If Lula wins again, what happens to the price of Brazilian risk?

1. Abstract

Brazil holds a presidential runoff on 25 October 2026. This paper does not forecast who wins. It fixes a Lula victory as a condition and asks what a renewed left-fiscal-expansionist mandate would do to the market price of Brazilian fiscal risk over the following ten months, measured on seven falsifiable thresholds dated 30 June 2027: the five-year sovereign CDS, the long real (NTN-B) yield, the real-curve slope, the exchange rate, IPCA inflation against its target ceiling, the Selic policy rate, and PNAD unemployment. A ten-agent research committee produced the forecasts under a freeze-before-read discipline: every probability anchored to a named reference class before any mechanism argument, all 93 variants registered, and every number that could not be grounded in warehouse data refused rather than guessed. The headline is deliberately unheroic. At a freeze where the CDS sits in the bottom six percent of its own decade, the committee finds the Lula scenario largely pre-priced: 49 percent on a 15 percent CDS widening, 12 percent on a 10 percent depreciation against a counted 12 percent base rate, 70 percent on inflation breaching the ceiling, and coin flips elsewhere. Every forecast is stamped exploratory and sizes to zero; the contribution is the discipline made visible, with counts, refusals, and base rates on the record before the answer is known.

2. Introduction

Every Brazilian election since 2002 has arrived with the same warning: if the left wins, the market will make the country pay for it. The warning has a famous receipt. In 2002, as Lula first approached the presidency, Brazil’s sovereign risk premium exploded, the currency lost a third of its value, and the IMF stood by with a rescue package. It also has a famous counterexample. In 2022, Lula won again, and the market yawned: the CDS fell through his first year on a commodity tailwind and a set of orthodox early signals.

Those two episodes bracket nearly the entire outcome space, and that is the problem this paper takes seriously. The reference class for “what Brazilian markets do after a PT election win” contains four members: Lula 2002, Dilma 2010, Dilma 2014, and Lula 2022. Two of the four saw the risk premium widen materially. Any confident story about October 2026 is therefore a story told against a sample of four, and an honest forecaster has to say so before saying anything else.

We ran the question through the studio’s forecasting committee anyway, because the discipline is the point. The committee does not forecast the election. The listed markets already price the condition, and a forecast of the winner would bind the wrong claim. Instead, a Lula win is fixed as a condition, and the committee prices the consequence: seven dated, falsifiable thresholds on the instruments that mark Brazilian fiscal risk to market, each frozen against a named reference class before the outcome can be known, each carrying a registered trial count, and each scored on 30 June 2027 against the freeze.

Section 3 places the exercise beside the forecasting literature and the ATOL line it continues. Section 4 describes the machine and its rules. Section 5 reports the results: where Brazilian risk was priced at the 13 August 2026 freeze, the three dynamics the committee synthesized, the seven probabilities with the base rate beside each one, and what the machine refused to do. Section 6 reads the result: a committee that hugs the base rate is telling you where the information already lives, and the one place it counted a deep sample is the one place it dared to lean. Sections 7 and 8 state the limitations and what resolution will settle.

3. Background and Related Work

The evaluation machinery is the standard one. Reference-class forecasting, taking the outside view from a named class of comparable episodes before any inside-view adjustment, and pre-registration, freezing the claim, the benchmark, and the probability before the outcome is known, are the operational core of the superforecasting program (Tetlock and Gardner, 2015). The interval treatment of counted classes follows the standard Jeffreys construction for a binomial proportion (Brown, Cai, and DasGupta, 2001). Nothing in this paper’s statistics is novel; the contribution is enforcement in code, by a platform that refuses a number it cannot ground rather than estimating one.

On the substance, the transmission this paper prices, from a left-fiscal-expansionist mandate into sovereign spreads, real yields, the currency, and the policy rate, is the classic emerging-market political-risk channel, and Brazil supplies its canonical episodes: the 2002 confidence crisis and rescue, and the 2022 transition that repriced almost nothing. The legal architecture the committee leans on is public record: central-bank autonomy under Lei Complementar 179/2021 and the fiscal framework (arcabouço) under Lei 14.791/2023.

Within the ATOL corpus, this paper continues the market-benchmarked forecasting line of MKT-013 (ATOL Research, 2026), which froze a single market-priced question against a devigged mid. MKT-015 extends the same freeze-before-read discipline to a conditional set with no tradable venue: the null is not a market mid but a menu of declared reference classes, and the honest stamp for the whole set is exploratory. The forecasts were produced on command-center, the studio’s forecasting platform (command-center, 2026), whose committee, grounding, and trial-registry mechanics Section 4 describes.

