{} of the \fomcTimeN{} statements and within five minutes on \fomcTimeWithinFivePct{} (median offset40\fomcTimeMedian{} minute); on most of the remaining days the largest minute falls about 35 minutes after the statement, during the41press conference, so the rule is right about the statement and the conference is sometimes the larger event.4243\paragraph{Events.} Window shares use the day's realized variance from one-minute sampling in the denominator. With the44five-minute realized variance as denominator and the squared window change as numerator the FOMC shares are45\rvFiveFomcLevelPct{}, \rvFiveFomcSlopePct{} and \rvFiveFomcCurvPct{} for level, slope and curvature, the CPI shares46\rvFiveCpiLevelPct{}, \rvFiveCpiSlopePct{} and \rvFiveCpiCurvPct{}, and the Employment shares \rvFiveEmpLevelPct{},47\rvFiveEmpSlopePct{} and \rvFiveEmpCurvPct{}; the levels move by a few points, the Employment report remains the largest for level and slope, and two close pairs swap order48(the FOMC and CPI level shares, and the FOMC and Employment curvature shares), so the ranking of the FOMC statement against the49CPI release is not robust to the denominator while the ranking against the Employment report is.50The placebo windows are drawn from non-announcement days in our calendar, which contains only three release types; other 08:3051releases fall on some placebo days, so the placebo standard deviations for the 08:30 events include some announcement variance52and the ratios in Table~\ref{tab:eventsa} are, for those events, lower bounds.5354\paragraph{What the construction is not.} The intraday factors are changes, not levels, and are defined relative to weights that55absorb the futures basis, the cheapest-to-deliver option and duration drift. They do not see the 1- to 12-month segment or the56Nelson--Siegel short rate. They inherit the microstructure of six contracts with different tick sizes and liquidity. The57local projections and variance decompositions describe associations and spillovers in a reduced form; nothing in the paper58identifies a monetary policy shock. They are a measurement device for where and when yield-curve factor risk is realized, and59for how long an announcement move lasts, not an identification device for the sources of that risk.60