--- license: - cc-by-4.0 - apache-2.0 language: - en - es - pt pretty_name: "AFOS · USA 2024 Electoral Divergence" tags: - elections - usa - prediction-markets - polls - press - political-risk - divergence - open-data --- ![AFOS · USA 2024 Electoral Divergence](banner.png) # AFOS · USA 2024 Electoral Divergence Dataset [![Harvard Dataverse DOI](https://img.shields.io/badge/Harvard%20Dataverse-10.7910%2FDVN%2F3DJCW5-a51c30)](https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/3DJCW5) 🌐 **[English](#english) · [Español](#español) · [Português](#português)** Open dataset cross-referencing **opinion polls × prediction markets × press coverage** for the **2024 U.S. presidential election** (election day 5 November 2024, Kamala Harris vs Donald Trump), built in the same spirit as the AFOS Brazil 2026 dataset: sources are reported side by side with **explicit divergence**, not blended into a single average, **including where the market was wrong.** Maintained by **[AFOS Analytics](https://afos-analytics.com)**. Part of AFOS's expansion of its electoral-divergence method beyond Brazil. *No personal data, only public electoral information.* 📚 **Academic release:** this dataset also has a citable, versioned deposit on **Harvard Dataverse**, DOI **[10.7910/DVN/3DJCW5](https://doi.org/10.7910/DVN/3DJCW5)**, in the **[AFOS Analytics](https://dataverse.harvard.edu/dataverse/afos-analytics)** collection (alongside the Brazil 2026 dataset). Cite: *Felipe, Andre, 2026, "AFOS - USA 2024 Electoral Divergence Dataset", https://doi.org/10.7910/DVN/3DJCW5, Harvard Dataverse, V1.* --- ## English ![Polymarket implied probability of winning the presidency (electoral college) over the Harris-Trump race](odds-trajectory.png) ![Winner-market probability of winning versus national poll vote share on the eve of the vote](odds-snapshot.png) **What makes 2024 a uniquely honest case: two markets disagreed.** On the eve of the vote (4 Nov 2024), Polymarket's two big election markets pointed in **opposite directions**, and the result vindicated one while refuting the other: | Signal | Eve of vote (4 Nov) | Result | Verdict | |--------|--------------------|--------|---------| | **Winner market** (electoral college, **US$ 3.7 bn**, the largest election market in history) | **Trump 56.4%** × Harris 43.5% | Trump won the presidency (312-226) | ✅ **right** | | **Popular-vote market** (US$ 628 M) | Trump 25.7% × **Harris 74.1%** | Trump won the popular vote | ❌ **wrong** | | **National polls** (7-day avg vote share) | Trump 47.8% × Harris 48.4% (near-tie) | Trump +1.5pp (49.8% × 48.3%) | within margin of error | The winner market's **conviction that Trump would win** ran well ahead of a poll near-tie, and was validated. The popular-vote market did the opposite, favoring Harris, and **missed** (a Republican had not won the popular vote since 2004). The polls, often maligned, landed close. **We report all three, the market's hit and its miss, because the validator is the real result, not a flattering narrative.** **Contents (start with the polls):** | Path | Rows | Content | |------|------|---------| | `polls/usa-national-polls.csv` | 3,691 | National general-election voting intention (**Harris & Trump**), long format (one row per candidate × poll), from **FiveThirtyEight's** 2024 poll database. Full provenance: pollster, rating, methodology, sample, population (LV/RV/A), source URL. | | `data/usa-winner-market-timeseries.csv` | 613 | **PRIMARY AXIS.** Daily Polymarket win-probability (Trump & Harris) from the **"Presidential Election Winner 2024"** market, electoral college, **US$ 3.7 bn** total volume. | | `data/usa-popularvote-market-timeseries.csv` | 615 | **COUNTERPOINT.** Daily win-probability from the **"Popular Vote Winner 2024"** market (US$ 628 M), the market that favored Harris and erred. | | `data/usa-poll-aggregate-timeseries.csv` | 1,322 | 7-day trailing average of national poll vote share (Trump & Harris). | | `data/usa-divergence-timeseries.csv` | 218 | **Market × poll divergence**, winner-market win-probability vs poll vote share, daily, for the Harris-Trump head-to-head period (after Biden withdrew, 21 Jul 2024). | | `press/usa-2024-press-timeline.csv` | 7 | **Curated, Wayback-archived press timeline** cross-referencing the market×poll tension as it unfolded (the AFOS press axis). | | `data/usa-case-summary.json` | n/a | Validated-case summary (result, both markets, polls, honesty note). | | `data/usa-winner-poly-raw.json`, `data/usa-popularvote-poly-raw.json` | n/a | Raw Polymarket payloads (event + per-candidate price histories), kept for provenance. | Market data fetched from Polymarket's gamma-api + clob (US-resolving function). The divergence axis uses the **winner (electoral-college) market**, because that is where the market diverged from the poll near-tie and was validated; the popular-vote market is included in full as an honest counterpoint. `naive_gap_pp` = market *win-probability* − poll *vote share*: different units (probability vs share), reported raw and **not scale-reconciled**. ### ⚖️ The reading, why divergence (and honesty) beats the average - **Winner market × polls, the validated signal.** Through the Harris-Trump race the winner market kept Trump favored to *win* (probability about 49-63%) while national polls measured a near-tie in *vote share* (Trump about 45-48%). The market's conviction led the poll near-tie by roughly **+3 to +16 points** of win-probability, and the result (Trump) validated it. - **Popular-vote market, the honest miss.** The same platform's popular-vote market favored **Harris (74% on the eve)** and was wrong: Trump won the popular vote (49.8% × 48.3%). Two markets, opposite calls, one right and one wrong, a caution that prediction markets are **not monolithic and not infallible.** - **The polls were close.** 2024 national polls clustered around a tie; Trump's +1.5pp popular-vote win sat within normal polling error. This was **not** a "polling failure", a point the post-election coverage makes explicitly (see the press timeline). ### 📰 Press timeline, the third axis (market × poll × **press**) The defining meta-story of 2024 was the **market-versus-polls divergence itself**, covered in real time. Each anchor below was verified live and **preserved in the Internet Archive Wayback Machine** for permanent citation. Primary sources (AP race call, WSJ "Théo" original) are attributed even where linked via an accessible carrier (NPR/PBS/NBC). | Date | Outlet | Maps to | Headline | |------|--------|---------|----------| | 18 Oct 2024 | Fortune | market × poll diverge | "A handful of pro-Trump bettors have dumped more than $25 million into Polymarket, swaying election odds in his favor" | | 21 Oct 2024 | Al Jazeera | poll near-tie | "Early voting in full swing… Harris and Trump tied in polls" | | 24 Oct 2024 | NBC News *(orig. WSJ)* | market conviction | "French trader bet over $28 million on Trump election win using 4 Polymarket accounts" | | 6 Nov 2024 | NPR *(orig. AP)* | result (validator) | "Donald Trump wins the White House, according to AP race call" | | 6 Nov 2024 | PBS NewsHour *(orig. AP)* | result (validator) | "AP Race Call: Donald Trump wins Pennsylvania" | | 8 Nov 2024 | CNN | market vindicated | "How prediction markets saw something the polls and pundits didn't" | | 30 Nov 2024 | NBC News | poll post-mortem (honesty) | "What the 2024 polls got right, and what they got wrong" | Full URLs and Wayback snapshots in [`press/usa-2024-press-timeline.csv`](press/usa-2024-press-timeline.csv) / [`.json`](press/usa-2024-press-timeline.json). ### Provenance, attribution & license **Polls, FiveThirtyEight (CC BY 4.0).** Poll toplines come from FiveThirtyEight's public 2024 president-poll database. FiveThirtyEight was wound down in 2025 and its live endpoint no longer serves the CSV; this dataset uses the **Internet Archive Wayback Machine snapshot of 4 Nov 2024** ([archived CSV](https://web.archive.org/web/20241104120000id_/https://projects.fivethirtyeight.com/polls/data/president_polls.csv)), a frozen, permanently citable copy. The CC BY 4.0 attribution obligation does not lapse with the site: please credit **FiveThirtyEight / ABC News** (compiler) and the **original polling firms** (e.g. NYT/Siena, Marquette, Emerson, Quinnipiac). Nothing is imputed or smoothed; missing values are left blank. **Market odds** come from the public Polymarket markets. **Press** anchors are public articles, archived in the Wayback Machine. **License (dual):** **data** → CC BY 4.0 (`LICENSE-CC-BY-4.0`); **code/scripts** → Apache 2.0 (`LICENSE-APACHE-2.0`), matching the AFOS repo root and Hugging Face mirror. Please attribute **AFOS Analytics**, **FiveThirtyEight / ABC News**, and the **original pollsters**. **Cite:** *AFOS Analytics. USA 2024 Electoral Divergence Dataset. Hugging Face, 2026. CC BY 4.0.