The GSCPI improves on the AR(1)
The GSCPI, alone or with the prices-paid index and read at its value at the forecast origin, improves on the AR(1) in forecasts of Core PCE inflation one through four quarters ahead. The evaluation is pseudo-out-of-sample and reads the current vintage of the data, the GSCPI’s among them, which its publisher revises (Section 3).
Not established by this sample
Neither the output gap, added to the AR(1), nor the federal funds rate, in a test of Granger causality, adds forecasting content distinguishable from zero, and neither the prices-paid index, the producer price index, nor unit labor costs adds any next to the GSCPI, on either vintage where a first release is read. No asymmetric GSCPI specification of the four the paper tests, among them one whose coefficient differs between quarters in which the index rises and quarters in which it falls, improves on the forecasts of the symmetric two-indicator model by a gain distinguishable from zero. A two-regime Markov-switching model is not identified by the sample under the floors of 0.025 and 0.05 percentage points on its residual standard deviations, the second of which is the floor under which the paper evaluates the model out of sample, and under no floor examined does the evidence establish that the sample identifies the model (Section 12).
The line of sight
Two rails run down each section’s left edge, the left for the first theme, “The GSCPI improves on the AR(1),” and the right for the second, “Not established by this sample.” A filled mark ( ) shows that the section belongs to the theme, and a hollow mark ( ) that the theme passes through it.
1 Introduction
- Establishes
- Adding the output gap to an autoregressive model raises R², the share of variance explained, by 0.001 and worsens the Akaike information criterion (AIC). Under the feasible treatment, the GSCPI and the prices-paid index held at their current-vintage values at the forecast origin, the improvement of Model R on the AR(1), the percentage by which its root mean squared error (RMSE) is below the AR(1)’s, rises from 13.3 percent at one quarter to 33.5 percent at four. The gains are distinguishable from zero at the 5 percent level at every horizon when the test’s simulated errors follow the larger model’s residuals and at none when they follow the AR(1)’s; Model R’s gain over the GSCPI alone is distinguishable from zero at no horizon.
- Leads to
- Is the Phillips curve flat on this sample, and why?
2 The Flat Phillips Curve: Motivation
- Establishes
- On quarterly data from 2012:Q1 through 2025:Q4, 54 observations after lags, the output gap’s coefficient in a backward-looking Phillips curve is 0.0098, not distinguishable from zero at the 5 percent level, while the coefficients on the two lags of inflation sum to 0.928: the fit is carried by inflation persistence. The literature’s explanations are anchored expectations, globalization, the identification effect of a stabilizing central bank, and two unusual episodes of the sample, COVID-19 (coronavirus disease 2019) and the slow recovery after the Global Financial Crisis.
- Exhibits
- Equation (1): backward-looking Phillips curve
- Leads to
- Which series do their publishers revise, and on which data are the models evaluated?
3 Data Revisions and Data Versions
- Establishes
- The GSCPI is revised: its publisher revises its latest months as realized data replace imputed values, and the index takes in revisions of some of its inputs up to twelve months back; its releases accessed on 9 October 2026 and 7 November 2026 differ in every quarter of the paper’s sample, and so also in quarters further back than the twelve months the publisher’s account of its revisions covers. Core PCE, unit labor costs, the producer price index, import prices, the trade-weighted dollar, and the Food and Agriculture Organization (FAO) food price index are revised as well; the revision status of the output gap’s series is not established. No revision is found for the federal funds rate or for the prices-paid index, read not seasonally adjusted, and none is established for crude oil. The models are evaluated on two data versions: the revised data (every series as published by 9 October 2026, its current vintage) and the first releases of Core PCE, unit labor costs, and the producer price index. The out-of-sample forecasts of the linear models are pseudo-out-of-sample: no value dated after the forecast origin enters a model’s coefficients or its forecast, the conditional diagnostic of Section 8 excepted, but the values are those published by 9 October 2026 rather than those available at the origin, the GSCPI and the prices-paid index were selected as regressors on the full sample, and the paper does not evaluate the models on the data as they stood at each origin. The regime-switching model’s out-of-sample RMSE, reported as a range (Section 12), rests on optima found on the full sample.
- Exhibits
- Table 1: revision status of the series
- Leads to
- Which supply-side indicators lead consumer prices, and by how much?
4 Supply-Side Leading Indicators
- Establishes
- Eight indicators measure upstream cost pressure that passes through the pricing chain; among them only the prices-paid index, read not seasonally adjusted, shows no revision, and crude oil’s revision status is not established. The GSCPI reaches its highest value in the sample, 4.3 standard deviations above the index’s average value, in 2021:Q4, one quarter before Core PCE inflation peaks at 5.5 percent in 2022:Q1.
- Leads to
- How are the candidate models specified, estimated, and selected?
