SPECIAL FEATUREFORECAST LABApril 12, 2026|16 min read

Keiko Fujimori, the 49.5% Ceiling, and a Polymarket That Got It Almost Exactly Right

Peru votes today. Keiko Fujimori is running for president for the fourth time — and she has lost the previous three runoffs by an average of about 1.5 points. We ran five forecasting lenses and seven model runs against the Polymarket price of 33%. The headline: no edge. The interesting part is why.

A note on where this comes from
This post is a little different from our usual equity deep dives. It comes from an internal project we’ve been calling the Forecast Lab — a parallel research system that runs our multi-LLM committee methodology against geopolitical, macroeconomic, and other non-equity questions on platforms like Polymarket, Metaculus, and Kalshi. It shares most of the machinery with the equity side of Runchey Research (signal ladders, evidence grading, persona-based discourse, pre-mortems) but swaps the lens library for a set built around reference classes, agent dynamics, and crowd reading rather than filings and moats. The Lab isn’t published yet. This analysis was interesting enough that we wanted to bring it over to the main blog as a one-off feature.

The Numbers That Matter

Polymarket Price
33%

Keiko wins overall

Our Central Estimate
33%

7-run ensemble median

Keiko Runoff Record
0 / 3

2011, 2016, 2021

Avg Losing Margin
~1.5 pts

Across all three losses

Runoff Vote-Share Ceiling
~49.5%

Consistent across 3 opponents

Polymarket Volume
$14.7M

$1.9M in the last 24h

First-Round Candidates
35

Unprecedented fragmentation

Undecided / Null Bloc
40-53%

Per last pre-blackout polls

The Central Question

Peru is voting today in a presidential election that the country’s leading pollster, Datum’s Urpi Torrado, has called “the most unpredictable on record.” The ballot has 35 first-round candidates. Polling has been in blackout since April 7. Somewhere between 40% and 53% of voters were still undecided or planning to spoil their ballots at the last measurement. Eight presidents have rotated through the Casa de Pizarro in the last six years. The homicide rate has more than doubled in a decade.

Against that backdrop, one candidate has been the quiet favorite the whole way: Keiko Fujimori, daughter of the former authoritarian president Alberto Fujimori, who died in September 2024 while serving a sentence for human rights abuses. She leads the first-round polls at roughly 18.6% of the valid vote. She runs the country’s strongest party machine. And, if the prediction market is right, she still only has about a one-in-three chance of winning the presidency.

What the committee examined
Is Polymarket’s 33% on Keiko Fujimori a mispricing — too low given her institutional advantages, or too high given her perfect 0/3 runoff record? We ran five forecasting lenses and a seven-run model ensemble to find out whether there was any edge to be had before tonight’s first-round results.

The Single Most Important Fact: The 49.5% Ceiling

You cannot analyze this market without engaging with the pattern. Keiko Fujimori has reached the runoff in three consecutive Peruvian presidential elections. She has lost all three. The margins, in order: roughly 3 points to Ollanta Humala in 2011, about 0.24 points to Pedro Pablo Kuczynski in 2016, and around 0.25 points to Pedro Castillo in 2021. Average losing margin: about 1.5 points. Average runoff vote share: somewhere near 49.5%.

The striking thing about that ceiling is the identity of the opponents. Humala ran as a left-populist ally of Hugo Chávez. Kuczynski was a center-right technocrat and former Wall Street banker. Castillo was a rural schoolteacher from the Marxist Peru Libre party. Three ideologically incompatible candidates, one consistent outcome: the anti-Fujimori vote coalesced around whichever one of them was still standing on the second ballot.

Our base-rate-anchor lens and our historical-analog-matcher lens arrived at this finding independently. The base-rate lens took Keiko’s personal 0/3 record, credited the razor-thin margins and the structural break of her father’s death, and landed at roughly 28%. The analog lens looked for cross-national parallels (Marine Le Pen’s 0/3 in France is the closest analog) and landed at roughly 35%. Blended 40/60 at the anchor stage, these gave us an outside-view prior of 32%.

Beware of N = 3
Three elections is not a lot of data. The difference between a “structural ceiling” and “three close elections that happened to go the same way” is, statistically, very hard to distinguish from N = 3 observations decided by an average margin smaller than typical polling error. We’ll come back to this. It is the single biggest reason an honest analyst cannot walk away from this market with high confidence.