4. Methods

The forecasts come from command-center, the studio’s forecasting platform, run as cycle 7 on frame fr_6d66bb420fb169e6 with a freeze date of 13 August 2026. The frame fixes the condition (Lula wins the 25 October 2026 runoff), the seven sub-questions, the resolution date (30 June 2027), and the measurement rule: every threshold is read against the freeze-date close of the named warehouse series, so the committee scores the conditional fiscal consequence, never the election itself.

Ten specialist agents sat the committee, weighted by relevance: emerging-market macro and rates-and-credit at full weight; domestic politics at 0.9; capital flows at 0.75; US macro, ideology, commodities, labor demographics, social psychology, and China macro filling out the roster at 0.3 to 0.5. Their work was synthesized into a junction: a force graph of ten mechanisms, three dynamics, and eight forecasts, seven scoreable and one exploratory structural tail.

Three rules bind every number in this paper.

First, outside view before inside view. Every forecast is anchored to a named reference class and its base rate before any mechanism argument is allowed to move it. The classes are declared in the frame: the four PT-transition windows for the Brazil-specific questions, a counted twenty-year sample of USDBRL windows for the currency, and a twenty-episode class of emerging-market inflation-targeters for the policy rate.

Second, the engine computes and the model names. Where a base rate could be counted from point-in-time warehouse data, it was counted: the currency question ran against 868 rolling ten-month windows of USDBRL history, with the count deflated for window overlap and the class falsified against out-of-sample retrodiction. Where the data could not support a count, the engine refused. The NTN-B series is not certified point-in-time, so its crossing rate was never counted; the derived curve slope has no single-series estimator; an ambiguous event type on the inflation question was refused rather than guessed, because a touch-resolved question scored against a terminal null flatters the forecaster. Of eighteen grounding attempts across the eight forecasts, two engaged and sixteen ended in a structured refusal. Every refusal is on the record, and every refused forecast fell back to its thin declared class with the thinness stamped on it.

Third, the trial registry. The synthesis registered 93 variants: 80 forecast variants, 34 narrative variants, 46 class searches. Selection pressure across variants is corrected by shrinking every inside-view deviation toward the base rate, with a shrinkage weight of about 0.31 computed from the registered counts. This is why the committee’s mechanism stories, whatever their color, move the final probabilities by fractions of a point: under a thin class and a wide search, the honest posterior barely leaves the prior.

The forecasts are pre-registered in the operator ledger with kill-criteria and content-addressed class freezes. Under the platform’s verdict rule, a conditional set of seven with no tradable venue and mostly thin classes is exploratory by construction and sizes to zero. The analysis in this paper requires no software beyond the platform itself and the committed data: notebook.ipynb runs top to bottom on the CSVs under data/ with the Python standard library only, and every figure reads those same files.

5. Results

5.1 The freeze: where Brazilian risk was priced on 13 August 2026

The committee froze its snapshot on 13 August 2026, seventy-three days before the runoff. The single most consequential fact in this paper is where the market already was on that date.

The five-year sovereign CDS closed at 123.58 basis points (Fig. 1). Across its own history from April 2017 to the freeze, 3,288 trading days, the median is 175.35 and only 5.6 percent of days closed lower than the freeze level. The instrument that is supposed to panic about a left-fiscal-expansionist mandate is priced nearer its decade low than at any election in the sample. For scale, the same series touched 374.9 in the COVID panic of March 2020 and 91.8 at its 2020 low.

The real curve is inverted (Fig. 2). At the freeze the NTN-B 2029 yields 8.04 percent in real terms, the 2035 yields 7.97, and the 2045 yields 7.45: the 2045-minus-2029 slope is minus 0.59 percentage points. Brazil pays more, in inflation-linked terms, to borrow for three years than for nineteen. Two readings coexist. The front end carries a restrictive Selic of 14.0 percent, down from a 15.0 peak, held against inflation near the ceiling. And a long-run investor is being offered roughly an 8 percent real yield, about 555 basis points over US TIPS, which is the carry anchor that has capped every BRL sell-off of the past several years.

