The Brazilian Macro Regime Score: A Monthly Composite of Domestic Macro-Financial Rotation

Radar Perene Working Paper Series, No. RPWP-005 · Version 2.0 · July 2026 Observers of the Brazilian economy face a persistent measurement problem: the instruments that describe the domestic macroeconomic environment are either official statistics — published with lags of one to three months and subject to revision — or financial-conditions frameworks imported from economies whose monetary structure Brazil does not share. Neither answers, at a monthly cadence and without hindsight revision, the question that sits beneath every allocation discussion in the country: is the Brazilian market currently rotating as it has historically rotated in benign domestic macro environments, or in hostile ones? This paper documents the Brazilian Macro Regime Score, the instrument with which the Radar Perene research program reads the domestic macro-financial surface — the "Brazil from the inside" axis of the program's orthogonal triad. Version 2.0 also documents, with the same prominence as the instrument itself, the evidence that the score only partially fills that gap: its categorical labels are hindsight-informed, a fact this version quantifies rather than implies away. The score is a monthly, month-end composite on a 0–100 scale, summarized publicly as a categorical regime label — defensive, neutral, or pro-risk. Rather than aggregating official macroeconomic statistics, it reads the macro environment through the market's own cross-sectional rotation: its inputs are year-over-year performance spreads of economically meaningful market segments against the broad index, together with macroeconomic derivative series, organized into eight component categories — size preference, interest rates (inverted), foreign exchange (inverted), liquidity, breadth, defensive rotation (inverted), real assets, and credit stress. Components are normalized as percentiles of the full available sample and aggregated over the components valid in each month. The production history comprises 198 monthly observations from January 2010 through June 2026 with zero gaps, zero null scores, a single calculation version, and a recorded interpretive narrative attached to every month — an audit discipline this paper documents. Two construction facts are disclosed prominently: the composite operated with a reduced component set from 2010 to 2012 (four to six of the eight categories), so early readings are not strictly comparable with the full-composition period; and the lowest of the five public bands has never been observed in the sample. The score is for locating the present month within the recorded history of Brazil's domestic macro-financial rotation: naming the regime and providing a documented monthly conditioning variable for research. It is not a forecast or a timing signal. The validation section reports, with all statistics verified against the production database as of July 2026, that the relationship between regime labels and subsequent index returns in this sample is inverse at longer horizons — defensive regimes preceded the strongest twelve-month windows and pro-risk regimes the weakest — and prints both candidate explanations: mean reversion of the year-over-year spreads on which the score is built, and confounding between the early pro-risk cluster, the reduced-composition period, and the specific path of the 2010s. The effective number of independent regime spells is far smaller than the monthly observation count, and every statistic is presented as descriptive of one decade and a half of Brazilian history, not as a property of the instrument. Version 2.0 records the outcomes of the two pre-declared experiments the instrument's remaining claims required, and both outcomes are adverse; they are reported here with the same prominence as anything favourable. First, real-time reproducibility (Section 7.5): under strictly recursive (vintage) recomputation of the normalization, the score's level nearly survives — Pearson correlation 0.905 with the published series, mean absolute difference 4.91 points on the primary 2013-onward window — but the categorical label, the instrument's declared public resolution, changes band in 36.4% of months (kappa 0.425), failing the pre-declared 30% tolerance; the recursive computation produces 45 band transitions where the published record has 34, and only 21 of the 34 published transitions have a recursive counterpart within two months. The full-sample record is therefore hindsight-informed wherever the label is read, and this version says so at every point where labels appear. Second, the trivial-benchmark race (Section 7.6): on the shared-threshold descriptive task of dating four pre-specified macro episodes, the composite classifies 78.2% of episode months correctly while a single one-line benchmark — the small-cap/index spread percentile, one of the composite's own components — classifies 93.6% and correlates +0.846 with the composite; the composite's remaining defences are label stability (median spell 4.0 months against the trivial benchmark's 2.0) and one episode (2024–25). The claim that the eight-category architecture adds information beyond a trivial one-line spread is accordingly withdrawn. Third, a hard fact the race exposed: the composite largely missed its own showcase episode — during the 2015–16 crisis only 6 of its 18 months printed in the defensive band, and most of 2015 printed neutral; earlier statements that the crisis "resided in the defensive band while it was happening" are retracted in this version. On the other side of the ledger, an independent replication reproduced the entire published series to within 0.05 points with zero label divergences, and this version corrects the paper's earlier proprietary-weights rhetoric: the production formula is an equal-weight mean of full-sample percentiles of eight public series, maintained in versioned SQL — the most fully disclosed instrument in this series, stated now as a property rather than an admission. What the paper claims after these results is narrower and, the program holds, more durable: a sixteen-and-a-half-year governed monthly regime register with its defects documented, and a measurement lesson — under real-time computation, the level of a full-sample-percentile composite is robust and its categorical label is not. Version 2.0 is the first Zenodo deposit of this paper (deposited together with the Version 2.0 batch of RPWP-003, RPWP-004, and RPWP-101). It is the adverse-experiments round: both pre-declared existential tests are reported in full — the vintage experiment (the categorical label changes band in 36.4% of months under strictly recursive computation, failing the pre-declared 30% tolerance; the recorded label history is hindsight-informed) and the trivial-benchmark race (78.2% against 93.6% for a one-line component spread; the architectural-superiority claim is withdrawn) — together with the 2015-16 defensive-band retraction, the corrected proprietary-weights rhetoric (the formula is an equal-weight mean of percentiles of eight public series, fully disclosed), and the corrected spell statistics. All statistics verified against the production database (two byte-identical runs).