Pancis 1.0
AI model
Understand the outlook. Inspect the evidence.
Pancis 1.0 brings Coldstreet’s macroeconomic forecasting research, data library and methodology into one place. It is a suite of statistical and machine-learning research workflows, with separate models for policy rates, inflation and employment.
Coverage is broader than forecast availability
August origin → October 2026 month-end
No model promoted to production
Mixture of experts
Three specialist models. One evidence-based routing decision.
How routing uses historical evidence
Annual selection uses only outcomes before the forecast year. Minimum support is checked for authority + cycle, authority, cycle and global evidence. Small samples are adjusted toward global performance. The selected expert supplies all three scores.
Authority & cycle
Policy history
Calendar & macro signals
Current cycle segment
Select an expert
Best supported, adjusted balanced accuracy
Selected scores
Use the selected expert’s three direction scores.
Next-calendar-month research targetBaseline champion
The frozen recent-own-calendar / PPI-missingness expert supplies the baseline direction scores. It is also the fallback when the router lacks sufficient historical support.
Click an expert above to inspect its role. This highlights the architecture; it does not change a live forecast.Annual routing uses matched out-of-fold predictions with target dates strictly before January 1 of the forecast year.
Local class recall is adjusted toward global evidence. Each routing level must pass minimum sample and cut/hold/hike support requirements.
This experiment uses hard routing: it selects a single expert for an authority and cycle segment rather than averaging all three score vectors.
Architecture evidence & evaluation
The configured expert priority is champion → compact → full. Minimum support is 72 rows and 4 cases per class for authority + cycle; 96 / 5 for authority; and 180 / 10 for cycle or global routing. Local estimates use a global prior equivalent to 25 rows per class.
On the reused 7,583-case, 39-authority development set, this MoE candidate reached 68.36% balanced accuracy, compared with 68.76% for the frozen champion. Recent-period scores were 59.06% and 58.69%, respectively. These next-month results do not validate the current two-month outlook.
Download architecture & source references ↗Documented support-shrunk mixture-of-experts experiment. It has not been promoted and does not replace the current country forecast selector.
How Pancis works
- 01
Gather the evidence
Policy-rate histories, inflation, producer prices and unemployment form the forecasting evidence. Source records retain their dates, units and available provenance. Country fundamentals and the event calendar add context; their presence in the library does not mean every field is a model input.
- 02
Align countries and time
Countries are mapped to their monetary authority and policy instrument. Members of a currency union share an authority rather than becoming independent rate observations. Histories are aligned to a forecast origin; missing or stale inputs can block issuance.
- 03
Generate candidate forecasts
Separate models answer separate questions: the direction of a rate move, the distribution of possible move sizes, and future macroeconomic values. Local forecasts, cross-country candidates and persistence baselines are compared where supporting evidence exists.
- 04
Select using historical evidence
Research comparisons assess cuts, holds and hikes separately, examine recent performance and compare alternatives with simple baselines. The selected CPI output can come from a local, global or persistence forecast. Historical relationship strength alone does not select a winning forecast.
- 05
Publish an inspectable snapshot
Country pages show the issued forecast, its horizon, available probabilities and evidence. The library exposes charts, tables and downloads. Missing outputs remain visible. Forecast snapshots need an explicit import and publication; a calendar refresh does not retrain the models.
Direction and size are different questions
The direction-focused model issues a cut, hold or hike outlook and a rate estimate. A separate size model assigns probabilities to 11 outcomes: hold, four named cut sizes, four hike sizes, and other cuts or hikes.
The most likely size category can differ from the point forecast. Probabilities are raw and uncalibrated, so a displayed 70% is not an established 70% success frequency.
Inflation and employment over time
Local CPI, PPI and unemployment forecasts cover 1, 2, 3 and 6 months. Each output retains its candidate model and input date. The selected CPI comparison uses historical validation to choose among local, global and persistence predictions.
Intervals from a raw global candidate do not describe a different, selected forecast. Dashed chart lines connect forecast endpoints; they do not predict the path between them.
Explore associations across economies
Relationship charts compare historical changes in national inflation and unique monetary-authority rate series. They show the sign and strength of a lagged association alongside the number of paired months.
Correlation does not establish causation, model feature importance or predictive value. These networks are descriptive research, and the raw global rate experiment remains unpreferred.
Explore inflation relationships ↗Read the score with its denominator
Balanced accuracy averages recall for cuts, holds and hikes. This prevents frequent holds from dominating the headline as much as ordinary accuracy. Country scores share evidence when they share a monetary authority.
Current historical benchmarks concern next-month direction. They do not validate the two-month live outlook. Magnitude errors in the country evidence ledger describe the zero-change persistence comparator.
Inspect model comparisons ↗Read a forecast in four steps
- Check the instrument and dates. A month-end benchmark forecast can span several meetings; 25 basis points means 0.25 percentage points.
- Separate the estimate from the scenarios. The point estimate and size probabilities answer different questions.
- Inspect the supporting history. Check sample size, recent performance, input freshness and source coverage.
- Keep uncertainty in view. Missing data, revised histories and uncalibrated probabilities limit what can be concluded.
What “1.0” means. This names the Coldstreet model hub and its current research edition, powered by the supplied Global Macro Engine outputs. It is not evidence of a newly trained or independently validated model. Historical development periods were reused, verified point-in-time vintages are incomplete, and the worldwide 90% accuracy goal remains unmet.