Methodology & data
A forecast is more useful when you can see what it measures, what supports it, and where it stops.
This edition
RESEARCH ONLY- Forecast snapshot
- 15 September 2026, 02:54 UTC
- Rate baseline → target
- 31 August → 31 October 2026
- Country coverage
- 195 in the registry; 54 issued rate forecasts across 33 authorities
- Unavailable rate forecasts
- 126 missing or stale origin rates; 15 separate policy frameworks
- Selected CPI comparison
- 56 countries with research output
- Source run
- 05f514887c465dbb85f48cd10dcf51e35f9b6b3105b975bcfaeb53add4eb4ace
- Updates
- Forecasts and the data library are imported snapshots. Country facts retain their observation years. Connected calendar sources are refreshed at most every 15 minutes on request, with last-success times and failed-feed states shown. An engine refresh alone does not republish these snapshots.
Month-end, not meeting-day
Rate forecasts describe the October month-end benchmark relative to August month-end. The cumulative change may span multiple central-bank meetings. A 25bp move is a 0.25 percentage-point change.
Instrument definitions differ. For example, the US series is a target-band midpoint, and euro-area members share the ECB deposit facility rate. The instrument and its evidence appear on each country page.
Different questions, different models
The direction-focused point model chooses a cut, hold, or hike and a rate estimate. The primary size model separately estimates 11 outcomes, including four named cut sizes, four hike sizes, hold, and other cuts or hikes.
Probabilities are raw and uncalibrated. Prior holdout calibration was not applied where per-class support was insufficient. The most likely size outcome may differ from the point forecast.
Accuracy needs a denominator
The country scorecard maps historical next-month performance to current monetary authorities. 59 countries have mapped scores, 121 are not backtested, and 15 use inapplicable scalar-rate frameworks. Shared-authority results are not independent country evidence.
Balanced accuracy averages cut, hold, and hike recall. Overall accuracy weights the observed cases, which are dominated by holds. Magnitude metrics in the current country evidence ledger describe the zero-change persistence comparator, separately from direction-model performance.
Research is not validation
The benchmark uses revised historical data and assumed publication lags. It does not reconstruct verified historical vintages. Development history has been reused, and within-month updates need historical replay.
No model is promoted for production, and the worldwide 90% accuracy goal remains unmet. A one-month historical score cannot establish the reliability of a two-month forecast or guarantee future accuracy.
Inflation and employment
Local CPI, PPI, and unemployment models provide forecasts at 1, 2, 3, and 6 months from the August origin. Stale or missing data blocks issuance. Each prediction retains its input date and candidate model.
The October CPI comparison uses prior validation to select local, global, or persistence predictions. Raw global CPI candidates failed their initial benchmark. Raw-model intervals must never be applied to the selected forecast.
Relationships are not causation
The global network treats unique monetary authorities as rate donors, and national CPI series separately. More donor series or supported associations do not establish stronger predictive performance.
Alternative-cutoff results hold the model and local inputs fixed. They measure sensitivity to information and aging, not causal contributions or changes from a previously issued forecast. The raw global rate experiment remains unpreferred.
Explore the underlying data
CSV & JSONThe exports retain the original research fields and missing values. The snapshot manifest records SHA-256 hashes for the forecast exports. Country pages also provide a complete country record.
Source-grounded, with gaps visible.Coldstreet presents the supplied Global Macro Engine research outputs. Official-source references and instrument caveats are retained on country pages; inclusion in the registry does not imply complete or verified source coverage.
Explore all countries