Ready to save as PDF

Research report · Colombia 2026

Democracy at the Bifurcation Point

In the new politics, elections are not decided by majorities — they are decided by margins. Colombia 2026 is a case study in how modern democracies reach inflection points where a few percentage points, a single running mate, or a two-week lag in public attention can determine the trajectory of a country for a generation.

Model updated 14 March 2026  ·  Data through 13 March 2026  ·  First round: 31 May 2026  ·  Second round (if required): 21 June 2026

1st round leader

~34%

Cepeda · range 29–39%

Valencia post-consulta

~26%

Up from 4–10% pre-Mar 8

Runoff probability

~97%

No candidate near 50%

Polymarket volume

$8.5M

Cepeda 42% · Valencia 35%

---

The new competitive politics

Something has changed in how democracies work. The era of comfortable majorities is over. From Brazil in 2022 — decided by 1.8 percentage points, the closest in that country's democratic history — to Colombia's own 2022 first round, where three candidates arrived within 17 points of each other, elections in Latin America and beyond are converging on a structural condition of near-parity. The center of gravity of the electorate is fragmenting, and the old coalitions that once produced commanding leads are dissolving faster than new ones can form.

Colombia 2026 takes this condition to its logical extreme. A left-wing senator consolidates the Petro coalition. A viral outsider-lawyer holds the right. A centrist technocrat who won a surprise primary with a data-director running mate emerges as the dark horse. And a respected academic from Medellín, running for president for the third consecutive time, holds the ideological center. No one is close to 50%. The runoff probability sits at 97%. Whoever wins will do so with something in the low forties — a plurality mandate in a polarized country.

This is not a Colombian anomaly. It is the new normal of democratic competition, and it creates a specific analytical challenge: in a world where elections are decided by 3 points rather than 30, the margin of error of our measurement tools starts to matter as much as the measurement itself.

Data signals calibrated · March 8 ranking

Different data sources carry different predictive weights. Calibration is updated empirically after each electoral event — the March 8 consultas provided new ground truth.

The information lag: how fast does reality travel?

In a bifurcated race decided by 3pp, the 14-day gap between when the electorate moves and when we observe it is not a technical detail — it is the window where elections are won and lost.

"In a world where elections are decided by 3 points rather than 30, the information lag between when the electorate moves and when we observe it becomes the decisive variable."

The new data environment offers a partial answer. Prediction markets, search behavior, and social media engagement are not perfect substitutes for a well-designed survey — but they update in hours rather than weeks, and they aggregate the expectations of participants who have real stakes in being right. When traditional measurement and digital signals diverge, that divergence is itself a signal: it marks the exact moment where the official narrative has not yet caught up to where the electorate already is.

In Colombia 2026, the divergence was spectacular. Before March 8, every traditional measurement had Paloma Valencia polling between 4% and 10% for the first round. Meanwhile Polymarket — the largest political prediction market in the world with $8.5M in volume — had already begun pricing her as a serious contender. The market was right. The divergence was the signal.

---

A new signal stack for a new politics

The model we built is not a critique of traditional measurement — it is an attempt to assemble every available signal into a coherent picture of an electorate in motion. Three layers, each capturing a different dimension of political reality.

A · Bayesian-Decay

The real-time layer. Blends prediction market odds (55%) with calibrated Google Trends (45%) into a digital signal, then anchors it against field surveys weighted by their information freshness — not just their sample size.

Historical 38% · Projection 30%

B · Kalman Filter

The uncertainty layer. Every data point — a Polymarket update, a GT reading, a survey — is treated as a noisy observation of a hidden truth. The filter continuously updates its best estimate and its confidence, narrowing as election day approaches.

Historical 37% · Projection 22%

C · Gravity / Momentum

The trajectory layer. From the last known market position, fits velocity and acceleration to each candidate's trend. Valencia: +2.97pp/day accelerating. Espriella: −2.1pp/day falling. Momentum in politics, like physics, tends to continue until something stops it.

Historical 25% · Projection 48%

The information lag as a political variable

In competitive politics, timing is not just operationally important — it is analytically decisive. The model measures empirically how long it takes for a real movement in public attention to propagate through the information chain: from the moment voters change their minds, through their search behavior, into a pollster's field, through data processing, to publication. That total lag is approximately 14 days. In a race where the margin is 3 points, 14 days is an eternity. It is the window where the future is already happening and the official record has not yet noticed.

---

What search behavior reveals about political identity

Google Trends is not a poll. It does not ask anyone how they intend to vote. What it captures is attention — and in modern politics, sustained attention is one of the most reliable precursors to actual votes. But not all attention is the same. The model's most important empirical finding is that Google Trends systematically underrepresents certain political identities, and the pattern of underrepresentation is consistent and measurable across elections.

This is not a technical artifact. It reflects something real about who uses Google to express political curiosity, and who does not. Urban, younger, more educated voters are overrepresented in search behavior. Older, rural, lower-income voters — who in Colombia as elsewhere have proven to be decisive swing constituencies — leave a much smaller digital footprint. In 2022, that asymmetry allowed Rodolfo Hernández to approach election day nearly invisible in the data, and then receive 28% of the vote. The model learned from that experience.

Empirical correction factors · CO 2018 + 2022

CF > 1 means GT underestimated the candidate. Hernández CF ≈ 1.69 — Google dramatically missed him in 2022. Valencia carries the Hernández proxy for 2026.

