Macro Regime Classification: Methodology and Data

What this classification computes

For any given month, the classification answers: what macro regime is the US economy in, based only on public data and a verifiable rule?

Every term is verifiable: the regime name follows from the growth × inflation grid (see Architecture), the financial conditions qualifier from the Chicago Fed’s NFCI, the global context from OECD CLIs and the ECB’s CISS.

What this classification is NOT

• A proprietary score (no opaque weighting)
• A trading signal or investment recommendation
• A forecast of future conditions
• An official cycle dating (≠ NBER)
• A portfolio timing or allocation system

Three independent axes

The classification uses three independent axes — real activity, inflation, financial conditions — rather than a flat list that conflates orthogonal dimensions. These three axes are the inputs of the calculation: what the engine measures to produce a verdict. They must not be confused with the three layers of the Atlas (see below), which describe the nature of the regime objects, at different horizons and evidentiary statuses. Two distinct reading grids: the axes are the ingredients, the layers are the dishes.

The 3×3 growth × inflation grid maps to seven named regimes plus a transition state.

A hierarchical world view, not an aggregated score

A single aggregated world score masks inter-zone divergence (e.g., 2022: US tightening vs China easing) — a misleading indicator. The architecture is hierarchical:

A eurozone core on the same architecture (ECB and Eurostat inputs) is on the roadmap and is not part of v1.1.0, which calibrated the US core only. The euro area currently enters the classification through one channel: the ECB’s CISS, in the global context layer.

The seven regimes and the transition state

Each regime below is followed by its verified occurrences in the published series (regime_history_v1.1.0.csv, 283 months, January 2003 → July 2026). These are the months the classifier actually assigns, not illustrative labels — every one can be checked by filtering the file. Episodes before 2003 fall outside the published window and are flagged as illustrative.

The overlay prefixes the name: “Slowdown under acute financial stress” (the actual reading for October 2008 to April 2009), “Balanced Expansion under accommodating conditions” (2021).

Three axes in, three layers out

The previous section describes the seven regimes of the G × I grid: this is layer 1 of the Atlas, the one the engine computes in real time. But the Atlas of regimes contains two other kinds of objects, read at different time horizons and with different evidentiary status. To avoid confusion: the three axes above are the inputs of the calculation (growth, inflation, financial conditions); the three layers below are the nature of the regime objects. They are not two names for the same thing.

US core and global context inputs

US Core

* HY OAS: optional corroboration only. Since April 2026, FRED truncates ICE BofA series to 3 rolling years. Included only when a full pre-April 2026 historical fixture is available locally. Never written to the published output.

† SOS: specified as an early-warning gate on the growth axis, but the Richmond Fed series is not reliably retrievable programmatically. The column is empty on all 283 rows of the published series and the gate has never fired. Retained in the schema so the gap is documented rather than hidden.

Global context layer

The OECD CLIs, the Cboe VIX and the ICE BofA series are used as computation inputs but are never written to the published output — only the categorical labels derived from them survive, in global_sync and global_qualifiers. None of the global context inputs can change the regime name.

Several further inputs were named in the original design and are not implemented in v1.1.0: net liquidity (WALCL − WTREGEN − RRPONTSYD), the Wu-Xia shadow rate (Atlanta Fed updates suspended since April 2022), the copper/gold ratio, and EM high-yield spreads. None is fetched, computed or published. They are listed here so the gap between the design and the implementation is explicit rather than discoverable.

Fixed institutional thresholds

These thresholds are not Eco3min editorial choices — they are defined by the producing institutions:

Calibrated thresholds

The remaining thresholds — CFNAI-MA3 levels, Trimmed Mean PCE bands, NFCI overlay levels — are calibrated by sweeping candidate values on a grid over the full 1968–2026 history and scoring each against a ground-truth matrix registered before any tuning: NBER recession dates, per-episode expectations, and pre-declared metrics for false positives, entry latency and stability. The percentile reported for each retained value is descriptive — it says where the value sits in the distribution — not the selection criterion.

The rejected candidates and the reason each was rejected are published alongside the retained values, in the changelog of thresholds.json. Any threshold change requires a version bump and a changelog entry: reproducing a given verdict depends on threshold immutability for that version.

Hysteresis: preventing regime oscillation

Without a hysteresis mechanism, an indicator oscillating around a threshold would produce unreadable regime changes month after month. The rule:

Two resolution windows

What this system cannot do

• Revisions: the current month’s regime may change after data source revisions (CFNAI is regularly revised). The NFCI in particular is re-estimated weekly over its entire history, which is enough to flip a label sitting on a boundary. The JSON timestamps the computation and flags lagged inputs.
• Publication lags: the CFNAI publishes with ~3-week lag. The classification uses the last known value and documents the pending input in the JSON.
• OECD CLI lag: ~6 weeks. The global context layer updates slower than the US core and may delay detection of inter-zone divergence.
• Not a forecast: the classification describes the currently measured state, not future conditions.
• Energy supply shocks: the Trimmed Mean PCE excludes volatile components. An oil spike can be invisible on the inflation axis while headline CPI surges. The headline_underlying_divergence flag in the JSON signals this case automatically.
• SOS gate inert: the early-warning gate on the growth axis is specified but its source is not reliably retrievable. It has never fired, and no verdict in the published series depends on it.
• Structural regimes not computed: layer 3 regimes (secular stagnation, financial repression, fiscal dominance) are not produced by the engine. They are long-run editorial frameworks and must not be read as a measured monthly verdict.

Any third party can recalculate the regime

Any third party can recalculate the regime from primary institutional sources and the published threshold table, with no opaque component. The classification code, the threshold table and the full monthly output are published:

github.com/eco3min/macro-regime-classifier

Running it needs a free FRED API key and nothing else. No source series is redistributed in the repository: the code fetches them at runtime, and the three inputs under a restrictive licence — ICE BofA, Cboe, OECD — are excluded from every published file by design. Reproducing a past verdict exactly is not always possible, because the sources are revised: the NFCI in particular is re-estimated weekly over its entire history, which is enough to flip a label sitting on a boundary. That is why each threshold version ships a frozen snapshot rather than relying on a re-run.