← Taylor Rule

The Fed against its rules, rolling 12 and 24 months ahead

An Econvitals research note, October 24, 2026. Prepared with AI, directed by Jason Furman; the companion piece is the Occasional Post of October 26, 2026, and the live tool is the Taylor Rule page.

Three rolling versions of the Econvitals “Fed minus rule” comparison, every set of FOMC projections since January 2012 and every scheduled FOMC meeting since 1960, scored 12 and 24 months on, with the realized-data versions run on the data as it was known at the time; and the distance between the funds rate and each rule's level, today and at every meeting since 1961. Data as of 2026-10-23; the latest projections are September 2026.

Contents: Summary · What was built · 1. The committee against its own forecast revisions · 2. The funds rate against the realized data · 3. How far the rate is from the rules' levels · 4. Projections against outcomes · 5. How unusual is the gap now, and what it would take to be typical · 6. Which perspectives have worked best, and what they imply now · 7. Does the Fed have a reaction function, and where to read it · 8. Which rule, and what to put on the page · Methods and data · Appendix: the series

Summary

What was built

The Econvitals Taylor Rule page compares one earlier set of projections with the latest, for one year-end. Everything here rolls that comparison through time and adds two cousins:

• 1. The committee against its own forecast revisions (the page's method, rolled). For every set of projections, take the set 12 months later (and 24 months later), and for the nearest year-end both cover compute the change in the median funds-rate dot against the change a rule with no smoothing implies from the revisions to the core PCE and unemployment projections for that year-end plus the move in the longer-run rate: Δdot − [Δr_LR + 1.5 Δπ − 2 (Δu − Δu_LR)]. Positive means the committee moved its path by more than its own forecast changes call for (a hawkish deviation), negative by less. The sample is every SEP from January 2012, when the funds-rate projections began, to September 2026: 58 sets, 52 twelve-month pairs and 49 twenty-four-month pairs. The medians before September 2015, when the Fed began publishing one, are computed from the per-participant submissions in the Fed's SEP compilations; the method reproduces the published September 2015 medians exactly (see Methods).
• 2. The funds rate against the realized data, as known at the time (a delta Taylor rule). For every scheduled FOMC meeting since 1960, take the meeting nearest 12 (24) months later and compare the change in the target rate with 1.5 × the change in 12-month core inflation minus 2 × the change in the unemployment rate, each measured on the data in hand at the two meetings in the vintage current on each meeting day (ALFRED). Real-time core PCE exists from August 2000; the NSA core CPI is never revised and carries the real-time history back to 1960; today's vintage of core PCE is run alongside for comparison. 651 twelve-month pairs and 643 twenty-four-month pairs. Variants use the Taylor weight on slack, CBO's output gap, and changes in r* and u*.
• 3. The level: how far the rate is from each rule. At every scheduled meeting since 1961 and today, the target rate minus the level prescribed by the Taylor (1993) rule, the balanced approach, the balanced approach with shortfalls, the Board's inertial rule and the Monetary Policy Report's first-difference rule, with the neutral real rate either fixed (Taylor's 2 percent) or rolling (Laubach–Williams, Holston–Laubach–Williams, or the committee's own longer-run funds rate less 2, each as known at the time where vintages exist), CBO's or the committee's u*, unemployment or CBO's output gap as the slack measure, on today's vintage and as known at the time; then where today's gap sits in that history.
• 4. Projections against outcomes. Each SEP's projection for next year (and the year after) against what the year delivered: the actual year-end rate minus the dot, against the rule run on the inflation and unemployment surprises. This is the page's “outcome” scoring, rolled.

For each version the questions are the same: how well does the rule track what the Fed did (correlation, the slope of the Fed's move on the rule's, R², the root-mean-square miss against the benchmark of assuming no change at all); what the Fed's own fitted weights on inflation and slack are; whether deviations are systematic (a bias, an asymmetry between tightening and easing, a dependence on where the rate started relative to the rule's level); and where the latest reading sits in the distribution. Rolling pairs overlap, so neighboring observations are not independent; standard errors are Newey–West and the December-to-December subset is reported as the non-overlapping check.

1. The committee against its own forecast revisions

Three things stand out. First, the pandemic pairs are not a reaction-function story: the June 2019 → June 2020 pair for end-2020 reads +10.8 because the rule asked for −12.8 points of cuts (the unemployment projection for 2020 jumped +5.6 points) and the floor allowed −2.0; the June 2020 → June 2021 pair for end-2021 reads −6.1 because the forecasts recovered and the dot stayed at zero. Every statistic below is shown with and without them. Second, the 2021 → 2023 pairs at 24 months read +3 to +5: the committee raised its end-2023 path by four to five points while its 2023 inflation and unemployment projections moved only enough to justify between a quarter point and a point and a half. That is the delta rule's blind spot (the path was far below the rule's level, and the move was catch-up), not evidence of hawkishness relative to the rule. Third, since the September 2022 anchor every 12-month reading has been negative, 13 in a row averaging −1.10: through the 2024–25 easing the committee cut its year-end paths by more than its forecast revisions implied, and through 2025–26 it has raised them by less than its inflation revisions imply.

How well the rule tracks the dots

Each row reruns the comparison for every pair of projections 12 months apart (24 where stated), scoring the nearest year-end both cover unless stated. “Slope of Fed on rule” is the regression of the change in the median dot on the change the rule implies; 1 would mean the committee moves its path exactly as the rule says, 0 that it ignores the revisions. The fitted weights come from regressing the dot change on the two forecast revisions directly; the balanced approach itself puts 1.5 on inflation and −2 on unemployment, the Taylor rule 1.5 and −1. Standard errors are Newey–West, allowing for the overlap between neighboring pairs. RMSE of “no change” is the error from assuming the dot does not move at all. The mean gap is shaded red when the Fed has on average moved by less than the rule (or cut by more), green the other way.

