Explainable by design. Deterministic by construction.

Kareg is generated by an automated system using statistical methods (rolling normalization and threshold-based classification). It is not a machine learning model and is reproducible given the same inputs.

Pipeline

Five deterministic stages — no secret sauce.

01

Ingest & normalize

Compute standardized features from macro and market inputs using rolling windows. Normalization methods are selected to remain stable across market conditions.

02

Stress mapping

Convert standardized features into a common non-negative stress scale using driver-specific directionality rules. This makes heterogeneous inputs comparable without changing their meaning.

03

Aggregate into risk score

Combine driver stress values into a single composite risk score. Driver contributions are decomposed so you can see what influenced the score on any given day.

04

State machine classification

Deterministic policy rules translate the composite score into a regime label. Buffers and minimum hold periods are used to reduce boundary churn near transitions.

05

Publish point-in-time artifact

Produce a self-contained artifact with integrity metadata and verification checks. No lookahead, and no later edits that rewrite history.

Scoring drivers

Five macro inputs. No hidden features.

Credit StressCredit
US, EU
Equity VolatilityRisk
US, EU
Inflation GapPrices
US, EU
Yield Curve SlopeRates
US, EU
Sovereign FragmentationFragmentation
EU only

Drivers, definitions, and methodology are documented per model version.

State machine

Buffers and minimum hold periods prevent boundary churn.

Transition logic (conceptual)

Enter AMBERwhen the score crosses a stress threshold
Exit AMBER → GREENwhen the score falls below a buffer
Enter REDwhen the score crosses a crisis threshold
Exit RED → AMBERwhen the score falls below a buffer

Minimum hold periods

Minimum hold periods prevent rapid flip-flopping during noisy transitions. Hold logic is versioned and documented.

Limitations (honest and useful)

Every model has edges. We document ours so you can plan around them.

Statistical relationships may not hold during unprecedented events.
If a data source is disrupted, fallback handling can reduce sensitivity to rapid changes.
A three-state label is intentionally simple and loses nuance at transitions.
Macro-level signals do not account for sector-specific or instrument-specific risks.
Some drivers may be correlated. We do not orthogonalize them to preserve interpretability.
View sample output