Digital Economy Intelligence

Layer 12 · Touch globe to explore · Tap country for detail

DRAG TO ROTATE  ·  TAP COUNTRY FOR DETAIL

✕

Model Methodology

L1 LOGISTIC REGRESSION · WALK-FORWARD VALIDATION · 29+ SIGNALS · 2015–2023

What ZEVKORA Measures

ZEVKORA assigns a calibrated probability score to every sovereign — the estimated likelihood of a crisis event within a 24-month horizon.

A crisis event is any of: currency collapse, sovereign debt default, IMF emergency programme, or acute political rupture.

The model detects structural conditions that historically precede crises — not when a crisis will occur, but whether the conditions for one are assembling.

Walk-Forward Validation — AUC by Fold

201620172018201920202021*20222023

* 2021 AUC 0.623 — COVID shock year. Exogenous, undetectable by structural model.

Signal Pipeline — Raw Data to Probability

Signal Architecture

SignalRelative WeightStatus

Why L1 Regularization

L1 forces redundant signals to exactly zero — it's a built-in feature selector. If two signals carry the same information, one is zeroed automatically.

In practice: all 5 FRED variables were zeroed because the structural family already captured them. The model retained only signals adding genuinely new information.

This is why ZEVKORA avoids overfitting with 29+ input signals — the model selects its own features.

Ablation Experiment Results — Lift vs Baseline

Industry Benchmark

ZEVKORA: Avg 0.756 · Peak 0.896 — TOP OF STRONG BAND

Key Design Decisions

Signal Convergence

MULTI-SIGNAL ALIGNMENT · PAST 5 YEARS · PRESENT · 24-MONTH FORWARD PROBABILITY CONE