Read across the first three rows: including the pandemic the rule's miss is more than twice the miss of assuming no change; excluding it the rule beats no change by about a third and explains two-thirds to seven-tenths of the variance. The slope of 0.74 says the committee moves its path about three-quarters as much as the balanced approach says it should. The fitted weights explain why: 1.56 on the inflation revision, right on the rule's 1.5, but −1.14 on the unemployment revision against the rule's −2. Put the Taylor (1993) weight of 1 on slack and the slope rises to 0.83, the R² to 0.74 and the miss falls to 0.62; it is the better description of the committee on every sample cut, at 24 months as well as 12 (R² 0.61 against 0.46). Holding the longer-run values fixed raises the slope to 0.96 but leaves a larger average gap (−0.61) and a larger miss: the committee's longer-run rate fell through 2012–19 and has risen since 2024, and its year-end dots moved with it, so a rule that ignores those moves reads the dots as biased. Headline PCE is more correlated with the dots (R² 0.76) but the committee moves only 0.58 of what headline revisions imply and the miss is larger: it looks through the energy component. December-to-December, the cleanest and smallest sample (11 pairs without the pandemic), gives the same picture: slope 0.87, R² 0.72, weights 1.79 and −1.56.

Scoring the following year-end instead of the nearest (for the Sept 2025 → Sept 2026 pair, end-2027 rather than end-2026) makes the rule much worse (R² 0.14 balanced, 0.45 Taylor). Two reasons: the following year's projections move less, so there is less to explain, and the committee's two-year-ahead dots carry its view of where neutral is as much as its forecast, so they move with the longer-run rate and the inflation projection and hardly at all with unemployment (fitted weights 3.25 and −0.48). The nearest year-end is where the reaction function lives.

What moved the rule, pair by pair, since 2021

From mid-2021 to mid-2022 the inflation revisions were enormous (+2 to +2.6 points on the next year's core PCE) and the committee delivered roughly the implied three to four points; the deviations were small or positive. From late 2022 the pattern reverses. In the 2023 → 2024 and 2024 → 2025 pairs the unemployment projections were revised down (the labor market held up better than the committee expected, which the rule reads as a reason to tighten) and inflation projections drifted up, while the committee cut its year-end paths: gaps of −0.9 to −1.5. In the 2025 → 2026 pairs the inflation revisions dominate: core PCE for 2026 was revised up 0.8 between September 2025 and September 2026, worth +1.2 under the rule, the unemployment revision added +0.6 and the longer-run rate +0.2, and the dot rose 0.7.

The most dovish and most hawkish pairs

The 12-month pairs with the largest gaps either way, pandemic pairs left out. The pandemic pairs themselves: June 2019 → June 2020 for end-2020 reads +10.8 (the rule wanted −12.8 as the unemployment forecast jumped 5.6 points; the Fed cut 2.0 to the floor), and June 2020 → June 2021 for end-2021 reads −6.1 (the rule wanted +6.1 as the forecasts recovered; the dot stayed at the floor).

December to December

December-to-December, scoring the later December's current year, whose Q4 values are nearly known by then: the cleanest version, and the smallest sample. 2019 → 2020 is the pandemic pair.

Does the committee close gaps to the rule's level?

A rule written in changes assumes the starting point was right. It was not always: in September 2015 the end-2016 dot was 1.40 percent while the balanced approach run on the committee's own 2016 forecast and longer-run values gave 3.25. Regressing each pair's deviation on that starting gap (the dot minus the rule's level, both from the anchor's own numbers) gives a coefficient of −0.37 (s.e. 0.04) over 12 months and −0.86 over 24, excluding the pandemic and the floor: when the dot starts a point above the rule's level, the committee lowers it by about a third of a point more than its revisions imply over the following year and most of a point over two. The starting gap alone explains 64 percent of the 12-month deviations. The committee behaves as if it is converging on the level rule, with a half-life of about two years. This matters for reading the current gap, below.

2. The funds rate against the realized data, as it was known at the time

The page runs its rules on today's data with real-time release timing. For a reaction-function question that is the wrong vintage: the committee reacted to the unemployment rate and the inflation rate in front of it, and both have since been revised. Here every meeting's inputs are the values in the vintage current on the decision day, from ALFRED. The unemployment rate is rarely revised by much (a revision of 0.1 or more at 34 percent of meetings, at most 0.4); 12-month core PCE inflation often is (an average absolute revision of 0.22 points between what the committee saw and what BEA now prints for the same month, 36 percent of meetings off by a quarter point or more, the largest 1.0 point). ALFRED's core PCE vintages begin in August 2000, and the price index the Fed watched before that was a different, fixed-weight animal; the NSA core CPI, which the BLS never revises, is the one core inflation measure with a genuine real-time history back to 1960, so it carries the long series below.

Pair by pair, the vintage moves the 12-month reading by 0.40 points on average and by half a point or more in 28 percent of the pairs since 2001; the largest difference is 2.3 points, for Jun 2009 → Jun 2010, when the 2009 comprehensive revision took the core PCE inflation the committee had seen in 2008 and 2009 down by about a point. What the vintage does not do is change the fit. The table below runs the same pairs both ways: correlation, slope, R² and the Fed's fitted weights barely move. What it does change is particular readings, and the most recent one is a case in point: at the September 2026 meeting 12-month core PCE stood at 3.3 percent (July, as first published) against 3.0 on today's vintage; a year earlier it stood at 2.9 against 2.8; two years earlier at 2.6 against 2.8. As the committee experienced it, core inflation rose 0.5 points over the year to September 2026 and 0.7 over two years; as BEA now has it, 0.1 and 0.2.

Balanced-approach weights, unemployment as the slack measure, floor pairs excluded. "As known then" means the unemployment rate and the price index as they stood in the vintage current on each meeting day (ALFRED), with the months the page already picks by release date; today's vintage is the same months as published now. NSA core CPI is never revised, so its real-time and current values coincide to the rounding of the index.