The rest of the snapshot points the same way. The exchange rate stood at 5.1904 to the dollar, so the depreciation threshold of the currency question sits at 5.7094 (Fig. 3). IPCA inflation printed 4.44 percent year-on-year for July 2026, six basis points under the 4.5 percent ceiling of the central bank’s target band of 3.0 plus-or-minus 1.5. PNAD unemployment last printed 6.1 percent, for March 2026, a number that was already 165 days stale at the freeze and that had jumped from 5.1 in the prior quarterly reading. Leveraged funds held a crowded net long of 10,214 contracts in BRL futures. The global backdrop was benign: EMBI spreads at 1.38 percent, VIX at 14.55, the dollar index near 100.

Seventy-three days before a runoff the market is supposed to fear, Brazilian risk is priced calm to the point of complacency, or to the point of verdict. Which of those two words is right is what the seven questions test.

5.2 Three dynamics

The committee’s synthesis wired ten forces into three dynamics, and the central thesis is a collision.

The first dynamic is the fiscal-expansion risk-premium channel. A returning PT administration widens primary spending: the arcabouço fiscal spending floor, the social-transfer ratchet, real minimum-wage indexation. The market prices that impulse in the CDS and the long real end. But the transmission is blunted twice over, first because the win is largely pre-priced at a CDS in its bottom decile, and second because commodity terms-of-trade revenue cushions the fiscal accounts. On the committee’s read, the net conditional move is a modest re-pricing, and, from an inverted starting curve, more likely a steepening of the long end than a blow-out of the level.

The second dynamic is the commodity-carry stabilizer, and the committee rates it decisive, the only force in the graph with that strength. An 8 percent ex-ante real yield, 555 basis points over TIPS, makes Brazilian carry expensive to bet against; commodity export revenue and a benign global backdrop reinforce it. The counted history agrees: across 868 rolling ten-month windows since 2006, USDBRL ended at least 10 percent weaker in only 107, about 12 percent of the time. The stabilizer’s known fragility is the crowded long, 10,214 leveraged contracts that would unwind non-linearly on a China-demand or global-rates shock. Absent that shock, the anchor holds.

The third dynamic is the wage-push monetary loop. Minimum-wage indexation and transfer expansion feed services inflation toward the band ceiling; a central bank that is legally independent since 2021 answers by holding or hiking; the restrictive Selic then feeds the carry anchor of dynamic two. The loop biases inflation up and the policy path tight, with one warning attached: in 2023, the analog year of Lula’s last win, the BCB was cutting by mid-year and inflation undershot. The direction is a lean, never a law.

5.3 Seven probabilities against their base rates

Table 1 and Figure 4 report the committee’s conditional probabilities beside the base rate of the declared reference class for each question. The pattern is the paper’s finding: five of seven final probabilities sit within 1.2 points of their base rate, and the two that sit exactly on it are the two whose classes were too thin to earn any deviation at all.

Table 1. The seven scoreable questions, all conditional on a Lula runoff win and all measured on 30 June 2027 against the 13 August 2026 freeze. Probabilities as decimals. The class column names the declared reference class; k of n is its count. Only the currency question carries a counted, retrodiction-tested class; the rest are typed analog sets and are stamped thin where n is 4.

Question by question, in the committee’s own accounting.

On the CDS (question 1), the mechanism raises and the pre-pricing lowers, and they cancel to 49.4 percent against a 50 percent class. The red team notes that the low base cuts both ways: 15 percent of 123.58 is only about 18.5 basis points, a move the series clears routinely on noise, with no fiscal story required.

On the long real yield (question 2), the committee’s final answer is exactly the coin flip of its 2-of-4 class, because the two live mechanisms, a fiscal term-premium lift and an already-elevated 8 percent starting level, are graded moderate in both directions, and the class was refused a count on data-certification grounds. The 2022 analog is a direct disanalogy: real yields fell as the arcabouço was accepted.

On the slope (question 3), the one surviving lean is mechanical. The curve starts inverted at minus 0.59 points, fiscal risk concentrates in the long end, and mean-reversion tilts steeper; the committee’s 0.3-logit lean shrinks, under the trial registry, to a final 50.6 percent.

On the currency (question 4), the only counted class in the set, the committee lands at 11.7 percent against a grounded 12.4. The count is the real thing: 107 of 868 rolling windows, band deflated for overlap to an effective seven disjoint windows, then widened after a walk-forward retrodiction found the class’s mean prediction of 0.030 against a realized out-of-sample frequency of 0.144. The inside view leans below the class because the carry anchor and the 2022 analog, when the BRL appreciated on a Lula win, both point down, and the crowded positioning is the tail that keeps the lean marginal.

