The paper's principal results
Model R, the GSCPI and the prices-paid index added to the AR(1), records a lower out-of-sample error than the AR(1) at every horizon, by amounts a test distinguishes from zero at every horizon or at none, depending on the residuals its simulated errors follow (the Out of sample tab), while the output gap's coefficient is not resolved (not distinguishable from zero at the 5 percent level).
Core PCE inflation and the two indicators of Model R
Quarterly. Core PCE inflation is year over year in percent; the GSCPI is in standard deviations from the index's average value; the Philadelphia Fed prices-paid index, not seasonally adjusted, is a diffusion index, the share of firms reporting higher input prices less the share reporting lower (0 is neutral). Each series stands on its own panel and scale; the values are those published by 9 October 2026.
The model and its reading
One-quarter-ahead accuracy by model
Multi-horizon forecasts on a common target window
Under the feasible treatment the indicators are held at their origin-date values, so that no value dated after the origin enters, the values held being those of the current vintage; under the conditional treatment their realized path enters, a diagnostic that uses values dated after the origin.
RMSE and bias in percentage points, the bias as the note above defines it; the biases are the AR(1)'s and, under the feasible treatment, Model R's; G: the GSCPI-only Model G under the feasible treatment.
Five episodes
In the point estimates Model R is the more accurate of the two in 2022–2025 and the AR(1) in 2020 and 2021. The one-quarter errors of the recursive evaluation, in which each model is re-estimated at each origin on an expanding window, are decomposed here by target quarter; the 2025 RMSEs of this section come from the recursive evaluation, and those of the section below from models estimated once, through 2024:Q4.
RMSE in percentage points; the improvement is the percentage by which Model R's RMSE is below the AR(1)'s.
2025 out of sample, estimated through 2024:Q4
Of the paper's models estimated once on data through 2024:Q4, Model R records the lowest error over the four quarters of 2025, with the data at their current vintage and the regressors of Models G and R selected on the full sample, which contains 2025.
Percentage points, over four targets, on which the paper runs no test; the coefficients are not updated during 2025.
Coefficients across data versions
By the paper's bootstrap p-values the GSCPI's coefficient is resolved in every data version, and the prices-paid and unit-labor-cost coefficients in none. The table gives the one-quarter-ahead equations, with t-statistics in parentheses.
Revisions and first releases
Granger causality
Asymmetric GSCPI specifications
No asymmetric GSCPI specification lowers the one-quarter out-of-sample RMSE below Model R's by a margin that either of the two tests in the note below resolves.
The highest optimum found under each floor
The search under the floor of the out-of-sample evaluation
Out of sample, by initial optimum and restart rule
The quarter in progress
Forecasts from each completed quarter
Each row is a forecast of the following quarter, made from a completed quarter by Model R and by the AR(1).
Sources and method
Core PCE inflation is the year-over-year change of the Bureau of Economic Analysis Core PCE price index. The Global Supply Chain Pressure Index (GSCPI) is the Federal Reserve Bank of New York's, and the prices-paid index is the diffusion index of the Federal Reserve Bank of Philadelphia Manufacturing Business Outlook Survey. The output gap is 100 × (real GDP / potential GDP − 1), real GDP from the Bureau of Economic Analysis and potential GDP from the Congressional Budget Office; the producer price index for finished goods, unit labor costs of the nonfarm business sector, and import prices are the Bureau of Labor Statistics'; and the effective federal funds rate is the Federal Reserve Board's. These are read through Federal Reserve Economic Data (FRED) of the Federal Reserve Bank of St. Louis, the GSCPI and the prices-paid index from their publishers' files. First releases are from the Federal Reserve Bank of Philadelphia's Real-Time Data Set for Macroeconomists.
The equations forecast Core PCE inflation one quarter ahead from regressors dated at the origin, estimated by ordinary least squares on data from 2012:Q1 through 2025:Q4, the predicted quarters 2013:Q3 through 2025:Q4 (50); the out-of-sample evaluation re-estimates them on an expanding window at each origin. The Overview, Estimates, Out of sample, and Regime search tabs show the paper's values, computed from the data as published by 9 October 2026, the prices-paid index not seasonally adjusted; the Overview's chart plots the paper's data series.
From each completed quarter, Model R and the AR(1), with the paper's coefficients, not re-estimated, forecast the following quarter from the three series as published on the run date; the realized value enters when the target quarter completes. A quarter enters the table of forecasts when all three months of every series are published; the paper, by contrast, takes the quarter of a monthly series when at least two of its three months are published (Appendix A).