Over the whole span the delta rule is a middling description of policy: away from the floor and as known at the time, correlation 0.53, slope 0.39, R² 0.28, a miss of 2.56 points against 2.17 for no change (on today's core PCE: 0.62, 0.57, 0.38, 1.98). It is a much better description of some eras than others. Under Greenspan it is good (R² 0.63 as known then, 0.72 on today's core PCE; slope 0.80; fitted weights 0.50 on inflation and −2.03 on unemployment; a miss of 1.02 against 1.60): this is the period Taylor's rule was fitted to and the period in which the funds rate moved with the data rather than toward a distant level. Under Martin the Fed under-reacted to both (weights 0.37 and −1.05); under Burns and Miller it reacted strongly to unemployment (−2.10) and weakly to inflation (0.44), the textbook account of the 1970s; under Volcker the misses are huge because the level of rates was being reset (on today's core PCE the weights flip to 1.32 on inflation and −0.82 on unemployment; on the real-time CPI the inflation weight is only 0.26, partly because the CPI of 1979–82 still carried mortgage interest costs and the committee did not chase that one for one). Since 2014 the rule has had no explanatory power at all. Yellen's 18 non-floor pairs have an R² of 0.14: liftoff proceeded at a pace set in advance, a quarter point in each of 2015 and 2016, three-quarters in 2017, a point in 2018, whatever the data did within the year. Powell's 37 non-floor pairs have an R² of 0.03 and a mean deviation of +1.52: the 2022–23 pairs are +3.5 to +5 (catch-up), the 2024 pairs are near zero and the 2025–26 pairs are mostly −0.5 to −1.7 (cuts, then a single hike, against rising core inflation and falling unemployment as the committee saw them).

Twelve-month pairs of scheduled FOMC meetings, the rule run on the data as known at each meeting: the unemployment rate in the vintage current that day and the 12-month change in the NSA core CPI, which is never revised (the only core inflation measure with a real-time history back to 1960; real-time core PCE begins in August 2000 and is in the next table). The era is the chair at the first meeting of the pair. “Excluding the zero floor” drops any pair with the target at or below 0.25 percent at either end. The catch-up column regresses the deviation on the gap between the funds rate and the balanced-approach level at the start of the pair: a negative number means the Fed closes part of that gap over the following year, which a rule written in changes cannot see.

The variants

1990–2026, excluding floor pairs. The fitted weights are the same across the unemployment versions because they come from the same regression of the Fed's change on Δπ and Δu.

Across variants the fit is similar and the message the same. Since 1990 the Fed's 12-month moves carry a weight of about −2 on the change in unemployment, the balanced approach's own, and a weight of about zero (slightly negative) on the change in trailing core inflation, on either vintage. The result is not that the committee ignores inflation; it is that it reacts to its inflation forecast (section 1 finds a weight of 1.5 on the forecast revision), and over 1990–2019 the trailing twelve-month rate carried almost no news for the forecast. Using CBO's output gap instead of unemployment fits a little worse (R² 0.30 against 0.42); adding changes in r* and u* makes almost no difference.

Asymmetry: tightening and easing

Averages in percentage points. The question is whether the shortfall against the rule is symmetric: does the Fed under-deliver on tightening and over-deliver on easing, or damp both?

Both versions damp in both directions, which is the familiar interest-rate smoothing: a 12-month window is shorter than the time the committee takes to get where its rule points. On the dots the damping is one-sided: when the revisions call for easing the committee delivers nearly all of it (−0.62 against −0.73); when they call for tightening it delivers half (+0.73 against +1.41). On realized data since 1990 the Fed delivers about 71 percent of the implied tightening and 70 percent of the implied easing. In the Greenspan era alone the two were 81 and 101 percent: symmetric.

Levels: the delta rule's blind spot, measured

The slope is −0.47 per point at 12 months and −0.72 at 24 (−0.49 and −0.76 as known then): starting a point below the rule's level, the Fed tightens about half a point more over the next year than the change in the data calls for, and about three-quarters of a point more over two years. The 2022–23 pairs sit at the far left (gaps of −7 to −10 points at the start, deviations of +4 to +5) and are the clearest case, but the relation holds inside the Greenspan points too. A delta rule that adds a term for the starting level gap is in effect the Board's inertial rule with the smoothing parameter estimated; the estimate here is in the same range as the Board's 0.85 a quarter, which closes about half of a level gap in a year. Section 3 takes the level gap itself as the object.

Today's reading, every variant

“Today” pairs the meeting nearest 12 (24) months before the October 28, 2026 meeting with today's data (2026-10-23) and the current target midpoint of 3.875 percent; on today's data the two vintages coincide at the near end. Percentiles are against all non-floor pairs since 1990 at the same horizon (for the real-time core PCE rows, since 2001).

As the committee saw it, over the twelve months to the September 2026 meeting it cut a quarter point net (two quarter-point cuts in late 2025, then the September hike) while 12-month core PCE rose from 2.9 to 3.3 and unemployment fell from 4.3 to 4.1, so the rule said tighten +1.10: a gap of −1.35, the 9th percentile of the non-floor pairs since 2001 and about 1.0 standard deviations below their mean. On today's vintage, which has the same window's inflation rising only 0.1, the gap is −0.86 (20th percentile since 1990). Bring the window to today (the December 2025 meeting to the October 28, 2026 meeting on today's data) and the gap is −0.42 either way, the 36th percentile: the September hike closed part of it. On the output gap the rule asks for less (+0.19) because CBO's gap barely moved, so the deviation is −0.44. On core CPI, which fell from 3.1 to 2.4 over the same year, the rule says ease and the committee looks +0.35 hawkish. The 24-month window is more dovish on every variant except CPI, because it includes the full 2024–25 easing against data that, net, called for a little tightening: −2.29 as known then (the 3rd percentile), −1.46 on today's vintage.

