CPI forecast dashboard.

Short-horizon CPI forecasts — one month ahead — are increasingly mechanical exercises. The major investment banks build bottom-up models that estimate each major component (gasoline from retail price tracking, used cars from Manheim auction data, airfares from booking platforms, shelter from market rent indexes adjusted for the BLS lag) and aggregate them with current CPI weights. Variance around the consensus on the eve of a release is typically just a few basis points on month-over-month.

The hard forecasting work is at 6- to 12-month horizons, where forecasters must judge how persistent the current trend will be. Three approaches compete:

1. Statistical / time-series models

Univariate models like ARIMA or unobserved-components state-space models exploit the persistence in inflation. They work well when the regime is stable but fail at turning points. Most academic 'naive benchmarks' fall in this bucket.

2. Structural / Phillips-curve models

Inflation is modeled as a function of slack (unemployment gap, output gap), import prices, energy prices, and inflation expectations. These models explain medium-run dynamics well but have been famously poor at forecasting actual numbers post-2008. The "Phillips curve is flat" debate of the 2010s reflected this.

3. Nowcasting / high-frequency data

The Cleveland Fed's Inflation Nowcasting model combines available high-frequency data (retail gasoline prices, ApartmentList rents, used-car auction prints) with statistical models for the unreported components. It's been one of the most accurate publicly available forecasts in recent years.

The honest answer is mixed. Through the long Great Moderation (1983–2019), consensus 12-month CPI forecasts were quite accurate — typical mean absolute errors of 0.5 percentage points or less. The economy was stable, and a 'this year looks like last year' forecast worked.

The 2021–2022 inflation surge was a catastrophic forecast miss for the consensus. The median Bloomberg forecast for December 2021 CPI made in mid-2020 was around 2.0%. The actual figure was 7.0%. Practically no major forecaster anticipated the magnitude. The Federal Reserve's own forecasts — published in the Summary of Economic Projections — missed by similar amounts. The miss reflected both the unusual size of the fiscal stimulus and the unprecedented supply-side shocks; standard models had no template.

The 2022–2024 disinflation was easier to forecast in direction (consensus correctly anticipated declining headline CPI) but persistently wrong in pace. The market-implied breakeven path consistently saw faster disinflation than what materialized, which is one reason real yields rose so much over that period.

The general lesson: in stable regimes, forecasts are useful. At regime changes, the consensus reliably misses, and the best forecasters are those who recognize the regime shift early — a hard problem to solve systematically.

Four times a year, the FOMC publishes its Summary of Economic Projections (SEP). The headline 'dot plot' shows each participant's projected federal funds rate path, but the same document contains projections for PCE inflation, core PCE inflation, GDP growth, and unemployment.

The SEP projections are useful for two reasons. First, they reveal the FOMC's collective view of where inflation is going. Second, the dispersion across dots reveals disagreement among policymakers, which is itself information — a tight cluster suggests consensus, a wide spread suggests genuine uncertainty about the path.

The SEP uses PCE rather than CPI, but historically core PCE runs about 0.3 percentage points below core CPI on average. So a 2.4% core PCE projection roughly maps to 2.7% core CPI. See our CPI vs PPI vs PCE page for the full comparison.