* (see `CITATION.cff`) **Disclaimer:** observational research. Not investment advice, not voting guidance. --- ## Español Dataset abierto que cruza **encuestas × mercados de predicción × cobertura de prensa** para la **elección presidencial de EE. UU. 2024** (5 nov 2024, Kamala Harris vs Donald Trump), con **divergencia explícita** entre fuentes, **incluido dónde el mercado se equivocó.** **Lo que hace de 2024 un caso singularmente honesto: dos mercados se contradijeron.** En la víspera (4 nov): - **Mercado de ganador** (colegio electoral, **US$ 3,7 bn**, el mayor mercado electoral de la historia): **Trump 56,4%** × Harris 43,5% → **acertó** (Trump ganó la presidencia, 312-226). - **Mercado de voto popular** (US$ 628 M): Trump 25,7% × **Harris 74,1%** → **se equivocó** (Trump ganó el voto popular; ningún republicano lo ganaba desde 2004). - **Encuestas** (promedio 7 días): Trump 47,8% × Harris 48,4% (empate técnico); resultado Trump +1,5pp (49,8% × 48,3%), dentro del margen de error. El eje de divergencia usa el **mercado de ganador** (donde el mercado divergió del empate de las encuestas y fue validado); el de voto popular se incluye completo como **contrapunto honesto**. `naive_gap_pp` = *probabilidad de ganar* del mercado − *voto* de la encuesta (unidades distintas, sin reconciliar). - `polls/usa-national-polls.csv`, intención de voto nacional (Harris & Trump), formato largo, de la base de FiveThirtyEight (3.691 filas). - `data/usa-winner-market-timeseries.csv` / `usa-popularvote-market-timeseries.csv`, probabilidad diaria de Polymarket (colegio = eje; voto popular = contrapunto). - `data/usa-divergence-timeseries.csv`, divergencia mercado × encuesta (probabilidad de ganar vs voto), periodo Harris-Trump. - `press/usa-2024-press-timeline.csv`, **línea de tiempo de prensa curada y archivada en Wayback** (mercado × encuesta × prensa). **La lectura:** el mercado de ganador acertó frente al empate de las encuestas (señal validada); el de voto popular falló (contrapunto honesto); las encuestas quedaron cerca. **El validador es el resultado real, no un relato favorable.** **Fuentes/licencia:** encuestas de FiveThirtyEight / ABC News (CC BY 4.0, vía snapshot de Wayback del 4 nov 2024), atribuir a FiveThirtyEight y a las encuestadoras originales; odds de Polymarket; prensa archivada en Wayback. Licencia dual: datos CC BY 4.0, scripts Apache 2.0. Atribuir a **AFOS Analytics**. Investigación observacional; no es asesoría de inversión ni orientación de voto. --- ## Português Dataset aberto cruzando **pesquisas × mercados de previsão × cobertura de imprensa** para a **eleição presidencial dos EUA 2024** (5/nov/2024, Kamala Harris × Donald Trump), com **divergência explícita** entre fontes, **inclusive onde o mercado errou.** **O que torna 2024 um caso singularmente honesto: dois mercados se contradisseram.** Na véspera (4/nov): - **Mercado de vencedor** (colégio eleitoral, **US$ 3,7 bi**, o maior mercado eleitoral da história): **Trump 56,4%** × Harris 43,5% → **acertou** (Trump venceu a presidência, 312-226). - **Mercado de voto popular** (US$ 628 M): Trump 25,7% × **Harris 74,1%** → **errou** (Trump venceu o voto popular; republicano não ganhava o voto popular desde 2004). - **Pesquisas** (média de 7 dias): Trump 47,8% × Harris 48,4% (empate técnico); resultado Trump +1,5pp (49,8% × 48,3%), dentro da margem de erro. O eixo de divergência usa o **mercado de vencedor** (onde o mercado divergiu do empate das pesquisas e foi validado); o de voto popular entra completo como **contraponto honesto**. `naive_gap_pp` = *probabilidade de vencer* do mercado − *voto* da pesquisa (unidades diferentes, não reconciliado). - `polls/usa-national-polls.csv`, intenção de voto nacional (Harris & Trump), formato longo, da base do FiveThirtyEight (3.691 linhas). - `data/usa-winner-market-timeseries.csv` / `usa-popularvote-market-timeseries.csv`, probabilidade diária do Polymarket (colégio = eixo; voto popular = contraponto). - `data/usa-divergence-timeseries.csv`, divergência mercado × pesquisa (probabilidade de vencer vs voto), período Harris-Trump. - `press/usa-2024-press-timeline.csv`, **linha do tempo de imprensa curada e arquivada no Wayback** (mercado × pesquisa × imprensa). ### ⚖️ A leitura, por que a divergência (e a honestidade) supera a média - **Mercado de vencedor × pesquisas, o sinal validado.