Established left · CF ≈ 1.0 (Petro)

Search behavior accurately predicted vote share. An engaged, digitally active base that expresses political identity online as readily as at the ballot box. No correction needed — digital and physical align.

Outsider / dark horse · CF ≈ 1.69 (Hernández)

Search showed 18%, vote delivered 28.2%. A constituency invisible to digital measurement but decisive at the ballot box. This is the correction factor now applied to Valencia — and it is why the model was ahead of the published data on her rise.

Cross-correlation: GT leads polls by ~14 days (CO 2022)

The lag at peak correlation tells us how far GT leads polls for each candidate type. Hernández peak at lag 2 weeks (r = 0.87). Petro peaks at lag 0 — no lead. Valencia 2026 shows a similar delayed peak to Hernández, consistent with the dark-horse proxy assignment.

---

The bifurcation point

First round · 31 May 2026

The first round will not produce a winner. It will produce a choice — the two candidates who advance to decide Colombia's direction for the next four years in a head-to-head on June 21. The model projects Cepeda at approximately 34%, Valencia at 26%, Espriella at 21%, and Fajardo at 7.5%. The uncertainty bands are wide because at this distance, and in a race this competitive, small events produce large effects.

First round projection with uncertainty bands

Solid bar = central projection. Shaded band = ±1σ model uncertainty. Vertical line = 50% threshold (required for first-round victory).

"On March 8, a running mate announcement and a primary result moved the projected second-round margin by 30 points overnight. That is what competitive democracy looks like at the bifurcation point."

The single-event inversion

The most remarkable feature of this election is what happened on March 8 and in the days immediately following. Every piece of traditional measurement taken before that date told a consistent story: Cepeda would win a second round against Valencia by 10 to 23 percentage points. Then Juan Daniel Oviedo — who had polled at 11% in the primary, and received 21% of the actual vote — accepted Valencia's invitation to join her ticket. One announcement, confirmed by one primary result, moved the projected second-round margin by approximately 30 points.

This is what bifurcation looks like in real time. The political system was in one equilibrium — left-vs-fragmented-right, Cepeda-dominant — and moved suddenly and discontinuously to a different one. The trigger was not a scandal or an economic shock. It was a personnel decision and a vote total that no published survey had anticipated correctly.

R2 poll evolution: Cepeda vs Valencia across all firms

Faded points = outlier firms (Invamer, CNC) — both ranked worst by La Silla Vacía after March 8. Starred point = post-consulta AtlasIntel, carrying weight 0.65 in the blended R2 estimate. Note the dramatic inversion at the final data point.

The second round: who decides?

In a low-margin democracy, the decisive voters are not the faithful of either camp — they are the people in the middle who arrive at the second round with no natural home. In Colombia 2026 that means approximately 7.5% of first-round votes going to Sergio Fajardo, and another 3.5% to Claudia López. Their choices — or their decision to stay home — will determine the presidency.

Fajardo voters, the model estimates, split roughly 25% to Cepeda, 45% to Valencia, and 30% to blank or abstention. This is the decisive arithmetic: a group that leans center-right but distrusts both poles, whose collective decision in the voting booth on June 21 is worth more than any single campaign event between now and then.

Cepeda 40.5%

Valencia 43.5%

blank 16%

Projected R2 result after vote transfers. Valencia wins by ~3pp. High uncertainty — within normal forecast spread. Confidence: 55%.

---

The 14-day signal window

The most important number in the model is the empirical GT-to-poll transmission lag: approximately 14 days. For the next 11 weeks, the model watches GT every week for a spike that polls won't confirm for a fortnight. If Valencia continues to accelerate in search behavior, that signal will arrive before any survey does.

Three vote blocs are decisive for the second round:

The centrist swing (~11%)

Fajardo and López voters combined represent the ideological middle of a polarized country. In the second round they face a binary choice neither was designed for. How they resolve it — and whether they resolve it at all, or simply stay home — is the single variable with the most leverage over the outcome.

The right-wing consolidation (~21%)

Espriella's first-round voters are the largest transferable bloc. The model estimates 75% go to Valencia — but this assumes active consolidation. In a competitive democracy, nothing consolidates automatically. Every point of that transfer requires a political decision.

Limits of this model: Correction factors derived from exactly n=2 Colombian presidential elections. Gravity model assumes current momentum holds for 11 weeks — rarely true at peak rates. Post-consulta data: one poll over three days. Uncertainty bands reflect model spread only, not unknown events. A debate, health crisis, scandal, or security incident can move things in ways no model anticipates. Colombia's electoral history contains several of these.

---

Methodology note

Ensemble of Bayesian-Decay (Sub-model A), Kalman Filter (B), and Gravity-Taylor momentum extrapolation (C). Weights shift from approximately equal in the historical phase to C-dominant (48%) in the projection phase. Digital signal = Polymarket 55% + Google Trends LOO-calibrated 45%, with dynamic GT weight boost to 65% on divergence. Survey information freshness computed as field lag + publication lag + response lag prior by methodology. Empirical GT-to-poll transmission lag = 14 days, derived via cross-correlation against CO 2018 and CO 2022 election series. Correction factors via leave-one-out calibration on both elections. 2026 candidate proxy mapping: Cepeda→Petro archetype, Espriella→Hernández, Valencia→50% Fajardo + 50% Hernández, Fajardo→direct historical. Signal weights post-March 8 updated from La Silla Vacía post-consulta accuracy audit. 13 surveys spanning January 5 – March 13, 2026. Polymarket data: $8.5M overall winner market, $578K first-round market, March 13 2026.