What a Treasury Yield Move Actually Measures
This chapter examines how researchers turn a scheduled macroeconomic release into a measured Treasury-yield response. The focus is the gap between the released number and the pre-release consensus, the short announcement window around it, and the information states that condition the response. Understanding those measurements lets a reader distinguish an observed interaction—how the news-response slope changes with a state—from a stronger causal claim that the state itself moved yields.
The distinction is easy to lose in market commentary. A report may say that “the payroll surprise lifted the two-year yield,” but that sentence compresses a sequence: forecasts existed before the report, the market formed a consensus, the actual number arrived, and investors revised the expected policy path. It also leaves out whether forecasters agreed, how uncertain the policy path already was, and which maturity is being discussed. The event-study design makes those hidden parts explicit.
From a release to an announcement-window response
The starting point is the surprise, not the data level. Suppose the median forecast for a release is 150,000 and the actual result is 200,000. The measured surprise is the difference between the two: 50,000 above expectations. For an inflation report, the units might be percentage points; for payrolls, they might be jobs; for GDP, a growth rate. The BIS study standardizes the surprises by the historical variability of each release so that the estimated coefficients describe responses to comparable surprise units rather than to incomparable raw measurements.
This construction reflects how markets trade. Much of the expected content of a release is already embedded in prices before publication. A high payroll number that everyone expected is not the same event as a high payroll number that no one expected. The gap between the actual figure and the consensus is an imperfect but practical measure of the new public information. It is also why the forecast distribution matters later: two releases with the same consensus gap do not necessarily carry the same information if one consensus was tightly clustered and the other was highly dispersed.
The BIS event study focuses on six scheduled U.S. releases: consumer prices, initial jobless claims, nonfarm payrolls, durable goods orders, retail sales, and GDP. It examines intraday changes in two-, five-, and ten-year U.S. Treasury futures. The yield change is measured over a 20-minute window, from five minutes before the announcement to fifteen minutes after it. The narrow window is intended to capture the repricing associated with the release rather than the many other developments that move the bond market over a full day.
That timing choice changes what the outcome means. The study is not estimating how a release changed the level of a Treasury yield over a month, a quarter, or a business cycle. It is estimating the immediate response of a particular maturity around a scheduled information event. The two-year contract is more closely tied to the near-term policy path; the five- and ten-year contracts incorporate expectations over longer horizons and can reflect changing beliefs about inflation as well as short rates. The three maturities therefore show whether a state changes only the front-end reaction or changes the shape of the curve.
A positive surprise also does not have a universal yield sign. A stronger-than-expected release that normally points toward tighter policy can raise yields. A higher-than-expected initial-claims number can instead suggest weaker demand and more policy easing, so its response can be negative. The BIS results preserve that distinction: the central question is how the sensitivity to a relevant surprise changes with the state, not whether every surprise produces the same directional move.
The diagram's order is the essential measurement logic. Forecasts and uncertainty exist before the data arrive. The surprise is formed when the actual number is compared with the consensus. The yield change is then observed in a tightly defined window. That order prevents a common mistake: treating forecast disagreement or policy uncertainty as outcomes created by the announcement itself. In the study, they are pre-existing conditions used to ask why the same surprise has a different slope in different observations.
Two kinds of uncertainty, two different questions
The most important measurement distinction is between survey forecast dispersion and short-rate uncertainty. Forecast dispersion is calculated from the cross-sectional spread of forecasts for the release. It asks how much forecasters disagree about the number that is about to be published. The BIS paper uses Bloomberg survey forecasts for this purpose. A wide spread can indicate that private signals differ, forecasting models differ, or the release is intrinsically difficult to predict.
Short-rate uncertainty asks a different question: how uncertain is the future path of short-term interest rates? The BIS measure is a forward-looking daily measure derived from Eurodollar futures and options, with the benchmark using a 12-month horizon. It is lagged by one day in the regression so that the uncertainty measure is not mechanically contaminated by the same announcement-window price move it is meant to condition. In plain language, survey dispersion concerns the quality of the information about the economic release; short-rate uncertainty concerns the market's uncertainty about what monetary policy will do with the information.
Those states can move independently. Imagine that forecasters are tightly clustered around a payroll estimate, but investors disagree about whether the central bank would respond aggressively or patiently to a result near that estimate. Forecast disagreement would be low while policy uncertainty could be high. The opposite can also occur: forecasters may disagree about a release, while the market has a settled view that the central bank will react only modestly to whatever number appears. The two variables should not be collapsed into a single label such as “volatility.” They describe different points in the path from data to policy to yields.
The regression makes that distinction visible by interacting the standardized surprise with each conditioning variable. A coefficient on the surprise gives the baseline slope for a maturity. An interaction between the surprise and disagreement asks whether that slope changes when forecast dispersion is higher. An interaction between the surprise and short-rate uncertainty asks whether the slope changes when the policy-rate path is harder to predict. The interaction is not the same as saying that a one-unit increase in disagreement or uncertainty directly moved the yield by a fixed amount.