3. How far the rate is from the rules' levels

Sections 1 and 2 ask whether the Fed moved as a rule would have. This section asks the simpler question people usually mean: where is the rate relative to the level each rule prescribes, and is today's distance unusual? At every scheduled meeting since 1961 (when the Laubach–Williams r* begins) the target set at the meeting is compared with the level the rule gives on the data in hand that day; today's reading uses the current target of 3.875 percent and the data as of 2026-10-23. The rules are the Board's own definitions: Taylor (1993), r* + π + 0.5(π − 2) + 0.5 × gap; the balanced approach, the same with 1 on the gap; the balanced approach with shortfalls, which drops the gap term when unemployment is below u*; the inertial rule, 0.85 on the rate going in (per quarter, converted to the interval between meetings) and 0.15 on the balanced approach; and the Monetary Policy Report's first-difference rule, the rate going in plus 0.5(π − 2) plus the one-year change in the unemployment gap. The gap is Okun × (u* − u) with Okun 2, or CBO's output gap where stated.

The history has a shape everyone knows and a fact fewer people carry around. The shape: the rate ran far below both rules through the Great Inflation (the median Taylor gap was −3.01 under Martin and −4.53 under Burns and Miller, and the recession meetings of 1974–75 took it to −10.4 in Jan 1975), far above them under Volcker (median +1.06), close to them under Greenspan (median −0.60, with the 2003–04 episode at −2 to −3), pinned below them by the floor in 2009–15 and 2020–21, and in 2021–22 about as far below the rules as at any time in the record (−9.5 on the Taylor rule in May 2022, −10.2 on the balanced approach). The fact: on these inputs the rate has been below the Taylor rule at 76 percent of meetings since 1961 and 78 percent of non-floor meetings since 1990, with a median shortfall of 1.08 points; against the balanced approach, 73 percent and 0.98. A funds rate a point or so below the unsmoothed rules is the normal state of affairs, not a deviation. Today's −1.48 against the Taylor rule is the 38th percentile of the post-1990 distribution and −1.67 against the balanced approach the 29th: a little further below than the median meeting, well inside the middle of the range (−2.57 to +0.91 is the 10th-to-90th band for the Taylor rule). The rate has been below the Taylor rule at 17 consecutive meetings, since Sep 2024.

Today's rule levels use the data in hand on 2026-10-23 (unemployment for September, core PCE for August) and the current target midpoint of 3.875 percent; the historical gaps use the target set at each scheduled meeting against the rule on the data in hand that day. Percentiles count the share of past meetings where the rate sat further below its rule than today (a low percentile = unusually far below). The gap to the inertial rule is small by construction and is reported for completeness.

The spread across the rows is the real information. Today's gap runs from near zero on the two rules that carry the current rate inside them (the inertial rule, −0.12; the first-difference rule, −0.32) to −2.82 on the balanced approach with CBO's output gap. Three choices account for most of it. The slack measure: CBO's output gap is +1.5 percent of potential, while the unemployment gap is only +0.4 (2 × (u* 4.4 − u 4.2)), a difference worth 1.2 points of prescribed rate under the balanced approach. The neutral rate: Laubach–Williams says 1.65, Holston–Laubach–Williams 1.01, the committee's own longer-run dot implies 1.2; that is a 0.6-point swing between the LW and HLW rows. And the rule's weight on slack, which matters less today because the unemployment gap is small. With the committee's own r* and u*, the gap is −0.84, the 66th percentile of 2012–26, a period in which the rate sat far below the rules for years at the floor; against that benchmark today's shortfall is smaller than usual.

As known at the time the history is a little different and the verdict on today is the same. Through 2005–08 the Laubach–Williams r* the committee could have read was 2.2 to 2.5 percent against the 1.5 to 2 the model now assigns to those years, so in real time the rate looked further below the rules in the mid-2000s than it does in hindsight; through 2019–21 the NSA core CPI ran a point or more above core PCE, so the real-time CPI rule puts the Fed further below in 2019 and far further below in 2021–22. On the fully real-time versions (the committee's own longer-run r* and u*, real-time unemployment and core PCE, 2012 on), the rate has been below the balanced approach at 66 percent of meetings, with a median gap of −1.18; today's −0.84 is the 66th percentile excluding the floor.

The same, on inputs as they were known at the time: the unemployment rate and price index in the vintage current on the meeting day (ALFRED), the Laubach–Williams r* from the vintage the New York Fed had published by then (its real-time workbook begins with 2005Q1; earlier meetings use today's estimate), CBO's natural rate from the vintage current then (2011 on; today's before), and for the SEP versions the committee's own longer-run funds rate less 2 and longer-run unemployment, which are real-time by construction. "Taylor (1993): r* = 2" is Taylor's own constant, so with the NSA core CPI that row is fully real-time back to 1960.

Fixed or rolling neutral rate

Taylor wrote his rule with a constant neutral real rate of 2 percent, and whether to keep it fixed or let it move is the single choice that most changes what the rule says about the last fifteen years. Here the Taylor (1993) rule is run both ways on the same inflation, unemployment and u* inputs: with r* fixed at 2 (and at 1, as a sensitivity), and with a rolling r* from three sources, the Laubach–Williams one-sided estimate the page uses, Holston–Laubach–Williams, and the committee's own longer-run funds rate less 2 from the SEP, which exists from January 2012. The model estimates come in today's vintage and as the New York Fed had published them at the time (Laubach–Williams from 2005, Holston–Laubach–Williams from the fourth quarter of 2015); the SEP figure is real-time by construction.

