AFOS-Analytics1/france-2024-electoral-divergence

France 2024 — Electoral Divergence Dataset (Legislative / National Assembly)

Open, auditable cross-reference of prediction market (Polymarket) x opinion polls x press for the 2024 French legislative election (two rounds, 30 June and 7 July 2024), validated against the official result. Part of the AFOS Analytics validated-cases collection. Built to FAIR + tidy-data + AAPOR-disclosure norms with a deterministic, re-runnable poll parser. Available in EN, FR, PT, ES below.

🇬🇧 English

By single party, Marine Le Pen's Rassemblement National (RN) was the largest group in the National Assembly (143 seats), and the deepest market (~US$917k, "which single party wins the most seats") priced this at nearly 99% on the eve, start to finish. It was right. What flipped was the government, not the largest party: the left-wing Nouveau Front Populaire (NFP) held the most seats by coalition (182) thanks to the front républicain (134 NFP and 82 Ensemble candidates withdrew from three-way runoffs to block the RN). The sensational near-RN-majority projections (230 to 270 seats in early polls) lived in polls and thin markets and did not survive in the highest-volume market. The AFOS lesson: deep money did not buy the hype, and a divergence that does not appear at high volume is not robust.

🇫🇷 Français

Par parti isolé, le Rassemblement National (RN) de Marine Le Pen a été le plus grand groupe de l'Assemblée (143 sièges), et le marché le plus profond (~917 k US$, « quel parti isolé obtient le plus de sièges ») l'a valorisé à près de 99 % à la veille, du début à la fin. Il a eu raison. Ce qui a basculé, c'est le gouvernement, pas le plus grand parti : la coalition de gauche Nouveau Front Populaire (NFP) a obtenu le plus de sièges par coalition (182) grâce au front républicain (134 candidats NFP et 82 Ensemble se sont retirés des triangulaires pour bloquer le RN). Les projections d'une quasi-majorité du RN (230 à 270 sièges dans les premiers sondages) ont vécu dans les sondages et les marchés peu liquides et n'ont pas survécu sur le marché au plus fort volume. La leçon AFOS : l'argent profond n'a pas acheté le hype, et une divergence qui n'apparaît pas à fort volume n'est pas robuste.

🇧🇷 Português

Por partido isolado, o Rassemblement National (RN) de Marine Le Pen foi a maior bancada da Assembleia (143 cadeiras), e o mercado mais fundo (~US$ 917 mil, "qual partido isolado tem a maior bancada") precificou isso em quase 99% na véspera, do começo ao fim. Acertou. O que virou foi o governo, não o maior partido: a coligação de esquerda Nova Frente Popular (NFP) fez a maior bancada por coligação (182) graças ao front républicain (134 candidatos da NFP e 82 do Ensemble se retiraram dos triangulares para bloquear o RN). A projeção de uma quase-maioria do RN (230 a 270 cadeiras nas pesquisas iniciais) viveu nas pesquisas e nos mercados rasos e não sobreviveu no mercado de maior volume. A lição do AFOS: o dinheiro fundo não comprou o hype, e uma divergência que não aparece no alto volume não é robusta.

🇪🇸 Español

Por partido individual, la Agrupación Nacional (RN) de Marine Le Pen fue el mayor grupo de la Asamblea (143 escaños), y el mercado más profundo (~917 mil US$, "qué partido individual obtiene más escaños") lo valoró en casi 99% en la víspera, de principio a fin. Acertó. Lo que se dio vuelta fue el gobierno, no el mayor partido: la coalición de izquierda Nuevo Frente Popular (NFP) obtuvo más escaños por coalición (182) gracias al front républicain (134 candidatos de la NFP y 82 de Ensemble se retiraron de los balotajes a tres para bloquear al RN). La proyección de una casi-mayoría del RN (230 a 270 escaños en las encuestas iniciales) vivió en las encuestas y los mercados finos y no sobrevivió en el mercado de mayor volumen. La lección de AFOS: el dinero profundo no compró el hype, y una divergencia que no aparece en alto volumen no es robusta.

Poll series (the audited core)

The opinion-poll leg is the full national series, in tidy long format (one row per poll x bloc), parsed deterministically from the raw Wikipedia aggregation preserved in raw/:

• 52 first-round vote-intention polls (376 rows; 7 pollsters: Elabe, Harris Interactive, Ifop, Ipsos, OpinionWay, Odoxa, Cluster17).
• 44 seat-projection polls (193 rows). French pollsters largely withheld seat projections until the campaign's final days, so this series is dense late and sparse early (source-documented).
• Each poll row carries pollster, fieldwork window, sample size (when disclosed), and a primary source URL. The official 2024 result and the 2022 baseline are included and tagged via row_type, never silently mixed with polls.

Files

README.md                 this file (EN/FR/PT/ES)
DATASHEET.md              Datasheets-for-Datasets (Gebru) 7-section datasheet
CODEBOOK.md               per-variable data dictionary (DDI spirit)
CITATION.cff              how to cite
datapackage.json          Frictionless Data Package + Table Schema
croissant.json            MLCommons Croissant metadata (schema.org/Dataset)
CHECKSUMS.txt             SHA-256 for every data file
LICENSE / LICENSE-APACHE  CC BY 4.0 (data) / Apache 2.0 (code)
banner.png, og-social-card.png
raw/      opinion-polling-2024-french-legislative.wikitext   immutable source of the poll tables
          france-2024-polymarket-snapshot.json               eve-of-election market snapshot
          france-2024-official-result.json                   Ministry of the Interior official result
polls/    france-2024-first-round-vote-long.csv              52 vote-intention polls, long format
          france-2024-seat-projections-long.csv              44 seat-projection polls, long format
market/   france-2024-market-odds-timeseries.csv             daily implied probability per party
derived/  france-2024-divergence.csv                         party x poll-vote x market-prob x divergence
          france-2024-structural-context.csv                 World Bank WGI + WDI indicators
news/     france-2024-press-coverage.csv                     press items (EN/PT/ES/FR notes)

Provenance and reproducibility

• Raw vs. derived are separated. Everything in raw/ is an immutable source pull; everything else is regenerable from it.
• Deterministic parser. The poll tables are produced by a rowspan/colspan-aware grid reconstruction of the raw wikitext (parse-france-polls.mjs, in the AFOS code repository). Re-running it against the same raw/ wikitext reproduces the CSVs byte-for-byte. Validation anchors: the parsed 2024 official row matches the Ministry count (first-round vote NFP 28.21 / ENS 21.28 / RN 29.26; seats NFP 180 / ENS 159 / RN 125 / RN+allies 142 / LR 39).
• Integrity. CHECKSUMS.txt carries a SHA-256 for every data file; the same hashes are embedded in datapackage.json and croissant.json.
• No imputation, no smoothing. Missing source values are left empty; seat ranges are kept as low/high, never averaged.

Sources and method

Polls: Wikipedia aggregation "Opinion polling for the 2024 French legislative election" (CC BY-SA), with the original pollsters credited per row. Market: Polymarket on-chain (live odds via the AFOS server-side proxy). Result: French Ministry of the Interior. Context: World Bank (WGI + WDI). Press: named international outlets. Method note: the market probability (of most single-party seats) is not the same quantity as a vote share or a coalition-seat outcome; divergence_pp is a signal, not a like-for-like error. Please attribute AFOS Analytics and the original pollsters.

How to cite

See CITATION.cff. A DOI (Harvard Dataverse) is being minted; until then, cite this Hugging Face repository URL and the version date.

License: data CC BY 4.0; code Apache 2.0. afos-analytics.com