thePredictables/kalshi-live-news

Kalshi live questions with news (October 2026)

Private, research use only. Every row carries the full text of the news articles retrieved for it; those texts are copyrighted by their publishers, and part of the retrieval used Bing News RSS, whose terms allow private, non-commercial use. Do not make this repository public with the text fields included.

1,069 open binary Kalshi markets (one per event). Each row is a forecasting question with two instants:

• news_cutoff — no article in the row is newer (seen time from GDELT / Bing and the page's own date are both at or before it).
• t_0 — the forecast time: the market price market_mid_t0 is read at t_0, and the proxy label is the market price at t_0 + 24 h (labels.jsonl). Here t_0 = news_cutoff for every row.

Rows	news_cutoff = t_0 (UTC)	News sources	Label t_0 + 24 h available
203	2026-10-28 08:30	GDELT only	yes (2026-10-29 08:30)
866	2026-10-29 09:00-10:26 (minute of the Bing retrieval)	GDELT (to 2026-10-28 08:30) + Bing News	from 2026-10-30 ~10:30

Each row also has two ready prompts (with and without news; "Today's date" = the date of t_0), at least three distinct articles a judge model rated relevant in the information block, and the complete list of everything retrieved. Real outcomes arrive as the markets settle over the coming weeks and months. Built by team The Predictables (LMU, EU AI hackathon 2026); code and runbook in the team repository (docs/PLAN_live_news.md, "Runbook as built").

Contents

File	What
data/kalshi_news3.jsonl	1,069 rows (~220 MB), one per market
data/summary.json	counts per selection, category and deadline month; rows per relevance threshold
data/labels.jsonl	per row: mid_t_0, mid_t_0_plus_24h (null until that instant has passed and a refresh ran), latest quote, status, resolution once settled

Rows with at least k relevant articles in their information block (news.n_in_block >= k): 10 → 223, 8 → 363, 6 → 560, 5 → 694, 4 → 866, 3 → 1,069.

Categories: elections 409, politics 168, sports 146, economics 87, entertainment 81, financials 78, science and technology 36, climate and weather 21, companies 13, crypto 12, other 18. 447 rows come from a stratified random sample of 1,000 events, 622 from 3,000 events picked for news coverage — the set over-represents well-covered topics.

How it was built

1. Universe: every open Kalshi event on 2026-10-27 13:40 UTC (14,605 events / 150,665 markets). Kept: binary, active, a quote on the book, deadline at least 48 h after 2026-10-28 08:30, no multivariate combos; one market per event (highest volume, mid between 0.05 and 0.95 in the dump), at most 10 events per series.
2. News, per selected market (not per event: multi-market events would otherwise share off-target articles):
   • round 1: GDELT GKG 2.1 bulk files of the 30 days before 2026-10-28 08:30 (2.9 M English articles), local index over titles and named entities; the market-specific terms must match;
   • round 2 (random-sample rows): gpt-oss-20b writes four keyword queries (aliases, organisations, topic) for markets still short of relevant articles, run on the same GDELT index;
   • round 3: Bing News RSS with the market question (random-sample rows also with three of the model queries), no lower date limit; the row's news_cutoff (and t_0) becomes the retrieval minute.
3. Full text with trafilatura; judge: gpt-oss-20b rates every article 1-6 for relevance to the market question and its rules and writes a 1-3 sentence summary about the question.
4. Information block (FAL format): the 10 most relevant articles with relevance >= 3, at most two per domain, without syndicated copies (duplicate_of: same canonical URL, same title without site suffix, or >= 60 % shared 5-word shingles; the most relevant / longest copy is kept) and without articles that quote prediction-market odds (mentions_market, which would leak the price label).
5. Prices from Kalshi's minute candles by market ID, read after every t_0 (refresh 2026-10-29 13:01 UTC): mid = (yes bid + yes ask) / 2 of the last candle at or before the instant.

Fields (one row)

• id (Kalshi ticker), event_ticker, series_ticker, category, is_sports, event_title, event_sub_title
• question (market question; the event title is prefixed when the market title alone is terse), resolution_criteria (Kalshi rules), resolution_deadline_utc, early_close_condition, settlement_sources
• news_cutoff, t_0, market_mid_t0 (market mid at t_0), deadline_before_label (true if the market's deadline is before t_0 + 24 h, so that label may be a settlement price); market_raw (full Kalshi market object from the universe dump)
• prompt_news (FAL-style prompt with the information block), prompt_nonews (same without articles), information_block
• news: n_in_block, n_judged, n_text, n_candidates, relevance_hist, n_domains, rounds (queries and windows per round), and articles — the complete list, each with round, rank, url, title, domain, seen_at, published_at, retrieval_score, query, text, text_chars, fetched, fetch_note, relevance, summary, in_information_block, duplicate_of, mentions_market
• resolution (null until the market settles), status, source_selection (kalshi_1k or kalshi_rich)

Caveats

• Targets are prices after t_0 (labels.jsonl: mid_t_0_plus_24h); market_mid_t0 is the market's view at forecast time — a baseline, not a label.
• Relevance comes from one 20b judge; relevance 2 summaries almost always say the article does not address the question, hence the threshold of 3.
• Page texts were fetched on 2026-10-28/29; a page can have been edited after publication.
• Coverage is uneven: sports props, earnings mentions and streaming counts rarely have relevant news.