** O mercado manteve Trump favorito a *vencer* (cerca de 49-63%) enquanto as pesquisas mediam empate no *voto* (cerca de 45-48%). A convicção do mercado ia +3 a +16 pontos à frente do empate, e o resultado (Trump) validou. - **Mercado de voto popular, o erro honesto.** O mesmo Polymarket dava **Harris (74% na véspera)** e errou. Dois mercados, chamadas opostas, um certo e um errado, alerta de que mercados de previsão **não são monolíticos nem infalíveis.** - **As pesquisas ficaram perto.** Trump +1,5pp no voto popular ficou dentro do erro amostral; não foi "pesquisa furada" (a imprensa pós-eleição diz isso explicitamente). **Fontes/licença:** pesquisas do FiveThirtyEight / ABC News (CC BY 4.0, via snapshot do Wayback de 4/nov/2024), atribuir ao FiveThirtyEight e aos institutos originais; odds do Polymarket; imprensa arquivada no Wayback. Licença dual: dados CC BY 4.0, scripts Apache 2.0. Atribuir a **AFOS Analytics**. Pesquisa observacional; não é recomendação de investimento nem orientação de voto. --- **Sources / Fuentes / Fontes:** [FiveThirtyEight 2024 president polls](https://web.archive.org/web/20241104120000id_/https://projects.fivethirtyeight.com/polls/data/president_polls.csv) (CC BY 4.0, via Wayback) · Polymarket (winner + popular-vote markets) · curated press (AP/NPR/PBS, Fortune, CNN, NBC, Al Jazeera, Wayback-archived). Column definitions in [`DATA_DICTIONARY.md`](DATA_DICTIONARY.md). ## Structural context (World Bank) Beyond the divergence data, this dataset ships `data/usa-structural-context.csv`: official, open World Bank indicators that frame the country, **governance** (Worldwide Governance Indicators, 0-100 scale) plus **economy & education** (World Development Indicators: population, GDP, GDP per capita, inflation, public education spending, expected years of schooling). These are annual structural indicators that contextualize the country; they do **not** predict the electoral outcome. Columns are documented in `DATA_DICTIONARY.md`. ## Provenance and reproducibility (v1.2.0) - **raw/ vs. derived.** `raw/` holds immutable source pulls: the Polymarket winner-market and popular-vote-market snapshots (`*-poly-raw.json`) and the official result (`usa-2024-official-result.json`). The CSVs under `data/` are derived series regenerated from those pulls plus the FiveThirtyEight poll file by the shipped `build-usa-divergence.mjs`. - **Poll disclosure (AAPOR-style).** `polls/usa-national-polls.csv` carries per-row `pollster`, `end_date`, `sample_size`, `population`, `methodology`, a FiveThirtyEight `numeric_grade` (pollster quality), and a `source_url` for every poll. - **Validated against the official result.** Trump won both the Electoral College (312-226) and the popular vote (49.8% to 48.3%); the deep college-winner market favored Trump throughout. See `raw/usa-2024-official-result.json`. - **Integrity.** `datapackage.json` (Frictionless, validated), `croissant.json` (MLCommons), and `CHECKSUMS.txt` (SHA-256) describe and verify every file. --- ## 📬 The method, and where it runs live **English** · This is one of the elections AFOS has validated against the official result. How the validation works and what qualifies as a case: [the method](https://www.afos-analytics.com/en/how-it-works) and [methodology](https://www.afos-analytics.com/en/methodology). The same method runs live on Brazil 2026, published every day in [AFOS Daily](https://www.afos-analytics.com/en/daily), free by email. **Português** · Esta é uma das eleições que o AFOS validou contra o resultado oficial. Como a validação funciona e o que qualifica um caso: [o método](https://www.afos-analytics.com/pt-BR/how-it-works) e a [metodologia](https://www.afos-analytics.com/pt-BR/methodology). O mesmo método roda ao vivo no Brasil 2026, publicado todos os dias no [AFOS Daily](https://www.afos-analytics.com/pt-BR/daily), gratuito por e-mail. **Español** · Esta es una de las elecciones que AFOS validó contra el resultado oficial. Cómo funciona la validación y qué califica como caso: [el método](https://www.afos-analytics.com/es/how-it-works) y la [metodología](https://www.afos-analytics.com/es/methodology). El mismo método corre en vivo en Brasil 2026, publicado todos los días en [AFOS Daily](https://www.afos-analytics.com/es/daily), gratis por correo.