What might happen next — and how sure can we be?
Forecast Center
Twelve-month forecasts for inflation, unemployment, and payroll growth — each backtested against a naïve baseline with rolling-origin evaluation, published with empirical prediction intervals, and selected on evidence rather than sophistication.
How to read this page
Every target compares simple baselines against classical time-series models. Models are refit at 60 historical origins and asked to forecast 1, 3, 6, and 12 months ahead; the published model is the one with the lowest average error that beats the naïve baseline — and when nothing beats the baseline, the baseline is published. Shaded intervals are quantiles of the errors the selected model actually made, so “80%” is a measured claim, not an assumption. Forecasts are statistical estimates, not guarantees.
CPI inflation
The selected model — Autoregression (AIC-selected order) — puts CPI inflation at 2.8 % year-over-year in Jun 2027, with an 80% interval of 0.4 to 6.7. Intervals are empirical — they reproduce the errors this model actually made in 60 historical trial runs.
CPI inflation — history and 12-month forecast
Autoregression (AIC-selected order) · trained Jan 1948–Jun 2026 · % year-over-year
Source: CPIAUCSL (BLS via FRED) · Model v1.0.0 · Run 2026-07-18
Backtest — which model earned the spot?
Autoregression (AIC-selected order) achieved the lowest average MAE across horizons, 21% below the Naïve (last value) baseline. Errors below are mean absolute error from 60 rolling-origin trials at each horizon — read the selected model's 12-month column as “ historically off by about 2.0 % year-over-yearon average”.
| Model | MAE, 1mo | MAE, 3mo | MAE, 6mo | MAE, 12mo | Avg MAE |
|---|---|---|---|---|---|
| Autoregression (AIC-selected order)Selected | 0.2 | 0.6 | 1.1 | 2.0 | 1.0 |
| Naïve (last value)baseline | 0.3 | 0.8 | 1.4 | 2.4 | 1.2 |
| Random walk with drift | 0.3 | 0.8 | 1.4 | 2.4 | 1.2 |
| SARIMA (1,0,1)(1,0,1,12) | 0.3 | 0.8 | 1.5 | 2.9 | 1.4 |
| Seasonal naïve (same month last year) | 2.6 | 2.6 | 2.5 | 2.4 | 2.5 |
Limitations of this forecast
- Univariate model: no exogenous drivers (policy, energy prices, fiscal shocks).
- Trained on current-vintage (revised) data; real-time accuracy would be lower.
- Backtest window includes the pandemic regime shift; errors reflect it.
- Missing federal releases interpolated for: 2025-10-01.
Unemployment rate
The selected model — Naïve (last value) — puts Unemployment rate at 4.2 % in Jun 2027, with an 80% interval of 1.8 to 4.6. Intervals are empirical — they reproduce the errors this model actually made in 60 historical trial runs.
Unemployment rate — history and 12-month forecast
Naïve (last value) · trained Jan 1948–Jun 2026 · %
Source: UNRATE (BLS via FRED) · Model v1.0.0 · Run 2026-07-18
Backtest — which model earned the spot?
No candidate beat the Naïve (last value) baseline on average MAE across horizons; the baseline is published. This is reported honestly rather than hidden. Errors below are mean absolute error from 60 rolling-origin trials at each horizon — read the selected model's 12-month column as “ historically off by about 0.9 %on average”.
| Model | MAE, 1mo | MAE, 3mo | MAE, 6mo | MAE, 12mo | Avg MAE |
|---|---|---|---|---|---|
| Naïve (last value)Selected | 0.2 | 0.4 | 0.6 | 0.9 | 0.5 |
| Random walk with drift | 0.2 | 0.4 | 0.6 | 0.9 | 0.5 |
| Autoregression (AIC-selected order) | 0.3 | 0.7 | 0.8 | 1.2 | 0.8 |
| Seasonal naïve (same month last year) | 1.7 | 1.5 | 1.3 | 0.9 | 1.4 |
Limitations of this forecast
- Univariate model: no exogenous drivers (policy, energy prices, fiscal shocks).
- Trained on current-vintage (revised) data; real-time accuracy would be lower.
- Backtest window includes the pandemic regime shift; errors reflect it.
- Missing federal releases interpolated for: 2025-10-01.
- The unemployment rate is bounded below; near historic lows, downside interval width is mechanical, not informative.
Payroll growth
The selected model — Naïve (last value) — puts Payroll growth at 57 thousands of jobs, month-over-month in Jun 2027, with an 80% interval of -430 to 205. Intervals are empirical — they reproduce the errors this model actually made in 60 historical trial runs.
Payroll growth — history and 12-month forecast
Naïve (last value) · trained Feb 1939–Jun 2026 · thousands of jobs, month-over-month
Source: PAYEMS (BLS via FRED) · Model v1.0.0 · Run 2026-07-18
Backtest — which model earned the spot?
No candidate beat the Naïve (last value) baseline on average MAE across horizons; the baseline is published. This is reported honestly rather than hidden. Errors below are mean absolute error from 60 rolling-origin trials at each horizon — read the selected model's 12-month column as “ historically off by about 283 thousands of jobs, month-over-monthon average”.
| Model | MAE, 1mo | MAE, 3mo | MAE, 6mo | MAE, 12mo | Avg MAE |
|---|---|---|---|---|---|
| Naïve (last value)Selected | 224 | 255 | 298 | 283 | 265 |
| Random walk with drift | 224 | 256 | 299 | 286 | 266 |
| Autoregression (AIC-selected order) | 828 | 382 | 412 | 244 | 466 |
| Seasonal naïve (same month last year) | 761 | 715 | 698 | 283 | 614 |
Limitations of this forecast
- Univariate model: no exogenous drivers (policy, energy prices, fiscal shocks).
- Trained on current-vintage (revised) data; real-time accuracy would be lower.
- Backtest window includes the pandemic regime shift; errors reflect it.
- No imputed observations.
- Payroll figures are revised twice after first release and benchmarked annually; month-one values differ from the history modeled here.