A Floor, Not a Spike — Stock-Bond Diversification in Brazil

Why Brazil has a floor

Textbook stock-bond correlation is set by which shock dominates: growth shocks push stocks and bonds apart, inflation shocks push them together. Brazil adds a third channel the framework leaves out — sovereign credit risk. When fiscal credibility slips, equity valuations and government bond prices fall together, because both are pricing the same probability of adjustment, restructuring or monetisation.

That channel doesn't make the correlation large. It makes it never reliably negative. Brazilian bonds are diversifiers, not hedges — and the advanced economies the IMF wrote about are losing a hedge Brazil never had to lose.

ρ daily · Ibov × NTN-B +0.122 Diversifier, not hedge

ρ monthly · same pair +0.405 Horizon triples it

Floor · 60-month low +0.073 Never negative in 21 years

Left-tail co-crash 2.9× More often than chance allows

Crises cushioned 4 of 6 Both failures were domestic

Ibovespa Sharpe vs CDI −0.07 Equities lost to cash

Fig. 1 No macro regime turns negative §4.2

Six regimes, two decades, one sovereign: none of them reaches hedge territory. The drift after 2013 looks like a trend but sits inside the noise band — no regime is distinguishable from the calmest. The floor is the finding; the level isn't.

Fig. 2 Brazil never had the hedge the G4 are losing §9.4

Lowest five-year correlation each market has ever printed. The G4 spent time below −0.45 and could go back if balance sheets re-expand — Brazil has no negative regime to revert to. That asymmetry, not today's level, is what makes Brazil the destination case.

Fig. 3 Most crisis spikes are volatility artefacts §6

Crisis windows carry five to ten times calm-period equity variance, and that alone inflates a measured correlation. Correct for it and only Joesley Day survives as real contagion — a domestic political shock. The familiar claim that diversification fails exactly when you need it rests mostly on the artefact.

Fig. 4 Duration is what co-moves — and it takes weeks §4.3

The daily number is the smallest one available. An allocator who rebalances monthly lives at +0.405, not +0.122 — which is also why daily Brazilian estimates and the IMF's monthly ones were never the same statistic. LFT stays flat at every horizon: no duration, nothing to transmit.

Fig. 5 The switch is shock type, not crisis §8.2

Global risk-off and the bond leg earns its place; a domestic fiscal-political shock and it doesn't. The two episodes where NTN-B lost to cash were both home-grown — so the question to ask of a drawdown is where the shock came from, not how large it is. Note the second series: the bond leg cushions, it does not rescue.

Fig. 6 Where it does fail: the left tail §5.2

Duration bonds crash with equities nearly three times as often as independence implies, and more often than they rally together — the failure is asymmetric and it is concentrated in the days that hurt. LFT is the exception, which is the point: it has no duration to be repriced.

What this changes about a Brazilian 60/40

Duration diversifies global risk-off and amplifies domestic fiscal stress, so no allocation inside domestic fixed income turns these bonds into a hedge. Against a home-grown fiscal shock the hedges are short LFT and international assets — and long LFT is not the risk-free instrument it is assumed to be, drawing down 1.33% at its longest maturity against 0.008% at one year.

The harder question the data raise is what the equity leg is for: the Ibovespa lost to cash over twenty-one years, at 26% volatility and a 60% drawdown. A domestic 60/40 was paying for equity risk out of a negative premium while holding a bond leg that co-crashes in the left tail.

Qualifications

No regime is statistically distinguishable from the calmest — all bootstrap p > 0.05, and regime boundaries are researcher-defined rather than estimated.
Convergence toward Brazil is directionally unanimous — four of four G4 gaps narrowed — but significant only for Germany. Four correlated bond markets moving together is closer to one observation than to four.
Joesley is an 11-day window, so that estimate is imprecise; the DCC time variation is weakly identified and reported as descriptive.
The earlier draft's real-rate and news-sentiment findings rested on a notebook that had never run and are not reported — those claims are open, not established.

Every figure comes from docs/04_final_paper.md, which scripts/run_analysis.py regenerates from free public data (IPEADATA, BCB SGS, Tesouro Transparente). Method, all nineteen tables and the full limitations are in the paper.