Claude FM: Claude Forecasting Model

Rustam JamilovEconomics Fellow

Peter McCroryAnthropic (Mentor)

October 4, 2026

Claude's Real-Time Forecasts of the US Economy

Outcome, first release SPF median Claude Claude 80 percent interval

Forecast accuracy summary statistics

MAE: Mean absolute error, the average size of the miss in the variable's own units. Lower is better.
RMSE: Root mean squared error, which weighs large misses more heavily than small ones.
Claude better: The share of forecasts where Claude's miss was smaller than the SPF's. It counts wins and says nothing about their size.

Live US Forecasting Track: Committed until 2027Q3

Outcome, first release SPF median Claude

Model Cutoff Dates

Claude model	Training cutoff	Obs.	Blind	Blind, with outcome
Haiku 4.5	31 July 2025	60	11	11
Sonnet 5	31 January 2026	60	11	11
Fable 5.0	31 January 2026	60	11	11
Opus 5	31 May 2026	60	8	8
Sonnet 5.5	30 June 2026	60	8	8
Opus 5.5	30 June 2026	60	8	8

Notes: Each model forecast 60 observations over the 3 rounds: four variables at five horizons in each round. A forecast is blind when the target's outcome was published after the model's training cutoff. Sonnet 5.5, Opus 5.5 and Opus 5 have the fewest because their cutoffs are the latest: the 2026Q1 outcomes predate them.

So far 57 forecasts have a published outcome, all from the backfilled rounds. The largest miss is CPI inflation for 2026Q2, which came in at 6.07 percent. Claude's average miss was 2.77 points and the SPF's 1.74.

Live track: forecasts against the published outcomes

Variable	Target	Obs.	Outcome	Claude error	SPF error	Winner
CPI inflation	2026Q2	12	6.07	2.772	1.735	SPF
Real GDP growth	2026Q1	3	1.99	0.085	0.660	Claude
Real GDP growth	2026Q2	12	1.50	0.457	0.601	Claude
Unemployment rate	2026Q1	3	4.33	0.006	0.067	Claude
Unemployment rate	2026Q2	12	4.27	0.091	0.153	Claude
3-month Treasury bill	2026Q1	3	3.59	0.008	0.057	Claude
3-month Treasury bill	2026Q2	12	3.62	0.094	0.045	SPF
All pooled		57		0.724	0.575	SPF
By model, the same observations
Sonnet 5.5		8		0.594	0.634	Claude
Opus 5.5		8		0.590	0.634	Claude
Fable 5.0		11		0.715	0.532	SPF
Opus 5		8		0.924	0.634	SPF
Sonnet 5		11		0.732	0.532	SPF
Haiku 4.5		11		0.771	0.532	SPF

Notes: Forecasts with a published outcome, from the 2026Q1 and 2026Q2 rounds. Each observation is one model's forecast of the target from one round, the three draws averaged, where the outcome was published after the model's training cutoff. Errors are mean absolute errors against the first published estimate. Claude is closer than the SPF median on 5 of the 7 targets, and pooled over the 57 observations its error is 0.724 against the SPF's 0.575.