On inflation (question 5), the committee’s 70 percent is the Jeffreys point of a 3-of-4 transition class, and it is the highest-conviction number on the board. The mechanism is the wage-push loop; the check on it is the freeze print of 4.44, six basis points under the ceiling, with the band’s own climatology giving the committee some outside comfort: our count of the committed series finds 65.8 percent of the 240 months since August 2006 printed above 4.5 (Fig. 5). The red team carries the 2023 analog, when an independent BCB held tight and the print landed inside the band.

On the policy rate (question 6), the class widens to twenty episodes of emerging-market inflation-targeters with inflation near the ceiling: twelve hiked or held higher ten months on, 59.5 percent, and the committee takes it unmoved. The red team notes the one most relevant analog, 2023, is a zero in the class: the BCB was cutting by mid-year. A Selic starting restrictive at 14.0 has more room down than up if inflation cooperates.

On unemployment (question 7), the committee reads 29.5 percent against a 1-of-4 class where the single yes, 2014-15, required a recession and an impeachment crisis. PT fiscal expansion historically pulls near-term unemployment down. The resolution carries a known defect, stated in Section 7: the September 2026 comparison baseline did not exist at freeze.

5.4 The structural tail

An eighth question is carried but never scored: over five to ten years, does the CDS mean-revert toward its 2015-2026 median or ratchet to a permanently higher plateau? The committee logs 49.6 percent on reversion below the median by end-2028, from a 6-of-12 class of EM reversion episodes, and names the distinguishing observable: whether the premium normalizes toward the long-run median or holds a structurally higher floor as the fiscal cycle turns. The red-team path to a ratchet runs through persistent primary deficits, an eroded arcabouço, or a China-led terms-of-trade downshift.

5.5 What the machine refused to do

The grounding ledger is part of the result. Eighteen grounding attempts were made across the eight forecasts; two engaged. The refusals, verbatim in kind: five because a forecast named no series base-rate request the engine could count; eight because no event class was requested; one because a relative threshold could not be resolved to a point-in-time reference level; one because the event type, touch versus terminal, was undeclared, and scoring a touch question against a terminal null flatters the forecaster; the NTN-B count because the source is not certified point-in-time. Each refusal downgraded a number from counted to typed, and each downgrade is visible in Table 1. The alternative, quietly guessing, would have produced more impressive-looking probabilities and a worse ledger.

6. Discussion

The committee walked into the most dramatized question in Brazilian markets and came back with probabilities that sit on top of their base rates. Read naively, that looks like the machine failed to have a view. Read correctly, it is the view.

The market has already voted. A five-year CDS in the bottom 6 percent of its decade, seventy-three days before the runoff the warning story is about, is the market saying the Lula scenario is not news. The committee’s job was to decide whether that verdict is complacency or information, and its synthesis answers: information, mostly. The fiscal impulse is real, and it is met by a decisive stabilizer, an 8 percent real carry anchor with commodity revenue behind it, and by the fact that the outcome is pre-priced. What remains is a knife-edge: an inflation print six basis points under the ceiling, a curve whose inversion encodes both the tight front end and the trusted long end, and a coin-flip on whether the risk premium widens at all.

The one number worth defending in an argument is the one the machine counted. Every Brazil-specific class in the set has four members, and the committee refused to pretend otherwise; the currency question alone ran against 868 windows of real history, survived a retrodiction test that widened its band, and produced the only confident number on the board: about 12 percent that the BRL is 10 percent weaker ten months after a Lula win, with the committee leaning marginally below its own class because the last Lula win saw the currency strengthen. Where the sample was deep, the committee dared to lean; where the sample was four, it declined. That asymmetry is the discipline working as designed.

The honest reading of the 2002 ghost completes the picture. The blow-out that anchors every “market will panic” story happened at a CDS regime this sample never revisits, before inflation targeting had credibility, before the central bank was independent by law, before the real-yield anchor existed in its current form. The 2022 analog, the only other Lula transition on record, ran the opposite direction. A four-member class whose two Lula members bracket the entire outcome space cannot license a dramatic forecast in either direction, and the committee declined to manufacture one.