5 Forecasting Models and Estimation
- Establishes
- The candidate models of Table 2, among them the AR(1) and Model E, which carries all eight indicators, forecast year-over-year Core PCE inflation one quarter ahead from regressors dated at the origin. An AIC forward stepwise procedure on revised data adds the GSCPI, then the prices-paid index, and stops: the two form Model R, the reference. Model F adds unit labor costs to the GSCPI.
- Values
- −98.7AIC, AR(1)−125.8AIC, Model R
- Exhibits
- Table 2: forecasting model specifications; Equations (2) through (4): the general framework and Models F and R
- Leads to
- How do the models fit in sample, on revised data and on first releases?
6 In-Sample Results: Revised and First-Vintage Data
- Establishes
- The output gap leaves the fit of the AR(1) all but unchanged and worsens the AIC, while the supply-side indicators improve both the fit and the criterion, Model R doing so with the GSCPI and the prices-paid index. The GSCPI’s coefficient is distinguishable from zero at the 5 percent level in every data version; where the simulated errors follow the residuals of the model without the GSCPI, the coefficient in Model R stands at the boundary of that level, and whether it is distinguishable from zero there is unsettled. The prices-paid index’s coefficient in Model R is not distinguishable from zero, and neither is that of unit labor costs, in revised data or as first released. Model E has the highest R² in sample but takes a worse AIC than Model R (−118.4 against −125.8).
- Values
- 0.924R², AR(1), revised data0.925R², AR(1) with the output gap (Model B), revised data0.959R², Model R, revised data0.963R², Model E, all eight indicators, revised data
- Exhibits
- Table 3: in-sample fit; Table 4: coefficient comparison across data versions
- Leads to
- Does the in-sample fit carry over to out-of-sample accuracy?
7 Out-of-Sample Forecast Evaluation
- Establishes
- Over the 24 one-quarter targets of 2020:Q1 through 2025:Q4, Model R’s RMSE is below the AR(1)’s; the margin is distinguishable from zero at the 5 percent level under the larger model’s residuals and not under the AR(1)’s. The output gap does not lower the AR(1)’s error, and its gain is not distinguishable from zero. Models D and F, which add the producer price index and unit labor costs, respectively, to the GSCPI, record RMSEs of 0.463 and 0.479 percentage points, below Model R’s, and Model D records 0.462 with the producer price index as first released, a difference from its current vintage the paper does not test. The GSCPI alone (Model G) records an RMSE of 0.459 percentage points, below all of them, an improvement of 17.5 percent on the AR(1), distinguishable from zero under either set of residuals; the gain of Model R over Model G is not distinguishable from zero, and next to the GSCPI the producer price index adds nothing distinguishable from zero on either vintage.
- Values
- 0.556RMSE, AR(1), percentage points0.506RMSE, Model R, percentage points9.1Improvement of Model R on the AR(1), one quarter ahead, 24 targets 2020:Q1–2025:Q4, percent0.557RMSE, AR(1) with the output gap (Model B), percentage points
- Exhibits
- Table 5: out-of-sample results, one quarter ahead
- Leads to
- Does the advantage hold, grow, or fade as the horizon lengthens?
8 Multi-Horizon Iterated Forecasts
- Establishes
- Under the feasible treatment, the indicators held at their current-vintage values at the forecast origin, on a common window of 21 targets, Model R’s improvement on the AR(1) rises from one quarter to four. The one-quarter figure differs from Section 7’s only through the targets covered. The improvement is distinguishable from zero at the 5 percent level at every horizon under Model R’s residuals and at none under the AR(1)’s. Under the conditional treatment, a diagnostic that enters the indicators’ realized values after the origin, it is larger still from two quarters on. The GSCPI alone improves on the AR(1) by between 23.7 and 27.6 percent across horizons, distinguishable from zero at every horizon under the larger model’s residuals and at one and two quarters under the AR(1)’s. In the point estimates the widening of Model R’s margin is the prices-paid index’s, and at no horizon is the gain of Model R over Model G distinguishable from zero. Both biases grow in absolute value with the horizon, in opposite directions: the AR(1) over-predicts and Model R under-predicts.
- Values
- 13.3Improvement of Model R on the AR(1), one quarter ahead, common window of 21 targets, feasible treatment, percent33.5Improvement of Model R on the AR(1), four quarters ahead, common window of 21 targets, feasible treatment, percent43.0Improvement of Model R on the AR(1), four quarters ahead, common window of 21 targets, conditional treatment (diagnostic), percent27.0Improvement of Model G on the AR(1), four quarters ahead, common window of 21 targets, feasible treatment, percent
- Exhibits
- Figure 1: multi-horizon RMSE; Table 6: forecast RMSE by horizon
- Leads to
- Is the four-quarter improvement a property of the indicators or of the iteration?