The marginal response is built from those pieces. At a particular level of disagreement and short-rate uncertainty, the estimated response combines the baseline surprise coefficient with the two relevant interaction terms. Moving from a low-dispersion observation to a high-dispersion observation can therefore change the slope even if the underlying release surprise is identical. Likewise, a high-uncertainty observation does not have a separate “uncertainty shock” hidden inside the announcement; it changes how much weight the regression assigns to the announcement's policy-relevant information. This is why the study can compare conditional responsiveness without claiming that the conditioning state was randomly assigned.
Standardization helps with comparison but does not erase interpretation. A one-standard-deviation surprise in payrolls is not economically identical to a one-standard-deviation surprise in CPI. It is a common statistical scale, not a claim that investors attach the same meaning to each release. The release category, its usual relationship to policy, and the maturity being measured remain part of the interpretation. The coefficients answer a narrower question: within the study's construction, how does a surprise of a comparable statistical size pass through to a chosen Treasury contract under different information states?
The surrounding controls matter because the conditioning variables may reflect broader market regimes. The BIS specification includes levels of disagreement and short-rate uncertainty, quadratic terms, measures such as the VIX, MOVE, economic policy uncertainty, and a macroeconomic uncertainty index, as well as a zero-lower-bound indicator and concurrent surprises. These controls do not turn the exercise into a randomized experiment. They make the comparison more specific: conditional on a set of observed states, how does the measured news-response slope vary?
The study's high-frequency identification assumption is that the announcement surprise is the dominant new information arriving during the narrow window, and that unobserved factors correlated with disagreement or policy uncertainty do not move yields differentially during those minutes. The narrow window helps, and the paper uses a permutation placebo exercise in which surprise values are shuffled across release dates while the yields, states, and controls remain fixed. The actual interaction coefficients are generally more extreme than the shuffled-surprise estimates. That supports the view that the measured interactions are related to the realized announcement news rather than being mere products of persistent regime variables.
Still, the authors are careful about what the coefficients establish. Forecast disagreement may proxy for unobserved features of the macroeconomic environment. Lagging short-rate uncertainty addresses same-day reverse causality but not every correlation created by persistent regimes. The reported interaction coefficients are therefore best read as state-contingent sensitivity slopes. They describe how the response differs across observations with different information states; they do not, by themselves, prove that changing disagreement or uncertainty in isolation would have caused the market to move.
The sample boundaries are part of the measurement too. The main sample runs from June 1998 to September 2024, with March through June 2020 removed because Treasury-market liquidity was severely disrupted and macro surprises were unusually extreme. The pre-2020 comparison covers 1998 through February 2020, and the post-COVID comparison begins in July 2020. Several exceptionally large NFP-dispersion observations are also excluded because they would distort the standardization of disagreement. These choices improve comparability within the study, but they mean the estimates should not be treated as a timeless response function for every crisis or every release.
What a conditional slope can—and cannot—say
The measurement framework gives the reader a disciplined way to read a reported yield response. Start with the release relative to consensus. Then ask which maturity moved and over what time window. Next ask whether the relevant pre-release state involved disagreement about the data, uncertainty about the policy path, or both. Only then should the direction and size of the yield move be connected to a policy interpretation.
For example, a stronger-than-expected payroll report and a higher-than-expected initial-claims report are not symmetric tests of the same idea. The first may imply a tighter policy path; the second may imply more easing. If both responses become more sensitive when a state variable changes, the sign of the interaction can differ because the underlying policy implication differs. An interaction with short-rate uncertainty that amplifies a tightening signal need not make an easing signal positive; it can make the negative response more pronounced. Looking only at whether the coefficient has a plus or minus sign would confuse direction with sensitivity.
The maturity comparison adds another safeguard. If two-year futures react strongly but ten-year futures barely move, the market may be revising the near-term policy path without changing its long-run view. If longer maturities move as well, the release may be changing beliefs about the persistence of policy, inflation, or the broader economic regime. The event study does not by itself identify which of those components dominates, but it tells us where along the curve the repricing occurred.
This is the point at which a statistical interaction becomes an economic judgment. The estimate does not say, “uncertainty raised yields.” It says something closer to: for this type of surprise, in this sample, at this maturity, the yield-response slope was different when the pre-existing policy-rate uncertainty was higher. The direction of the surprise still matters. The release type still matters. The maturity still matters. The event window and controls still matter.
The next question is why the two conditioning variables should have opposite effects. A noisy consensus can make a release a less reliable guide to the economic state. Yet uncertainty about the future policy mapping can make policy-relevant news more valuable, because investors have more to learn about which path the central bank may follow. The measurement design separates those questions; the learning mechanism explains how they can meet in one Treasury-yield response.