Today the choice is worth a point. With r* fixed at Taylor's 2 the rule gives 5.70 and the rate sits 1.83 points below it; with Holston–Laubach–Williams (1.01) the rule gives 4.71 and the gap is −0.84; with the committee's own longer-run rate (1.2) the gap is −1.03; Laubach–Williams (1.65) sits between, at −1.48. The percentiles move the other way from the gaps. Against the post-1990 history the fixed-r* gap is the 27th percentile (further below the rule than usual), while the rolling versions put today at the 56th (HLW) and 62nd (SEP) percentiles, less far below than usual. The history explains why. With a fixed 2 percent the rate has been below the Taylor rule at 100 percent of non-floor meetings since 2012, with a median shortfall of 2.37 points: on that reading the Fed has been easy for fifteen years without interruption, which is the Taylor critique of the 2010s in one number. Let r* roll and the shortfall since 2012 shrinks (1.74 with LW, 1.33 with HLW, 1.20 with the SEP) and on HLW and the SEP the rate is below the rule at 86 to 87 percent of meetings rather than all of them (on LW, whose post-COVID re-estimates of r* run high, still 99 percent). Before 2008 the choice mattered less, and in the late 1990s it ran the other way: the models put r* at 2.2 to 3.7 then (Laubach–Williams, today's estimates), so against them the Greenspan Fed looked tight while against a fixed 2 it looked about right. As known at the time the rolling versions were lower still (Laubach–Williams was negative in late 2015 and 0.4 at the 2022 liftoff), so in real time a committee following a rolling-r* Taylor rule in 2016–19 was closer to it than today's re-estimates suggest, and a rolling-r* reading of 2021–22 is a little less damning than the fixed one (−7.1 at worst on real-time HLW against −9.1 with a fixed 2).

Every row uses CBO's natural rate and the same inflation and unemployment inputs, so the rows within a block differ only by the neutral real rate. "Since 2012" is the common sample on which the SEP version exists; floor meetings are excluded from the percentiles and medians. A low percentile means the rate sits unusually far below the rule.

The same split can be put into the delta rule of section 2, where a rolling neutral rate adds Δr* (and, for the SEP version, 2 × Δu*) to the change the rule asks for over the window. It is a small term in most windows and today it widens the 12-month gap by about a fifth of a point on the Laubach–Williams and SEP versions, since both have the neutral rate rising over the past year, and not at all on Holston–Laubach–Williams.

Balanced-approach weights (Taylor where stated), core PCE on today's vintage unless stated. The r* and u* term is Δr* + 2 × Δu* over the window: the committee's longer-run funds rate rose from 3.0 to 3.2 between September 2025 and September 2026 while its longer-run unemployment rate was unchanged, and Laubach–Williams r* rose about a fifth of a point; Holston–Laubach–Williams barely moved. Percentiles are against each version's own non-floor pairs since 1990; the SEP rows exist only from 2012, so theirs are against the 2012–26 non-floor pairs, a smaller and more dovish-leaning sample.

By era

Today's-vintage inputs unless stated (LW r*, CBO u*, core PCE). The funds rate has sat below the Taylor rule at most meetings since the 1960s; the exceptions are the Volcker years and the late 1990s.

Some meetings people argue about

The funds rate minus the rule's level at some meetings people argue about, in percentage points.

Two things in this table bear on how to read today. First, the 2021–22 numbers are matched only by the worst meetings of the 1974–75 recession: eight to twelve points below the rules in late 2021 and early 2022, against a Volcker peak of four to five above, and a 1970s low of −10.4 on the Taylor rule (−9.8 on the balanced approach) in Jan 1975, when the committee cut into a recession with inflation above 10 percent. Second, the current episode is not 2003–04 in the rules' eyes. In June 2004, at the start of that tightening, the rate sat 2.88 points below the Taylor rule on today's inputs and 1.76 as known then; today it sits 1.48 and 1.16 below. On the balanced approach the comparison is 2.32 then against 1.67 now. The distance is similar; what differs is that in 2004 inflation was 1.7 percent and the gap came from a low rate, while today core inflation is 3 percent and the gap comes from a rule level above 5.

4. Projections against outcomes

Each SEP's projection for next year (or the year after) against the year's actual Q4/Q4 core PCE, Q4 unemployment and year-end target midpoint. “Fed” is the actual year-end rate minus the dot; “rule” is the balanced approach run on the surprises, longer-run values held at the anchor's. The floor flag marks pairs where both the dot and the outcome were at the floor; the pandemic flag marks 2020 and 2021 outcomes.

Against outcomes the picture is a strong fit with a steady bias. Excluding the floor and the pandemic the rule run on the surprises explains 92 percent of how far the Fed ended up from its own dot a year ahead, with a slope of 0.93: when inflation or unemployment surprise, the committee's year-end rate moves relative to its earlier dot almost exactly as the balanced approach says. But the average gap is −0.76: the Fed ends the year about three-quarters of a point below where the rule's response to the news would put it, and −1.25 two years out. In 2014–19 the dots were too high relative to the Fed's own subsequent behavior (unemployment undershot the projections, the rule reads that as a reason to tighten, and the committee instead raised rates more slowly than its dots had said); in 2021–22 the inflation surprise was three points and the committee delivered most but not all of the implied five; in 2023–25 small upward surprises to both met rate paths that came in below the dots. Only two of 35 clean pairs are positive. Part of this is the well-known upward bias of the dots in the 2010s; part is the same half-delivery on the tightening side as in section 1. With Taylor weights the bias is −0.55.

5. How unusual is the gap now, and what it would take to be typical

Percentiles are the reading's place among all earlier pairs at the same horizon, excluding the pandemic pairs and the pairs where both dots sat at the zero floor (a low percentile = among the most dovish readings on record). Changes in percentage points.

All in percentage points, balanced-approach weights (Taylor where stated), core PCE.

The three versions agree on the sign and disagree on the size, and the disagreement is informative.