9 Four-Quarter-Ahead Direct Forecasts
- Establishes
- Estimated directly four quarters ahead, Model R’s in-sample RMSE is lower than the direct AR(1)’s by a percentage close to its out-of-sample improvement at four quarters under the feasible treatment. Their agreement, across a difference in method, evaluation, and sample, is consistent with an improvement that comes from the indicators rather than from the iteration. At this horizon the output gap’s coefficient is not distinguishable from zero at the 5 percent level.
- Values
- 33.3In-sample improvement of direct Model R on the direct AR(1), percent33.5Out-of-sample improvement of Model R on the AR(1), four quarters ahead, feasible treatment, percent
- Exhibits
- Table 7: four-quarter-ahead direct forecasts
- Leads to
- Where in the pandemic era does the advantage come from, and would asymmetric GSCPI specifications have helped?
10 The Pandemic and Its Aftermath: A Sub-Period Decomposition
- Establishes
- In the point estimates Model R is more accurate than the AR(1) in 2022–2023, 2024, and 2025, and less accurate in 2020 and in the surge of 2021, when both models under-predicted, and the AR(1) by less. The episodes carry four targets each (eight in 2022–2023), and the paper does not test the differences. The four asymmetric GSCPI specifications move the one-quarter-ahead out-of-sample RMSE by between −0.0269 and +0.0121 percentage points relative to Model R (the specification’s RMSE minus Model R’s). The largest gain, that of the split between rising and falling quarters, is not distinguishable from zero at the 5 percent level, tested alone or across the four forms, and Model R, with its symmetric loading, remains the reference.
- Values
- −40.2Improvement of Model R on the AR(1), one quarter ahead, 2020, percent−11.3Improvement of Model R on the AR(1), one quarter ahead, 2021, percent52.7Improvement of Model R on the AR(1), one quarter ahead, 2022–2023, percent30.4Improvement of Model R on the AR(1), one quarter ahead, 2025, percent
- Exhibits
- Figure 2: sub-period RMSE; Table 8: sub-period RMSE and bias; Table 9: asymmetric GSCPI specifications
- Leads to
- What happens when the models are frozen before a year of tariff increases?
11 2025 Out-of-Sample Validation
- Establishes
- Model R, estimated through 2024:Q4, records the lowest error of the models compared over the four quarters of 2025, the year of the tariff increases; the GSCPI alone (Model G) records an RMSE above the AR(1)’s. The data are at their current vintage, and the regressors of Models G and R were selected on the full sample, which contains 2025. The AR(1) places the year above the outcome; Models D and F, which add the producer price index and unit labor costs, respectively, to the GSCPI, place it below. Model L, Model F with cumulative tightening of the federal funds rate (the policy stance), moves Model F’s forecasts part of the way toward the outcome, lowering the RMSE below Model F’s 0.144 percentage points, its RMSE remaining above the AR(1)’s and Model R’s. The comparisons rest on four targets, on which the paper runs no test.
- Values
- 0.068RMSE, Model R, 2025, percentage points0.100RMSE, AR(1), 2025, percentage points0.132RMSE, the GSCPI alone (Model G), 2025, percentage points0.111RMSE, Model L, 2025, percentage points
- Exhibits
- Table 10: 2025 validation
- Leads to
- Can a two-regime model separate a logistics channel from a trade and cost-push channel?
12 Regime-Switching Models
- Establishes
- A two-regime Markov-switching model is specified to separate a logistics channel, read through the GSCPI, from a trade and cost-push channel, read through import prices. Its estimation imposes a floor, whose value has no published source, on the regimes’ residual standard deviations. Under the four floors examined from 0.025 through 0.1 percentage points, among them 0.05, the floor under which the paper evaluates the model out of sample, the highest optimum a search over starting values finds lies on the floor. Under floors of 0.115, 0.125, and 0.13 the highest optimum found is interior, its standard deviations at least 0.01 above the floor and its stay probabilities, the probabilities that each regime persists from one quarter to the next, within [0.01, 0.99]; the search reaches that optimum from 3.8 percent, 3.8 percent, and 2.7 percent, floor by floor, of the 600 starting values it draws at random. Under the floors of 0.025 and 0.05 the highest optimum found exceeds every interior optimum by more than the margin of 1.92 in log-likelihood by which the search compares optima, and the sample does not identify the model. Under the other floors an interior optimum and an optimum on the floor stand within that margin of each other, so that the evidence does not establish that the sample identifies the model under these floors. Under the floor of 0.05 the best interior optimum and the optimum the search reaches from a default starting vector order the regimes’ residual standard deviations differently. Across the initial optima and restart rules of the model’s re-estimation, its one-quarter-ahead out-of-sample RMSE runs from 0.500 through 1.111 percentage points against the AR(1)’s 0.556, a range that rests on optima found on the full sample and so is not a pseudo-out-of-sample result as the other models’ are.
