# Datasheet: India 2024 (Lok Sabha) Electoral Cross-Reference Dataset
Following Datasheets for Datasets (Gebru et al., 2021; arXiv:1803.09010). Part of the AFOS Analytics validated-cases collection, cross-referencing prediction market × opinion polls × exit polls × press against the official result.
1. Motivation
Created by AFOS Analytics (independent, self-funded) to document, in an auditable form, how a prediction market, opinion polls, exit polls and press coverage compared with the official result of the 2024 Indian general election (the 18th Lok Sabha), the largest election in human history. This case: the market and the polls agreed with each other and overestimated the winner together, getting the winner right (Modi's NDA) but the seat magnitude very wrong (predicted 350-400, result ~293; the BJP alone lost its single-party majority). AFOS reports the crossing without any value judgment; the official result is the anchor.
2. Composition
Observational units, each in its own table: a poll × alliance seat projection (polls/, 17 opinion + 14 exit polls + the ECI anchor); a poll × alliance vote-share estimate (polls/, 13 polls + the ECI result + the 2019 result); a daily market-implied probability per contract (market/, 7 contracts); a per-tool comparison against the result (derived/india-2024-divergence.csv); the three largest states' 2019-vs-2024 swings and projections (derived/india-2024-state-swings.csv); press items (news/, 10 outlets with trilingual notes); and national structural context (structural/, World Bank governance + macro). Raw pulls in raw/ (official result + market snapshot). Instance counts and columns are enumerated in datapackage.json and CODEBOOK.md. Values are transcribed from public aggregates; nothing is imputed. No personal or sensitive data.
3. Collection process
Polls and exit polls were extracted deterministically from the raw wikitext of the Wikipedia opinion-polling aggregation (to avoid the truncation of summarised reads), each pollster credited, with a second source in the press. Market series from Polymarket (CLOB prices-history) via the AFOS server-side proxy, deterministic and on-chain-auditable. Press items are named outlets with primary links. Official result from the Election Commission of India, corroborated across four sources. Voting phases: 19 Apr, 26 Apr, 7 May, 13 May, 20 May, 25 May, 1 Jun 2024; result proclaimed 4 Jun 2024.
4. Preprocessing / cleaning
Poll and market series are emitted in tidy long format. Seat projections are kept as ranges where the source gave ranges. Market "expected seats" figures cited in the documentation are probability-weighted midpoints computed by AFOS (disclosed as such), not a single published number. These computed figures are reproducible from the shipped CSVs via scripts/compute_derived_figures.py. Missing source values are left empty; nothing is smoothed.
5. Uses
Supports the AFOS cross-reference narrative. Should NOT be read as a causal claim, nor as an official tally; a market-implied probability is not a vote share, and a seat projection is not a vote share. Excluded on purpose: informal betting (Delhi bookies, Phalodi satta bazaar) and community forecasts (Metaculus), which also over-projected the NDA but are not auditable and have no verifiable volume. Known limitation: state-wise polls exist for all ~28 states; only the three largest were folded in.
6. Distribution
Published on the Hugging Face Hub (https://huggingface.co/datasets/AFOS-Analytics1/india-2024-lok-sabha-electoral-divergence), version 1.0.0, and deposited on the Harvard Dataverse under the persistent DOI 10.7910/DVN/CNLXZJ, in the AFOS Analytics collection. License: data CC BY 4.0, code Apache 2.0. Upstream provenance and terms: the poll numbers are facts released by the named pollsters, and they were read from the Wikipedia polling aggregation, which is published under CC BY-SA. AFOS releases its own contribution, meaning the parsing, the joining and the divergence tables, under CC BY 4.0. Anyone reusing the poll tables should credit the original pollsters and observe the terms of the Wikipedia aggregation the tables were read from. No redistribution-restricted source is included.
Correction, 2026-07-25. This section previously said that a persistent DOI was "planned (Zenodo)". By the time the bundle was published the deposit had in fact been made, and on the Harvard Dataverse rather than Zenodo. The live DOI is stated above.
7. Maintenance
Maintained by AFOS Analytics (afos-analytics.com). Corrections and versioning via the repository. Country ISO3: IND. Election: national (Lok Sabha), 2024.
Reproduction. scripts/PIPELINE.md documents, file by file, how every table in this bundle was produced: the Wikipedia ?action=raw wikitext reads for the polls, exit polls and state results, the Polymarket endpoints for the market series, and the official ECI figures used as the anchor. scripts/compute_derived_figures.py recomputes the derived figures. Known limitation: unlike some other bundles in this collection, the raw wikitext snapshot itself is not shipped here, so a rebuild reads the live page, which may have changed since the extraction.