• On the committee's own forecasts the gap is large by the standards of 2012–26. −1.3 over twelve months is the 6th percentile of the non-pandemic pairs and 6th most dovish of all 52; only June 2025 → June 2026 (−1.65), June 2023 → June 2024 (−1.50) and the pandemic pairs are below it, December 2023 → December 2024 ties it, and the next company is the rest of 2023–24 (−0.9 to −1.2) and the 2013–14 liftoff delay (−0.65 to −0.95). Taylor weights make it −1.00 (14th percentile); holding the longer-run values fixed −1.10; headline PCE −1.75. Over 24 months, −1.60 (11th percentile).
• It is not catch-up. In the 2023 → 2024 and 2024 → 2025 pairs the year-end dots started 0.3 to 2.2 points above the balanced-approach level implied by the committee's own forecasts, and section 1's estimate says a third of that comes off each year regardless of revisions; those negative readings were part convergence. In September 2025 the end-2026 dot was 0.1 points from its rule level. The whole of the current gap is a response to revisions that did not happen.
• It is about 2026, not 2027. Scored on end-2027 the same pair reads −0.20: the committee moved its 2027 path (up 1.0) in line with its 2027 revisions (+1.2). The committee is signaling that it will get there, later.
• On realized data the lean is real but not extreme, and it depends on the vintage. As the committee saw the data, the twelve months to the September meeting read −1.35, the 9th percentile of 2001–26, and the 24 months −2.29 (3rd percentile), because core PCE as first published rose from 2.6 to 3.3 percent over those two years. On today's vintage, which has revised that rise down to 2.8 to 3.0, the same windows read −0.86 (20th percentile since 1990) and −1.46. To today the gap is −0.42 on either vintage; the Greenspan-era standard deviation of this gap is 0.88 and today's reading is inside it. Core CPI reverses the sign.
• In levels the rate is below every rule and that is normal. −1.48 against the Taylor (1993) rule and −1.67 against the balanced approach on today's inputs are the 38th and 29th percentiles of post-1990 meetings; the rate has been below those rules at 78 and 73 percent of them. The range across twenty rule-and-input choices is −2.82 to −0.84, with the committee's own longer-run values giving −0.84. The large gaps come from CBO's output gap (+1.5 percent of potential) and the Laubach–Williams r*; the small ones from the committee's own r* and from any rule that carries the current rate inside it. The neutral rate alone is worth a point: hold it at Taylor's 2 percent and the Taylor gap is −1.83 (27th percentile since 1990, and the rate has then been below the rule at every non-floor meeting since 2012); let it roll with Holston–Laubach–Williams or the SEP and it is −0.84 to −1.03 (56th to 62nd percentile).
• Why the versions differ. Between the September 2025 and September 2026 meetings trailing core PCE rose +0.1; the committee's projection for 2026 Q4/Q4 core PCE rose +0.8, to 3.4, which is an acceleration from the current 3.0 into year-end. The committee has raised its forecast and not its path: either it expects the acceleration to be transitory and is looking through it (tariffs are the obvious candidate), or the forecast is a committee median that the policy median does not share. Either way, on its own numbers it is −1.3 points easier than the rule it has described as a benchmark, and a 24-month view says the same (−1.6).

What it would take to reach the 50th percentile

Every measure puts the Fed on the easy side of its rules and they disagree about how much. The table takes each one, asks what its historical median is, and translates the distance into a change in the funds rate, with the October 28 and December 9 meetings in mind.

A positive change is a hike. For the level rules and the realized-data windows ending today, a quarter-point move changes the reading by a quarter point, so the translation is direct (the inertial and first-difference rules, which carry the rate inside them, are left out). For the dots version the object is the end-2026 median dot, which the December projections will set equal to the year-end rate: a typical response to the past year's forecast revisions would put the end-2026 rate near 5.1 percent against the 4.1 the September projections show and the 3.875 in force now. The LW-based level medians are about half a point lower on a post-2000 sample than on a post-1990 one, because the 2000s and 2010s carried larger shortfalls; on that sample the Taylor rule with LW r* says no move.

In October-and-December terms the answers fall into three groups. The level rules with a rolling neutral rate say nothing to two quarter points: with Holston–Laubach–Williams or the committee's own longer-run rate the Taylor rule already sits a shade above its median (−0.28 and −0.17 to reach it), and with Laubach–Williams it asks for +2 quarter points on today's vintage and +2 quarter points as then published. The realized-data delta rule as known at the time says about the same, +2 quarter points, since the window to today already contains the September hike. The committee's own forecast revisions say far more: a typical response to the past year's revisions would put the end-2026 rate near 5.1 percent, about +5 quarter points from here over the two meetings, where the September projections show one. The measures that use a fixed 2 percent neutral rate or CBO's output gap ask for a point or more, but they are also the measures that have called the Fed easy at every meeting since 2012, which is a statement about the neutral rate rather than about the committee.

6. Which perspectives have worked best, and what they imply now

Ranked by how well each has tracked what the Fed actually did:

• Best: the committee's dots as a function of its own forecast revisions. Excluding the pandemic, the revisions explain 68 percent of the variation in 12-month dot changes with balanced weights and 74 percent with Taylor weights, with a typical miss of 0.62 points against 1.09 for assuming no change. On this view the Fed is about a point short of its usual response: it raised its 2026 inflation forecast by 0.8 and its 2026 rate path by 0.7, where its own habits (a weight of about 1.5 on an inflation revision) would have produced about 1.7.
• Second: level rules with a rolling neutral rate, on the data as it was known at the time. Since 2000 the real-time Taylor rule with Holston–Laubach–Williams has stayed within a point of the actual rate at 44 percent of meetings and the Laubach–Williams version at 47 percent, with standard deviations of 1.31 and 1.37; a fixed 2 percent neutral rate manages 27 percent and the output-gap version 24. These say today is normal: between no move and one hike.
• Third: the delta rule on realized data. It explained 42 percent of 12-month changes since 1990 on today's core PCE (36 on the real-time core CPI), and about nothing since 2014. It currently reads one to two quarter points short.
• Least useful as a guide: anything with a fixed neutral rate or CBO's output gap. Those have said "too easy" without interruption since 2012, which makes them a statement about r* and potential output, not about the committee.