7. Limitations

1. Everything is conditional and may never score. If Lula does not win the 25 October 2026 runoff, no question resolves and the set is void. The committee deliberately did not forecast the condition.

2. The core reference class has four members. Lula 2002, Dilma 2010, Dilma 2014, Lula 2022. The intervals on the thin-class questions span most of the unit interval, evidence weight was capped at 0.25, and every such forecast is stamped exploratory, size zero.

3. The September 2026 unemployment baseline does not exist yet. PNAD unemployment was 165 days stale at freeze, last printed at 6.1 for March 2026, and question 7 resolves against a value that will only be observed after the election. The sign of its resolution can hinge on data not in hand.

4. The long-yield and slope questions rest on uncertified data. The NTN-B series is warehouse-sourced from Tesouro Direto and not certified point-in-time, so the engine refused to count its crossing rate, and the 2029 tenor has only 897 observations from January 2023.

5. The currency class carries proxy basis risk. The official ptax series has 63 observations, too thin to ground, so question 4 is counted on the deep Yahoo USDBRL series instead, with the proxy noted as a disanalogy.

6. The probabilities are one operator’s committee, unblended, and the trial registry records 93 variants behind the final set. The shrinkage that hugs these forecasts to their base rates is a correction for exactly that search.

7. The climatology count in Section 5.3 (65.8 percent of months above the ceiling since 2006) is the paper’s own count of the committed series, offered as context. The engine’s own attempt to count it was refused on an event-type ambiguity, and the operative 70 percent comes from the typed transition class.

8. The repository pre-registration flag is false by design. The forecasts are pre-registered where they live, in the operator ledger, with content-addressed class freezes and kill-criteria (Section 4); this paper does not commit a separate repository-level hypothesis file, so the repository flag stays false, matching the convention set by MKT-013.

8. Conclusion

Conditional on Lula winning the 25 October 2026 runoff, the committee prices the ten-month consequence for Brazilian fiscal risk as modest and largely pre-paid: a coin-flip on the CDS and the long real yield, a marginal lean to a steeper curve, 12 percent on a 10 percent depreciation against a counted 12.4 percent base, 70 percent that inflation is above the ceiling for June 2027, 60 percent that the Selic is higher, 30 percent that unemployment is. The freeze is on the record dated 13 August 2026, the classes are frozen content-addressed, the trial count is 93, and every number will be scored against this document on 30 June 2027 if the condition lands. The set is exploratory and sizes to zero, and its one honest claim is the same one the studio makes everywhere: it was recorded, with its base rates and its refusals beside it, before the answer was known.

9. Data Availability

All series are exports from the command-center indicator warehouse, inventoried in data/manifest.yaml. The committed copies are reduced-resolution for the market series, because the full-resolution feeds belong to their sources and are not ours to redistribute: br-cds-5y (weekly, from 3,288 daily closes, 2017-2026), br-ntnb-2029 / br-ntnb-2035 / br-ntnb-2045 (weekly real yields, Tesouro Direto), usdbrl (monthly, from 5,486 daily closes, Yahoo), and br-selic-meta (monthly policy target). The official statistical series are committed in full: br-ipca-yoy (391 monthly prints, IBGE) and br-pnad-desocupacao (57 quarterly prints, IBGE). Two derived tables complete the set: ntnb-curve-freeze (the three real yields at the freeze) and forecasts (the seven calls, base rates, and class counts from the operator ledger). Headline counts in the text (the 3,288-day percentile and median, the 868-window currency class) were computed on the full warehouse series; the notebook (notebook.ipynb) runs top to bottom on the committed copies with the standard library only, verifies the freeze snapshot, rebuilds the table behind each figure, and recomputes the percentile and climatology checks, reproducing each to within the resolution loss. The forecast set, classes, counts, and refusal ledger are from operator ledger frame fr_6d66bb420fb169e6, junction version 1, cycle 7, freeze 2026-08-13T22:15:58Z.

10. References

See references.yaml. The method and its verdict rule are the studio’s forecasting platform (command-center, 2026); the freeze-before- read discipline and market-null framing continue the line of MKT-013 (ATOL Research, 2026); reference-class forecasting and pre-registration follow Tetlock and Gardner (2015); the Jeffreys interval treatment of counted classes follows standard practice (Brown, Cai, and DasGupta, 2001); the central-bank autonomy and fiscal-framework statutes are Lei Complementar 179/2021 and Lei 14.791/2023.