- Exhibits
- Equations (5) and (6): the two regimes; Table 11: multistart search; Table 12: static versus regime-switching comparison; Table C.1: the search under each floor
- Leads to
- Does the federal funds rate add forecasting content that the supply-side models omit?
13 Does the Policy Rate Add Forecasting Content?
- Establishes
- In a regression of Core PCE inflation on its own two lags, the lags of the federal funds rate add nothing distinguishable from zero at the 5 percent level, and neither do those of the prices-paid index once the test’s distribution is simulated; those of the GSCPI do.
- Exhibits
- Table 13: Granger causality tests
- Leads to
- What does a better forecast of supply-driven inflation give a policymaker?
14 Implications for Monetary Policy
- Establishes
- Forecasting inflation and controlling it are different capabilities: in the reduced form the output gap’s coefficient on inflation is not distinguishable from zero, so the reduced form does not establish that a better forecast of supply-driven inflation gives a policymaker an instrument for preventing it through demand, nor what policy could achieve, since a flat curve can also reflect successful stabilization (Section 2). To forecast Core PCE inflation one through four quarters ahead, a forecaster can add the GSCPI, held at its value at the forecast origin, to an autoregressive model: on the common window and the current vintage of the data, the indicators selected on the full sample, it lowered the RMSE relative to the AR(1) by 23.7 percent at one quarter and 27.0 percent at four, gains distinguishable from zero at every horizon under the larger model’s residuals and at one and two quarters under the AR(1)’s. The model with both indicators lowered the RMSE relative to the AR(1) by 13.3 percent at one quarter and 33.5 percent at four, gains distinguishable from zero at every horizon under the larger model’s residuals and at none under the AR(1)’s; its gain over the GSCPI alone is distinguishable from zero at no horizon. One quarter ahead the model with both indicators had larger errors than the AR(1) in 2020 and 2021, differences the paper does not test. To watch for supply-chain pressure ahead of Core PCE inflation, a forecaster or a policymaker can follow the GSCPI, whose highest value in the sample, in 2021:Q4, preceded Core PCE inflation’s, in 2022:Q1, by one quarter, measured on its current vintage (Section 4), and whose lags a bootstrap test of Granger causality for Core PCE inflation distinguishes from zero over the sample (Section 13); the paper does not test whether the GSCPI separates supply shocks from demand shocks. To decide what weight Phillips-curve models carry in a forecasting suite, a forecaster can compare them out of sample with the supply-side models over the period at hand; on this sample the output gap did not lower the AR(1)’s one-quarter error (Model B’s RMSE 0.557 against the AR(1)’s 0.556 percentage points).
- Leads to
- What does the paper contribute to forecasting inflation when the Phillips curve is flat?
15 Conclusions
- Establishes
- The output gap adds no forecasting content distinguishable from zero in the inflation-targeting era. For targets within 2020:Q1 through 2025:Q4 the GSCPI alone and with the prices-paid index improve on the AR(1) at every horizon by gains distinguishable from zero under the larger model’s residuals; under the AR(1)’s, the GSCPI’s gains are distinguishable at one and two quarters and the two-indicator model’s at no horizon. The linear models’ out-of-sample forecasts are pseudo-out-of-sample, and the paper does not evaluate the models on the data as they stood at each forecast origin (Section 3). Next to the GSCPI, neither the producer price index nor the prices-paid index adds anything distinguishable from zero. Unit labor costs carry no forecasting content distinguishable from zero on either vintage, and the federal funds rate carries none. In the point estimates the advantage is largest in 2022–2023, present in 2024 and 2025, and reversed in 2020 and 2021, differences the paper does not test. No asymmetric GSCPI specification improves on the symmetric two-indicator model by a gain distinguishable from zero. The highest optimum a search over starting values finds for a two-regime Markov-switching model lies on the floor under the four lowest floors examined and is interior under the three highest. Under the floors of 0.025 and 0.05 percentage points the highest optimum found exceeds every interior optimum by more than the margin of 1.92 in log-likelihood by which the search compares optima, and the sample does not identify the model. Under the other floors examined an interior optimum and an optimum on the floor stand within that margin of each other, so that the evidence does not establish identification. Under the floor of 0.05 percentage points, the model’s one-quarter-ahead out-of-sample RMSE runs from 0.500 through 1.111 percentage points, against 0.556 for the AR(1), depending on the full-sample optimum its re-estimation starts from and on the restart rule. Every series is read from its publisher’s file or from the copy FRED distributes, with the file and its access date named.
- Leads to
- Appendix A: Data Sources and Provenance; Appendix B: Tests and Their Simulated Distributions; Appendix C: The Regime-Switching Search Under Each Floor