SuperAGI/llm-vs-prediction-markets

Price-blind LLM forecasts against prediction markets: main dataset

This is the data behind the paper Sorting, Not Calibration: A Price-Blind Comparison of Language Models and Prediction Markets (arXiv link forthcoming; code and paper PDF at https://github.com/TransformerOptimus/llm-vs-prediction-markets). Eight language models were asked for the probability that a prediction-market question resolves Yes, without being shown the market price. Their forecasts are compared with the market price at the same moment.

A prediction market sells a contract that pays $1 if an event happens. Its trading price is the crowd's probability. A market-moment is one market at one forecast time: the unit of every table here.

Full news text is in the separate news dataset (https://huggingface.co/datasets/SuperAGI/llm-vs-prediction-markets-news), linked by SHA-256 checksum. Code that rebuilds the SQLite databases the analysis reads, and the analysis itself, is in the paper's code repository (https://github.com/TransformerOptimus/llm-vs-prediction-markets; analysis/build_db.py).

What is in it

Folder or file	Rows	What it holds
benchmark/window_a/	2,997 market-moments	Polymarket, 6 to 12 February 2026, from the published PolyBench dataset. Own licence, see below.
benchmark/window_b/	3,619 market-moments	Polymarket, 15 July to 9 August 2026.
benchmark/window_c/	3,920 market-moments	Kalshi, 14 May to 10 June 2026. Identifiers and our own labels only, see below.
benchmark/quarantine_flags.parquet	0	Empty table kept for the schema.
runs/calls.parquet	107,214	Every model call, retries included: probability, reply text (not for Window C), token counts, host, settings.
runs/runs.parquet	41	One row per run (model x mode x window) with its full manifest.
analysis_out/		Fixed analysis inputs: depth cuts, calibration gate, memory-probe flag files (Windows A and B), Window C flag counts.
results/		The result tables for all three windows (aggregates per model, window, depth bucket and similar groups). Window C cells built on fewer than 10 markets keep their counts but have scores, returns and intervals blanked (suppressed_small_cell = 1); a few more Window C cells are blanked so the small ones cannot be worked out from totals (suppressed_complement = 1).
schema_benchmark.sql, schema_runs.sql		The exact SQLite schema the rebuild uses.
LICENSE		Full legal text of CC BY 4.0.
MANIFEST.sha256		SHA-256 of every file in this dataset.

One Parquet file per table per window. Column names and meanings follow the research databases exactly; the main ones are described below.

Market tables (Windows A and B)

• market_moments: question, event title, resolution rules (description), scheduled close, forecast time, price (price_yes; on Window A use mid_yes, the order-book midpoint, because PolyBench's stored price is a stale sweep), best bid and ask, the order book as JSON, depth in dollars on the five best prices each side (depth_usd), ladder tag, flags. news_text is replaced by news_text_sha256 and news_text_words.
• moment_outcomes: outcome (outcome_yes), resolution time, topic tags, calibration/test split, fee schedule per row, event id used for resampling.
• price_paths: the market price after the forecast moment (never shown to a model).
• news_articles (Window B): one row per attached article: URL, site, title, crawl capture time, claimed date, the Common Crawl News file it came from (warc_file), text checksum and word count. URL, capture time and checksum identify and verify each article.
• news_pointers (Window A): PolyBench snapshot id, capture time, text checksum and word count per row.

Every news checksum is the SHA-256 of the exact text the model saw.

Window C (Kalshi): what is and is not here

Included: row_id, moment_key, the Kalshi market ticker (venue_market_id), the forecast time, and labels we assigned: calibration/test split, ladder tag, news availability and article count, news capture time, late-dateline flag, in-play flag, end-date-revised flag. News pointers (news_articles) and each model's probability in runs/calls.parquet.

Not included: Kalshi's prices, order books, outcomes, questions, rules, topic categories, close times; model reply text; per-row scores or returns; per-row memory-probe flags (counts only, analysis_out/window_c_quarantine_counts.csv).

Window C results are in the results/ tables alongside A and B, as aggregates; for example results/expA_buckets.csv has every model's accuracy and return per depth bucket. Any Window C cell built on fewer than 10 markets keeps its market count but has its scores, returns and intervals left blank, marked suppressed_small_cell = 1, so no aggregate stands in for a single market. Where a blanked cell could otherwise be worked out by subtraction (for example a band total minus its other parts), a neighbouring cell is blanked too, marked suppressed_complement = 1.