Five Lenses, Two Directions, One Cancel-Out

After the triage step, we selected five lenses from the Forecast Lab library as most relevant to this market. Each runs as its own multi-model discourse — analysts, minority advocate, bullet hole critics, synthesizer, moderator, reporter. Each lens produces a direction (toward yes, toward no, or neutral), a magnitude, and a standalone probability estimate.

Base Rate Anchor
LEANS NO · ~28%

Keiko’s personal runoff record is 0/3 with a ~49.5% vote-share ceiling. Adjusted for razor-thin margins and Alberto’s death, the ceiling still holds as a reference class. The lens flags its own anchor as WEAK because N = 3.

Historical Analog Matcher
LEANS NO · ~35%

Cross-national anti-dynasty runoff patterns (Le Pen in France being the closest analog) corroborate the consolidation mechanism. Rates the analog quality as STRONG, signal as LEANS NO.

Principal-Agent Map
NEUTRAL · ~33%

Fuerza Popular has the strongest institutional machine in the race and is 92% to win both congressional chambers. But the anti-Keiko bloc has demonstrated equally credible coordination three times running. Institutional alignment is MIXED, commitment credibility is CREDIBLE; the two forces cancel.

Scenario Decomposer
LEANS YES · ~34%

Enumerates the runoff pathways. Conditional win probability ranges from 33% (vs. Belmont, a textbook consolidation vessel) to 48% (vs. López Aliaga, a right-vs-right matchup never seen in the N = 3 sample). The 2026 opponent field is the weakest Keiko has ever drawn. Leans YES weakly.

Crowd Dynamics Reader
LEANS YES · ~34%

$14.7M in Polymarket volume is MODERATELY INFORMATIVE. The salient “0/3 failure narrative” may carry more weight with an English-speaking international bettor base than N = 3 actually supports. Also leans YES weakly — and disagrees, interestingly, with the base-rate lens about which direction the narrative bias runs.

Directional ledger
4
Toward NO
2
Neutral
4
Toward YES

The weighted tally across lens signals comes out exactly balanced. Two backward-looking lenses pull toward NO through the 49.5% ceiling. Two forward-looking lenses pull toward YES through structural breaks (the death of Alberto Fujimori, the weakness of this year’s opponent field, N = 3 statistical noise). The principal-agent map sits in the middle. Net direction: neutral.

The Scenario Decomposition

The scenario decomposer lens is the one that most explicitly breaks the 33% aggregate into its constituent pathways. It enumerates the plausible runoff matchups, assigns a probability to each, and estimates Keiko’s conditional win probability given that matchup. Here is the full matrix, which is really the arithmetic behind the headline number:

Runoff scenario
P(matchup)
P(Keiko wins runoff)
Contribution
vs. Ricardo Belmont
Textbook consolidation vessel
25%
33%
8.3pp
vs. Carlos Álvarez
Comedian, no governance record
25%
36%
9.0pp
vs. Rafael López Aliaga
Right-vs-right — no N = 3 precedent
15%
48%
7.2pp
vs. Other candidate
Residual (fragmented 35-candidate field)
27%
35%
9.5pp
Keiko doesn’t advance
Fragmentation + 2021 Castillo precedent
8%
0%
0.0pp
Probability-weighted total
100%
~34%

Three observations. First, the matrix hands back essentially the market price once you weight across pathways — the scenario lens is not generating the edge on its own. Second, the highest-variance cell is the López Aliaga scenario (48% conditional win on 15% matchup probability), where the right-vs-right dynamic has no historical precedent; a slight re-weighting of that row can swing the total by several points. Third, the 8% non-advance cell is the single scenario the backward-looking lenses barely priced at all, and it is where a first-round surprise tonight would bite hardest.

Where the Lenses Disagree With Each Other

The meta-synthesis stage of our pipeline looks for places where one lens’s conclusion quietly depends on an assumption that another lens has already called into question. Four such tensions surfaced on this market. Two are worth highlighting because they point at the genuine epistemic fault lines here.

Tension 1 · Moderate severity

Is the 0/3 record a ceiling, or is it three close elections?

