Presidential calls for easier Fed policy bring a real, small repricing in front-month fed funds futures. Baretti, Collodel and Zacché put it at roughly a basis point on days when a sitting president publicly presses for easing. The paper does not establish how much of that move belongs to the pressure rather than the day's other news. The authors acknowledge that a residual information component "cannot be fully excluded within a reduced-form design." They answer with an interaction based on Board alignment: when the president's party holds a Board majority, each dovish event adds -1.56 bp at FFc1. They call this "a pattern the information channel cannot generate," a claim that asks too much of the test. Nor does the paper establish that a reader of the wire could have captured the move.
Counting the pressure
The premise is straightforward. Public demands from those who appoint the Board convey news about the policy path if Fed independence is less than complete. Futures should respond.
The authors collected Reuters newswires from Factiva covering 1988 to 2025, retaining articles that jointly mention a political actor, the Fed and monetary policy. Three open-weight models (Qwen3-235B, Llama-4-Maverick, DeepSeek-V3) operated at temperature zero. A 2-of-3 vote governed each step. First the models combined duplicate stories into events, then retained statements attributable to incumbent officials, and finally classified events as dovish, hawkish or not relevant. The roughly 11,500 raw articles yielded 923 relevant events and 666 labelled ones: 579 dovish and 87 hawkish.
The regression relates daily changes in fed funds futures to the daily number of dovish events, including two lags of the dependent variable. Controls comprise FOMC and month-turn dummies, monthly inflation and unemployment gaps, and selected Bauer and Swanson predictors.
The front-month estimate is -0.91 bp per dovish event (SE 0.33). Across the next three contracts, it shrinks to -0.62, -0.57 and -0.50 bp. A hold-or-cut-25 mapping translates the front-month response into a 3.7 percentage point increase in cut probability. Significance is borderline through FFc7 and disappears thereafter. The hawkish coefficients are positive and insignificant, unsurprisingly given 87 events. Dividing the sample by speaker leaves only the president with a detectable response, -0.96 bp at FFc1.
The response stops at the front of the strip
Extending the curve to Treasury futures (TU, FV, TY, US) produces coefficients near zero or positive, with none significant. The units matter here. The fed funds results measure rate changes; the Treasury results measure price returns in basis points, where a positive return means lower yields. Daily TU price returns have an SD of 9.78 bp, making a sub-basis-point yield response difficult to detect. Expressing the trade through 2-year notes gives a desk a different exposure.
Is this a Trump result?
No. Excluding both Trump terms increases the FFc1 estimate to -1.58 bp, with significance extending to six months. The authors describe the increase as "approximately 50% larger." At FFc1 the ratio is 1.74. At FFc2, the estimate more than doubles, moving from -0.62 to -1.37.
Habituation is their explanation. A single day accounts for 79.5% of non-Trump pressure streaks, compared with 48.4% during the Trump years. The authors interact the event count with dovish articles over the prior three days. As recent pressure accumulates, the marginal response declines and becomes insignificant at the highest intensities.
Saturation fits that result. So does administration-level heterogeneity. The authors acknowledge that distinguishing the explanations "is not straightforward in a reduced-form framework," and describe the split-sample evidence as "suggestive rather than conclusive." They nevertheless present the intensity interaction as evidence of saturation. It leaves the question open: the estimate uses the pooled sample, and most high-intensity days belong to the Trump years.
Pressure and news share the same daily window
Precise timestamps are available only from mid-1997, leaving the full design close-to-close. A presidential complaint on a soft payrolls morning shares an observation with the payrolls release. Monthly gaps carried forward and a few Bauer and Swanson series supply the controls.
Classification provides another connection to the macro state. When criticism gives no direction, the prompt defaults to "Criticism during RECESSION/HIGH UNEMPLOYMENT -> dovish." The economic backdrop therefore supplies part of the label.
The stress tests weaken the findings only beyond the front month. The baseline already includes a payrolls dummy, partly answering the payrolls example. Adding release dummies for CPI, advance GDP, PPI and retail sales leaves the estimates unchanged: -0.90 bp at FFc1 (SE 0.33) and -0.50 at FFc4 (SE 0.25). FFc1 remains significant across all nine variants, ranging from -0.82 to -0.98 bp.
Controlling for the Citi surprise index, whose sample starts in 2005, leaves FFc1 significant at -0.88 bp. FFc2 through FFc4 decline to -0.52, -0.42 and -0.39, respectively, and all become insignificant. The authors note that this control halves the sample. At FFc2, the SE rises from 0.27 to 0.36 while the estimate changes only from -0.62 to -0.52. With year-by-month fixed effects, FFc1 remains significant at -0.85 (SE 0.36). FFc4 comes in at -0.42 (SE 0.31), without significance.
The Board-alignment test is clever. Each dovish event adds -1.56 bp at FFc1 when the president's party commands a Board majority. Calling this a result the information channel cannot generate goes too far. Alignment changes across only a handful of eras, and the Trump split already demonstrates that the per-event coefficient varies by era. I take the interaction as supporting evidence, with era effects still a possible explanation.
The treatment of label noise deserves credit. Human validation finds 45 of 51 disagreements in articles that raters judged not relevant and models labelled directional. The authors construct a placebo using 678 draws of politician-and-Fed articles without pressure. Its distribution centres near -0.1 bp; the actual coefficients lie in the far left tail. Generic coverage of Fed politics does not reproduce the finding.
The credibility conclusion survives only partly. Full-sample 2Y inflation swaps respond by -0.22 bp per event (SE 0.27), tightly bounding any increase. Excluding the Trump years gives -1.2 bp (SE 1.07). The downside band is wide, though it still excludes a rise as large as the 1.58 bp no-Trump futures response. For MOVE, the no-Trump point estimate is +1.17 (SE 0.87), compared with -0.23 (SE 0.21) over the full sample. Neither is significant, and their signs conflict.
A wire reader still needs an entry point
The mean shift of 0.91 bp compares with a daily SD of 4.96 bp in the Kuttner-scaled front month. Including controls, the regression explains 1.8% of daily front-month variance. There are 0.054 dovish events on the average day. Measurement covers the day of publication, including any movement before the story appeared.
The local projections also include the event day. Their cumulative bands consequently combine the same-day response and subsequent drift, without a separate test of the post-day increment. The sole hint of drift is in FFc5 and FFc6, which become significant only after two to three trading days.
In the pricing-error regressions, each article brings futures 0.5 to 0.8 bp closer to the eventual outturn. Unconditionally, this represents 15 to 25% of the within-maturity SD. On days when the futures rate declines, the figure is 35 to 65%, a selection the authors acknowledge captures the upper tail. They call the pricing-error findings "suggestive evidence of accommodation rather than... a causal claim." We did not find a trading rule, costs or an out-of-sample test in the paper.
We could not backtest it either.
Our contract coverage excludes fed funds futures, and we have neither the Reuters archive nor its publication times. The paper's sample also begins years before our price history.
An intraday rerun using the post-1997 timestamped events would change my view if it showed repricing beginning after the wire time.