Scheduled meetings from 2000 excluding the floor, the common sample on which real-time core PCE exists (the SEP rows from 2012). A rule that tracks the committee stays within a point of the rate more often and has a smaller standard deviation; a fixed neutral rate scores badly on both because the level it implies drifted away from the committee's for fifteen years. The inertial and first-difference rules are left out: they track the rate by construction.

The two best perspectives disagree, and the disagreement is the finding. The forecast-based measure is large because the committee's 2026 core PCE projection rose 0.8 over the year while realized core PCE rose 0.1 on today's vintage and 0.5 as the committee saw it. If the forecast is right, the gap says more tightening is coming than the one hike the dots show. If the forecast is wrong, the gap closes itself at the December projections without any move. The outcome-based version of section 4 adds a warning about which way that usually goes: since 2014 the Fed has ended each year about 0.8 points below where a rule-consistent response to the surprises would have put it, so under-delivery relative to its own dots is the base case, not the exception.

7. Does the Fed have a reaction function, and where to read it

Many argue the Fed should publish its reaction function rather than its rate plans. The evidence here is that it already behaves as if it has one, and that the dots reveal it better than the data do. Three features have been stable since 2012:

• Taylor-type weights on its own forecasts. About 1.6 on a revision to next year's inflation and about 1.1 on a revision to unemployment, not the balanced approach's 2. The same weights hold on the December-to-December pairs (1.8 and 1.6) and at 24 months.
• Asymmetry. When its revisions call for easing it delivers nearly all of it (−0.62 against −0.73); when they call for tightening it delivers about half (+0.73 against +1.41).
• Gradual convergence to the level rule. When its path sits above or below the balanced-approach level implied by its own forecasts, it closes about a third of the gap each year (0.37 per point at 12 months, 0.86 at 24) regardless of new revisions. On realized data the figure is about half a point per year. That is the Board's inertial rule with the smoothing estimated rather than assumed, and the estimate lands close to the Board's 0.85 a quarter.

The realized data do not reveal any of this. Since 1990 the Fed's 12-month moves carry the full balanced-approach weight on changes in unemployment (−2.05 as known then) and no weight at all on changes in trailing core inflation (−0.18). That is not because it ignores inflation. It reacts to its inflation forecast, and until 2021 the trailing twelve-month rate carried almost no news for the forecast. A data-only reader of the 2014–26 Fed would conclude it had no reaction function (section 2's R² of 0.14 under Yellen and 0.03 under Powell), when what it had was a reaction function to levels and forecasts rather than to changes in the data. So to understand the reaction function, read the SEP. Use the data for a different question: whether the forecasts the committee is reacting to are turning out to be right, which is what the vintage comparison in section 2 and the outcome scoring in section 4 are for.

Two caveats on the SEP as the place to read it. The forecast median and the dot median come from different participants, so a revision to one is not a revision by the people who set the other, and each participant's forecast is conditional on his or her own policy path. And the 13 consecutive negative readings since late 2022 mean either that the weights have drifted dovish or that the committee is deliberately looking through what it forecasts, most plausibly tariff-driven inflation in 2025–26; the Fed-minus-rule arithmetic cannot tell those apart. A published reaction function would make that distinction explicit instead of leaving it to be inferred from the dots four times a year. On this evidence the committee already has one; stating it would mostly formalize what the dots reveal, and would force the current deviation to be explained rather than discovered.

8. Which rule, and what to put on the page

• Of the two delta rules, the forecast-revision version is the better reaction function and the realized-data version the better long history. The forecast version explains two-thirds of the committee's path changes once the pandemic is set aside, with fitted weights close to a rule; the realized version explains two-fifths since 1990 and nothing since 2014, because policy since then has been about levels. For the question “when its view of the data changed, did the committee change rates?”, use the dots.
• Taylor weights, not balanced-approach weights, describe the dots. On every sample the fitted weight on the unemployment revision is about 1, not 2. The page's updater defaults to the balanced approach; a Taylor default (or showing both) would put the historical gaps closer to zero and make the current one −1.00 rather than −1.30, still the 14th percentile.
• Rolling with all meetings, not December only. December-to-December gives 11 usable pairs and the same estimates; the rolling series gives 46 and shows the turning points (late 2022, mid-2025) that annual pairs straddle. Report the rolling series and use December pairs as the check; the overlap only matters for standard errors.
• Score the nearest year-end. The following year-end is a view of neutral, not a reaction function.
• Show the starting level gap beside the delta. Both delta rules miss the convergence to the level rule, which explains the 2021–23 readings entirely and the 2023–25 readings in part. One extra number, the dot minus the rule's level at the anchor, separates “moving toward the rule” from “moving away from it”; today it is near zero, which is what makes today's gap a real one.
• Flag the floor and the pandemic in the data, not the text. Any statistic across the sample is dominated by six pairs otherwise.

Methods and data

The projections panel

Fifty-eight sets of projections, January 2012 to September 2026. From September 2015 the medians are the Fed's published Table 1 medians, as cached by the Taylor Rule page (sep_history.json). Before that the Fed published only central tendencies and ranges; the medians used here are computed from each participant's submission in the Fed's SEP compilations (the FOMC<date>SEPcompilation files released with the transcripts: HTML for 2012, PDF from 2013), Table 2 of each, 16 to 19 participants per round. As a check the same code was run on the September 2015 compilation and reproduces the published medians exactly once rounded to one decimal (funds rate 0.38/1.38/2.63/3.38 against the published 0.4/1.4/2.6/3.4; every other variable identical). A participant's funds-rate submission at the zero floor is recorded as 0.13 in the compilations, and 0.1 in the published medians; the difference does not affect any change. The two 2013 rounds in which a single participant submitted a 2016 value have that year dropped. Core PCE has no longer-run projection by construction. The published medians round to one decimal and the compilation medians can end in .05 or .125; changes are taken as they come.