The base-rate lens treats 0/3 as a reference class. The crowd-dynamics lens argues that the salient failure narrative is overweighted relative to the statistical power of N = 3. Here is the problem: the base rate lens’s adjusted estimate (28%) is below the market price (33%). If 0/3 is a genuine ceiling, the crowd is pricing Keiko too high, not too low — and the crowd-dynamics lens’s own directional bet flips. These two lenses are reaching for opposite corrections of the same piece of data, and whichever one you believe determines which side of the market you want to be on.

Tension 2 · Moderate severity

Does opponent identity matter, or is the ceiling opponent-agnostic?

The scenario decomposer says yes: it models conditional runoff win probabilities ranging from 33% against Belmont to 48% against López Aliaga in a right-vs-right matchup. The analog matcher says no: the ceiling held against three ideologically incompatible opponents, so matchup identity is probably noise. The key observation: the N = 3 sample contains no right-vs-right runoff. If López Aliaga places second tonight, both lenses are extrapolating, and the real uncertainty widens considerably.

The Devil’s Advocate: Is 0/3 Really a Ceiling?

After the lens discourse, our ensemble runs a dedicated devil’s-advocate pass whose entire job is to find the strongest possible case against the consensus. On this market, the devil’s advocate reached 40% — 7 points above the consensus median and above the market price. The argument was the one the crowd-dynamics lens had already half-made:

“Three elections decided by an average of 1.5 percentage points are well within normal polling error and provide no statistical confidence of a structural barrier. If 0/3 is statistical noise rather than signal, Keiko’s true per-runoff win probability is closer to 45-48%, not the 28-35% the ceiling implies. The market at 33% is too low.”

The counter was provided, importantly, by the same historical record the devil’s advocate was attacking. The anti-Keiko consolidation operated against the weakest opponent in the sample (Castillo in 2021, a rural schoolteacher whom almost nobody expected to even make the runoff). If opponent quality isn’t the driver, then the “weak 2026 field” upside argument is softer than it looks.

The devil’s advocate case, in the language of our pipeline, did not survive the stress test. But it survived partially, and the 7-point delta it opened up is a real signal that the consensus is fragile. Our pre-mortem flagged the overall analysis as FRAGILE (score: 0.75) for precisely this reason: most of the load is being carried by one statistical claim about N = 3.

The Seven-Run Ensemble

After the lens work, we run a split-prompt model ensemble. Each model gets the same prediction context (distilled from the lens reports, not raw web research), writes its own base rate, logs each update with an evidence tier, and produces an independent probability. The devil’s advocate is one of the seven runs rather than a separate stage, so its estimate is visible inside the ensemble as a tail-risk signal.

Tier
Model / Role
Estimate
Tier 1
Opus · standard
Starts from 0.28 base rate; credits Alberto death; honors status quo.
32%
Tier 1
Opus · standard
Weights combined outside-view anchor; flags right-vs-right edge case.
34%
Tier 2
Sonnet · standard
Largest weight on death effect; tempered by 2021 Castillo precedent.
35%
Tier 2
Sonnet · standard
Pathway-weighted average of matchup probabilities.
33%
Tier 3
Haiku · standard
Minimal update from anchor; status quo dominates.
32%
Tier 3
Haiku · standard
Slight downward pull from consolidation-as-default framing.
31%
Tier 1
Opus · devil’s advocate
Attacks the N = 3 ceiling claim; reaches for the upside case.
40%
Ensemble Median
33%

Across all 7 runs

Model Agreement
94%

Very high

Std Dev
3.0pp

Tight consensus

Edge vs. Market
~0pp

No actionable gap

The Meta-Finding: The Market Is a Good Analyst

The whole machinery — five lenses, roughly fifty discourse steps, seven model runs, a pre-mortem, a devil’s advocate stress test — lands at a central estimate of ~33%. Polymarket is at 33%.

That is not a boring result. It is actually the most interesting possible result for a market like this one, because it tells you something specific about how the crowd is thinking. A naive reading of the polling data would have Keiko well above 33% (she leads the first-round field by more than 7 points). A naive reading of her runoff record would have her well below 33% (0/3 with an average vote share under 50%). The market is sitting almost exactly between those two naive readings, which is also where our structured analysis lands.

Translating the 33% back into its component parts is instructive. It implies roughly 80% probability that Keiko advances to the runoff (from a fragmented 35-candidate first round), and a conditional win probability in the runoff of around 42%. That 42% is above her historical 0/3 record but below a 50/50 prior. The crowd, in other words, is not naively anchoring to the polling lead, and it is not naively anchoring to the losing streak either. It has already absorbed both facts and traded them into a single number. Our job on this market is to tell you whether that number is wrong. On April 12, 2026, with voting underway and no first-round results yet in hand, our answer is: we don’t think so.

