1% ============================================================================2\section{Data}3\label{sec:data}4% ============================================================================56\subsection{Treasury futures}7\label{sec:data-futures}89The intraday data are one-minute bars for six CME Treasury futures: the 2-year note (ZT), the 5-year note (ZF), the 10-year note10(ZN), the ultra 10-year note (TN), the classic Treasury bond (US) and the ultra bond (UB). The bars come from a private lakehouse11built on FirstRate Data; each bar records the open, high, low, close and volume of the minute, time-stamped at the bar open in12UTC, and minutes without a trade are absent. We use the bars of the \emph{individual contracts} (every expiry, \nContractsZN{}13contracts for the 10-year note) rather than the vendor's continuous series. For each root and each CME session---18:00~ET to1417:00~ET the next day, with a one-hour maintenance break---the contract with the largest session volume is selected, and15one-minute log returns in percent are computed from consecutive bar closes of that contract within the session. No return is16ever computed across sessions or across contracts, so there is no roll splice and no adjustment. A first version of the paper17used the vendor's unadjusted continuous front-contract series; its bars are sparse in the two to three weeks before each18quarterly expiry, when the vendor's front contract is no longer the most traded, and the sparse weeks removed a third of the1908:30 releases in March, June, September and December from the event sample. The contract-based construction has no such gap:20the median number of active minutes per yield day is above 1{,}270 for every root; Table~\ref{tab:data} summarizes the panel.2122The 2-, 5- and 10-year notes and the bond are available from January 2008, the ultra bond from January 2010 and the ultra2310-year from its launch in January 2016; the three-month SOFR future (SR3), used as an additional instrument in24Section~\ref{sec:robustness}, has a median of \srThreeMedianBarsTwentyTwenty{} one-minute bars a day in 2020 and25\srThreeMedianBarsTwentyTwentyTwo{} in 2022, so it enters only from 2022. The sample ends on 6 November 2026. A segment of returns26is broken by the maintenance break and by any gap of more than five minutes between bars. A minute at which a contract has no27bar but has traded within the previous fifteen minutes is assigned a zero return; a contract that has not traded for fifteen28minutes is inactive and the cross-section for that minute is dropped. Isolated prints that are reversed in full in the next29minute are removed (\nBouncePairs{} pairs in total, \nBounceUB{} in the ultra bond and \nBounceUS{} in the bond). The resulting30panel has \nPanelMinutes{} million minute rows. The daily return of a contract is the sum of its minute returns over the yield31day defined below.3233\input{tables/tab_data}3435\subsection{Daily yields and factors}36\label{sec:data-yields}3738The daily inputs are the Treasury constant-maturity (CMT) yields published in the Federal Reserve's H.15 release and obtained39from FRED, at eleven maturities: 1, 3 and 6 months and 1, 2, 3, 5, 7, 10, 20 and 30 years. These yields are read from bid-side40quotes at approximately 15:30~ET. We therefore define the \emph{yield day} $t$ as the interval from 15:30~ET on the previous41business day to 15:30~ET on day $t$, assign every minute to the yield day that contains it, and compute all daily futures returns42over the same interval. This alignment is what makes the daily regression of Section~\ref{sec:methodology} and the intraday43cumulation comparable, and Section~\ref{sec:robustness} shows that the data detect a thirty-minute misalignment. As robustness44checks we use the zero-coupon curve of \citet{gurkaynak2007treasury}, which starts at one year and is published by the Federal45Reserve Board, and we re-estimate the factors on the CMT maturities of one year and above and of two years and above. The46effective federal funds rate (FRED) defines the regimes.4748\subsection{Announcement calendar}49\label{sec:data-events}5051Scheduled FOMC statement dates are parsed from the Federal Reserve's historical and current meeting calendars; the release time is5214:15~ET before April 2011, 12:30~ET at the eight press-conference meetings of 2011--2012 and 14:15~ET at the others, and5314:00~ET from January 2013. Four unscheduled policy statements (22 January and 8 October 2008, 3 and 15 March 2020) are kept with54their recorded times and excluded from the scheduled-event analysis. Release dates of the Consumer Price Index and of the55Employment Situation report are taken from the release-date lists of the ALFRED database, one release per calendar month (annual56revision releases are excluded), all at 08:30~ET. The 2025 lapse in appropriations delayed the September and October 202557releases; we use the actual release dates. Over 2011--2026, the period for which out-of-sample mimicking weights exist, the58calendar contains \nFomcSched{} scheduled FOMC statements, \nCpiEv{} CPI releases and \nEmpEv{} Employment releases, of which59\nFomcEventsUsed{}, \nCpiEventsUsed{} and \nEmpEventsUsed{} have complete minute coverage in the $[-5,+25)$-minute window and60enter the event tables.6162\subsection{Regimes}6364We report results for the full 2011--2026 period and for four sub-periods: January 2011 to 15 December 2015 (zero lower bound),6516 December 2015 to December 2019 (normalization), 2020--2021 (pandemic zero lower bound) and 2022--2026 (tightening and easing66cycle). Weights are estimated from 2008 onward, so the first out-of-sample quarter is 2011Q1.6768\begin{figure}[tp]69\centering70\includegraphics[width=\textwidth]{figures/fig_cartography.pdf}71\caption{\textbf{Data, pipeline and validation map.} Sources (private FinData OS lakehouse of FirstRate Data bars; public FRED,72Federal Reserve Board and ALFRED inputs with download date), the compute node on which all bar-level work is done (the dashed73boundary is the licence perimeter: no one-minute bar leaves it), the pipeline stages with the script number that implements each74(01 export; 03 return panel; 02 daily Nelson--Siegel factors; 04 quarterly out-of-sample mimicking weights; 06 minute factors and75realized measures; 07 events; 08 HAR forecasts; 13 impulse responses and connectedness), the validation loops of76Sections~\ref{sec:results} and~\ref{sec:robustness}, and the aggregate outputs that leave the node. Numbers inside boxes are sample77sizes of the final run.}78\label{fig:cartography}79\end{figure}80