To rebuild Window C, obtain the market data yourself and join it on ticker + forecast time (venue_market_id, forecast_ts):

• Metadata and outcomes: Kalshi's public REST API, https://api.elections.kalshi.com/trade-api/v2 (/markets, and /historical/markets/{ticker} for settled markets).
• Order books: the pmxt archive of hourly Kalshi order-book snapshots, https://r2kalshi.pmxt.dev/kalshi_orderbook_YYYY-MM-DDTHH.parquet (CC BY 4.0, credit "pmxt"). The forecast moment is the first book event at or after 18:00 UTC.
• Price after the forecast: Kalshi's hourly candlesticks (midpoint).

python3 analysis/build_db.py --main . --kalshi-columns prints the columns your Kalshi file must hold; build_db.py --kalshi DIR merges it.

Model calls (runs/calls.parquet)

One row per call. final = 1 marks the record that stands for a row; parsed_probability is the model's probability of Yes (NULL on failed and excluded rows); mode is normal (with news) or memory_probe (news removed). Prompts are not stored, only prompt_sha256; the template is in the code repository (harness/prompt.py). Local file paths in error messages and manifests were shortened; provider response ids were removed.

Personal data

Every email address and phone number in news text, article bodies and model replies was replaced by [email removed] or [phone removed], whoever it belonged to: private people, reporters, press offices and helplines alike. That touched 56 Window A market-moment rows, 21 of the 1,141 Window A texts and 73 of the 5,317 Window B and C texts in the news dataset, and one model reply. The affected rows carry a flag (news_text_redacted on market-moments and news pointers, text_redacted on news articles, redacted in the news dataset, raw_reply_redacted on calls). Their checksums remain those of the original text the model saw, so a redacted text does not re-hash to its stored checksum.

To ask for a correction or removal, open an issue on the code repository: https://github.com/TransformerOptimus/llm-vs-prediction-markets/issues.

How it was made

• Window A: the published PolyBench dataset (Cheng, Liu and Long, arXiv:2604.14199), including the news its pipeline fetched live at each forecast moment.
• Window B: Polymarket order books from the pmxt archive (CC BY 4.0, credit "pmxt"), market metadata from Polymarket's public API, price paths from Polymarket's public price-history endpoint. Each market is taken once, on the first day its 18:00 UTC book is live (two-sided, spread at most 0.10); 120 non-ladder rows per day, 24 from each of five depth groups, plus 20 ladder rows.
• Window C: as Window B, on Kalshi.
• News for Windows B and C: Common Crawl News pages crawled before the forecast moment and at most 14 days before it, matched by name and a reranker score of at least 0.3, top three per row. About seven in ten rows have no news.
• Models: every scored model was publicly released before the outcomes it forecasts. One sample per row, temperature 0 where the host allows it, a 4,000-token reply limit, a price-blind prompt (no price, bid, ask or book).

The primary analysis set drops ladder markets, rows with a revised end date or a late news dateline, and prices outside 0.05 to 0.95. Headline numbers use its test split.

Known limits

• On Windows B and C the question, rules and close date were captured after settlement; exchanges can edit them, so equality with the text at forecast time is assumed, not verified.
• Replies are model output and may contain anything the model wrote, including text quoted from news articles.
• News titles are kept in news_articles; article bodies are only in the news dataset.

Credits

• Market data: Polymarket (Windows A and B) and Kalshi (Window C identifiers).
• Order books for Windows B and C: the pmxt archive (CC BY 4.0, credit "pmxt").
• Window A: PolyBench (Cheng, Liu and Long, arXiv:2604.14199).
• News: Common Crawl News; articles remain the work of their publishers.

Licence

CC BY 4.0 (full text in LICENSE), except benchmark/window_a/, which is CC BY-NC-SA 4.0 because it derives from PolyBench (full text in its LICENSE).

Citation

@misc{krishnan2026sorting,
  title         = {Sorting, Not Calibration: A Price-Blind Comparison of Language Models and Prediction Markets},
  author        = {Krishnan, Adithyan and Kanchi, Ritu Rajesh and NS, Mukunda and Bhola, Ishaan},
  year          = {2026},
  note          = {arXiv preprint, forthcoming}
}