AFOS-Analytics1/india-2024-lok-sabha-electoral-divergence

[image]


DOI (Harvard Dataverse): https://doi.org/10.7910/DVN/CNLXZJ

India 2024 (Lok Sabha): Electoral Cross-Reference Dataset

An AFOS Analytics validated case. It cross-references a prediction market (Polymarket), opinion polls, exit polls and press coverage against the official result of the 2024 Indian general election, the largest election in human history. AFOS reports the crossing, convergence or divergence, and makes no value judgment; the official result of the Election Commission of India is the anchor.

The finding

EN. The prediction market and the polls agreed with each other and overestimated together: they got the winner right (Narendra Modi's NDA alliance would win) but the size very wrong. They predicted a landslide of 350 to 400 of the 543 seats; the alliance won only ~293, and Modi's own party, the BJP, lost the single-party majority it had governed with (240 seats), for the first time since 2014. What validates the case is the official result, not the tools getting it right.

PT. O mercado de previsão e as pesquisas concordaram entre si e superestimaram juntos: acertaram o vencedor (a aliança NDA, de Modi, venceria) mas erraram muito o tamanho. Previram um massacre de 350 a 400 das 543 cadeiras; a aliança fez só ~293, e o partido do próprio Modi, o BJP, perdeu a maioria que governava sozinho (240 cadeiras), pela primeira vez desde 2014. O que valida o caso é o resultado oficial, não as ferramentas terem acertado.

ES. El mercado de predicción y las encuestas coincidieron entre sí y sobreestimaron juntos: acertaron el ganador (la alianza NDA, de Modi, ganaría) pero erraron mucho el tamaño. Predijeron un arrasamiento de 350 a 400 de los 543 escaños; la alianza obtuvo solo ~293, y el partido del propio Modi, el BJP, perdió la mayoría propia con que gobernaba (240 escaños), por primera vez desde 2014. Lo que valida el caso es el resultado oficial, no que las herramientas acertaran.

The crossing, in two charts

[image]

The market was 80-96% sure the BJP would keep its majority (it did not) and priced the actual outcome (NDA under 300 seats) at only 3-15%.

[image]

Every tool clustered at 355-373 seats for the NDA; the official result was 293. They converged with each other and were wrong together.

Scale (why this election matters)

• —~980 million eligible voters (final electoral roll ~97.97 crore): the largest election in human history.
• —642 million votes cast (turnout 65.79%), a world record for votes cast.
• —Voting via Electronic Voting Machines (EVM) with a VVPAT paper trail (~5.5 million machines).
• —7 phases over six weeks (19 Apr to 1 Jun 2024), counted in a single day (4 Jun 2024), 543 seats, majority at 272.

Cross-referenced evidence base

• —Market: Polymarket, an on-chain real-money market with auditable volume (~US$835k across 3 markets; a daily price series of ~350 observations, Apr to Jun 2024).
• —Polls: 44 national poll data points (17 seat-projection + 13 vote-share + 14 exit polls), plus the three largest states (Uttar Pradesh, Maharashtra, West Bengal); ~100 polls read in total.
• —Press: 10 articles with verified links, spanning the campaign, exit polls, result and the stock-market reaction.
• —Anchor: the officially proclaimed result of the Election Commission of India.

What's inside

Methodology and scope

AFOS cross-references prediction markets and polls against the officially proclaimed result and reports the convergence or divergence, without any value judgment. This is a national case (the Lok Sabha, India's federal parliament); state elections are separate.

Notes

• —Electronic voting machines. India votes on EVMs (nationwide since 2004) with a VVPAT paper trail, a sample of which is cross-checked. EVM integrity is a live political debate in India: parts of the opposition question it; the Election Commission and the Supreme Court defend it (the Supreme Court rejected pleas for 100% VVPAT verification in April 2024). The 2024 national result was broadly accepted, with isolated seat-level complaints. AFOS takes no position on EVM integrity and anchors on the official ECI result.
• —Result variance. The NDA total is reported as 292-294 and the INDIA bloc as 232-234 by different sources (classification of small allies); this dataset adopts NDA 293, INDIA 234. Vote share is the ECI figure (NDA 43.8% / INDIA 41.48%, a 2.32-point lead); an alternative tabulation gives 42.5% / 40.6%. Both are a near-tie.
• —Exit polls were not unanimous. The 14 houses ranged from 221 (DB Live, which projected an NDA defeat) to 415 (Today's Chanakya); the closest to the real ~293 was Dainik Bhaskar (316).
• —Auditability. Only Polymarket is included as a market signal, because it is on-chain and auditable. Informal Indian betting (Delhi bookies, Phalodi satta bazaar) and community forecasts (Metaculus) also over-projected the NDA, but they are not auditable and have no verifiable volume, so they are deliberately excluded from this dataset.
• —State coverage. The three largest states (170 of 543 seats) are incorporated; smaller states are documented in the source but not folded in one by one.
• —Reproducibility. The AFOS-computed figures (exit-poll average, market expected seats, vote-share averages) are recomputed from the shipped CSVs by scripts/compute_derived_figures.py; see scripts/PIPELINE.md for the full pipeline and sources.

Full field definitions are in CODEBOOK.md; provenance and limitations in DATASHEET.md and KNOWN_ISSUES.md.

Citation

See CITATION.cff. Please credit the original pollsters, Polymarket, and the Election Commission of India.

License

Data: CC BY 4.0 (LICENSE). Accompanying code: Apache 2.0 (LICENSE-APACHE). No redistribution-restricted source is included.

AFOS Analytics: afos-analytics.com. Cross-referencing prediction markets, polls and press against the official result of elections.