Pairing and scoring

For version 1 each set of projections is paired with the set closest to 12 (24) months later, within two months, ties to the later (January 2012 → December 2012 is an 11-month pair; the March 2020 round was canceled, so March 2019 has no 12-month partner). The year-end scored is the later set's current year (“nearest”), or the year after (“following”); both sets must project it. Fed = the later median dot minus the earlier, for that year-end. Rule = Δ(longer-run funds rate) + φ Δ(core PCE projection) − β·Okun·[Δ(unemployment projection) − Δ(longer-run unemployment)], with φ 1.5, β 1, Okun 2 (balanced) or β 0.5 (Taylor), exactly the page's updater arithmetic; the page's own chart (every earlier set → September 2026 for end-2026) is reproduced to the decimal as a check. Pandemic pairs are those whose later set is dated 2020 or 2021; floor pairs those where both dots are at or below 0.25. For version 3 the later set is replaced by the outcome: Q4/Q4 core PCE from BEA, the Q4 average unemployment rate from BLS, the target midpoint on December 31, from the page's loader; longer-run values are the anchor's.

For version 2 every scheduled meeting in the page's meeting list (662 since January 1960; intermeeting moves show up in the rate at the next scheduled meeting) is paired with the scheduled meeting closest to 12 (24) months later within 45 days. Inputs are the page's: the latest 12-month core PCE inflation and unemployment rate released by each decision day (ALFRED first-release dates, today's vintage of the values), the target midpoint three days after the decision (the single target 1982–2008, the range midpoint since, the previous month's effective rate before September 1982). The “today” pairs use the data in hand on 2026-10-23 and the current target midpoint. The starting level gap uses the balanced approach with the New York Fed's Laubach–Williams r*, CBO's natural rate and an Okun coefficient of 2, so its scale is the page's unsmoothed featured rule less the output-gap hybrid. Floor pairs are those with the target at or below 0.25 at either end; the chair is the chair at the first meeting of the pair.

What is real-time here, and what is not

The month each input refers to is the one the page already chooses by first-release date; only the value changes with the vintage. "As known at the time" is therefore what a member of the committee could have computed on the decision day from published data, with the two exceptions in the table (the natural rate before 2011 and r* before 2005, where no real-time estimate exists and today's is used).

The level rules

Taylor (1993): r* + π + 0.5(π − 2) + 0.5 × gap. Balanced approach: r* + π + 0.5(π − 2) + gap. Balanced approach with shortfalls: the same with min(gap, 0). Inertial: ρ × (rate going in) + (1 − ρ) × balanced, ρ = 0.85 per quarter converted to the days between meetings, ρ^days/91.31, days clamped to 20–183 (the page's convention). First-difference: rate going in + [0.5(π − 2) + (u* − u) − (u*_−12m − u_−12m)] × days/91.31. The gap is 2 × (u* − u), or CBO's output gap for the latest quarter with a GDP estimate moved by 2 × the change in unemployment since that quarter (the page's hybrid). π is 12-month core PCE unless stated. The rate compared is the target set at the meeting (the midpoint of the range since December 2008; the single target 1982–2008; the next month's effective rate before September 1982), and today's is the current target midpoint. Percentiles of today's gap count the share of past meetings with a lower (more negative) gap, so a low percentile means the rate is unusually far below the rule; the "whole history" column excludes meetings at the floor, where the rate could not follow a rule down.

Statistics

“Slope” is from a regression of the Fed's change on the rule's change with an intercept; the fitted weights from a regression of the Fed's change on the two inputs' changes directly; “catch-up” from a regression of Fed minus rule on the starting level gap (through the origin in the figure, with an intercept in the tables). Standard errors are Newey–West with a lag of three (seven) pairs for the 12-month (24-month) projections series and eight (sixteen) meetings for the realized series, roughly the overlap. RMSE of “no change” is the root mean square of the Fed's change itself. Percentiles count the share of earlier pairs with a lower reading, in the stated sample. Nothing here is a forecast; every comparison is between things that had both happened.

Caveats

• The committee's medians come from different participants in different years and each participant's forecast is conditional on his or her own policy path, so a revision to the median forecast is not a revision by a fixed forecaster, and the dot is not a commitment.
• The current-year dot in a September or December round is constrained by the meetings left, which pulls the “nearest year-end” comparison toward what has already happened; the December-to-December table shows the extreme case.
• The realized-data rule is run on today's vintage of the data with real-time release timing. Revisions to core PCE and the unemployment rate are small; the Fed's actual information set differed more through its forecasts than through its data.
• A twelve-month window is shorter than the committee's smoothing horizon, so slopes below one partly measure gradualism rather than disagreement with the rule's destination; the 24-month results and the level-gap regressions are the correction for that.
• Balanced-approach weights are a choice. The Fed's Monetary Policy Report presents them as one rule among several; the Taylor weights fit the committee's behavior better, and the gaps under them are reported throughout.

Appendix: the series

Download: projections pairs (CSV) Download: meeting pairs since 1960 (CSV) Download: projections vs outcomes (CSV) Download: level gaps, every rule (CSV)

Built 2026-10-24 from the Econvitals Taylor Rule page's data (econvitals.org/tools/taylor-rule: data.json, sep_history.json, meetings.json) and the Federal Reserve's SEP compilations (federalreserve.gov/monetarypolicy/fomchistorical2012.htm through 2015). Rule definitions follow the Board's “Policy Rules and How Policymakers Use Them” and the Monetary Policy Report.