White Paper
Markets, polls, and the divergence signal: project goals and method
Working note, June 2026
Two decades of research on prediction markets and polls (Wolfers and Zitzewitz, 2004; Rothschild, POQ 2009) established that each carries information and each carries bias, and that de-biased markets often outperform de-biased polls early in the cycle. The practical question that follows has been harder to operationalize at scale:
When markets and polls disagree, what does the disagreement tell us, and does it hold up against real outcomes?
AFOS Analytics is an attempt to answer this continuously, across countries, in public, with the data left open for inspection.
Our working thesis is simple and deliberately falsifiable:
The divergence between what real-money markets price and what polls report is itself a signal, and its informativeness can be checked against the eventual result.
We do not claim divergence always favors the market. We claim it is a measurable quantity worth tracking, and that whether it pointed the right way is an empirical question we answer case by case, including the cases where it failed.
What we integrate
For each race we bring four layers into one continuously updated panel:
- Real-money prediction markets (e.g. Polymarket): implied probabilities.
- Traditional opinion polls from multiple institutes, with sampling and methodological metadata preserved where the source publishes it.
- Mainstream press coverage time-stamped and archived (Wayback) so each market move can be anchored to a real-world event.
- AI-generated synthesis that summarizes, never substitutes, the underlying numbers.
The divergence reading
The divergence reading is descriptive, not yet a calibrated estimator: for each candidate we track the gap between market-implied probability and poll-implied standing over the cycle, and we ask whether and when that gap anticipated the real result.
We are explicit that this is observational. Turning it into a properly calibrated forecast is one of the project’s open questions.
Validation, including the failures
The project began with practical validation rather than a model. We reconstructed the markets-versus-polls picture for elections whose results are now known and checked the divergence reading against the actual outcome. The validated set spans eight national elections across three continents (Brazil context aside): Peru, Colombia, Chile, Germany, Canada, Mexico 2024, the United Kingdom 2024, and the United States 2024.
Two points we treat as load-bearing:
- Convergence counts too. In Germany and Colombia the signal was near-zero divergence, and the result confirmed it. The validator is the real outcome, not the size of the divergence.
- We publish the misses. In the US 2024 case the electoral-college market read the result correctly while the popular-vote market did not. We document the failure rather than hide it, because a method that only displays its wins is not a method.
Data and openness
Everything is open and citable.
- Datasets are deposited in the Harvard Dataverse (collection afos-analytics) with their own DOIs, under CC BY 4.0, and mirrored on Hugging Face.
- The platform itself is open source.
- The aim is that any claim we make can be audited and reproduced from the deposited data.
Goals
We state the ambition plainly while keeping the near-term claims modest. The validated cases are the evidence; the global framing is the direction, not a finished result.
- Near term: harden the divergence reading from a descriptive indicator into something with explicit calibration and uncertainty, and widen the validated-case library.
- Medium term: a real-time, global reading of political risk that treats every election as a comparable instance of the same markets-versus-polls measurement problem, rather than a series of one-off national stories.
Open questions
A few questions we are genuinely unsure about and are working on:
- How best to de-bias thin or low-liquidity markets (Polymarket depth varies enormously by race) before comparing them to polls.
- Whether an MRP-style correction on the polling side changes which signal "wins" in the divergence.
- How to formalize "divergence as signal" into a calibrated probability rather than a descriptive gap.
- Selection effects: the races that attract market liquidity are not a random sample of elections.
Limitations
The validated set is still small. Market liquidity is uneven. Press anchoring is curated, not exhaustive. The divergence reading is, today, a descriptive instrument and we are careful not to oversell it. We would rather be corrected early than be impressive on paper.
Data: Harvard Dataverse (collection afos-analytics, CC BY 4.0). Open-source platform.
Contact: founder@afos-analytics.com