What Will Actually Move This Market

Every estimate in this post has a shelf life measured in hours. The first-round count starts tonight at 5pm Lima time and runs through the early morning. Three things will move the Keiko-wins-overall probability when it lands:

1

Who places second

Belmont (currently polling around 10%) is a textbook consolidation vessel for the anti-Keiko vote. López Aliaga is a hard-right candidate who would produce the first right-vs-right runoff in the sample, and the anti-Keiko mechanism might not apply. The scenario decomposer assigns this variable a swing of about 15 points depending on the matchup.

2

Keiko’s actual first-round share

Polls had her around 18.6%. If she clears 22-23%, the undecided bloc has broken her way and the bandwagon argument becomes live. If she lands at 15% or below, the fragmentation is biting her hardest, and the small but real “Keiko doesn’t advance” scenario (our pre-mortem priced it at 8%) becomes visible.

3

Whether anyone mounts a coordinated runoff campaign

The 49.5% ceiling was not a natural phenomenon. It was produced by coordinated action: media endorsements, party leader pacts, civil society mobilization. Our principal-agent map flagged one specific pre-mortem scenario in which extreme fragmentation breaks that coordination layer, not because the sentiment changes but because the organizers have no incentive to organize. That is the single scenario in which the consensus breaks toward Keiko, and the signals to watch for it are the post-first-round endorsement cascade.

Honest Blind Spots

The pre-mortem stage of the Forecast Lab pipeline exists to make us write down what we didn’t cover. On this market, five things didn’t get the weight they probably deserve, and readers should know about them:

  • No lens modeled the 56-day inter-round campaign dynamics. Peruvian runoffs have shifted 5-10 points on scandals, debate performance, and opposition research in the historical sample.
  • Keiko’s outstanding Odebrecht-linked legal exposure could be weaponized during the runoff and partially reverse any softening from Alberto’s death. We noted this but did not price it.
  • The role of social media and WhatsApp-driven information flows was pivotal in the 2021 Castillo surge. We have no structured way to price it for 2026.
  • Resolution mechanics. The Polymarket deadline is October 31, 2026. Peru has recently contested election results; a disputed count that drags past the deadline would resolve the market to “Other” regardless of who actually won. This is a small but real tail risk that none of our lenses priced explicitly.
  • Five additional lenses in our library (incentive analyzer, information asymmetry detector, temporal structure mapper, resolution mechanics auditor, contrarian signal scanner) were deprioritized in triage but could have changed the weighting. Lens selection is itself a judgment call we got to make, and it’s worth flagging.

What This Exercise Is Actually For

The Forecast Lab is not in the business of beating Polymarket. It is in the business of taking a well-traded market, running a structured methodology against it, and learning something about either the market or the methodology in the gap between them. On the Keiko Fujimori question, the gap is essentially zero. That is itself a finding. It tells us that on a high-volume geopolitical market where the structural evidence is both strong and contested, a sophisticated crowd and a disciplined committee of language models are converging on the same number from opposite directions.

The more interesting question is what the methodology found that the market price alone doesn’t express: the fragility rating, the location of the load-bearing assumption (N = 3), the devil’s-advocate delta, the specific scenario in which the consensus breaks. If you want to trade around tonight’s first round, those are the things to watch, and the 33% consensus is almost a distraction from them.

We’ll revisit this after the first-round results. In the meantime, if you want to see more of these — or if you have questions about how the Forecast Lab methodology differs from the equity-side lens library — reply to the newsletter and let us know. If the response is strong we’ll start publishing these more regularly.

Educational content, not trading advice
Nothing in this post is a recommendation to buy, sell, or hold any prediction market position. Probability estimates are the output of a research methodology, not guidance. Prediction markets are speculative and you should not put money into them that you aren’t prepared to lose. The Forecast Lab is an experimental research project; its outputs are for educational discussion of forecasting methodology.

This report was generated by the Runchey Research AI Ensemble using primary SEC data and reviewed by Matthew Runchey for accuracy.

This analysis is for educational purposes only and does not constitute investment advice. See our Editorial Integrity & Disclosure Policy and Terms of Service.