cfb-mainlines / README.md
SmartStake's picture
College football mainline odds (2026), the dataset behind The Closing Line issue 7
1e72f5f verified
|
Raw History Blame Contribute Delete
6.16 kB
metadata
license: cc-by-4.0
language:
  - en
pretty_name: SmartStake College Football Mainline Odds (2026)
tags:
  - sports-betting
  - sports-analytics
  - college-football
  - ncaaf
  - odds
  - prediction-markets
size_categories:
  - 100M<n<1B
task_categories:
  - time-series-forecasting
configs:
  - config_name: default
    data_files:
      - split: train
        path: kick_day=*/*.parquet

SmartStake College Football Mainline Odds (2026)

The exact dataset behind the study "Who's Actually Sharp on College Football Mainlines?": every pregame price change on college football moneylines, spreads, and totals at the twelve sportsbooks, exchanges, and prediction markets in the study, for the 1,067 games that kicked off from 27 August through 31 October 2026. A companion to SmartStake MLB Mainline Odds (2026).

Coverage

  • Games: 1,067, first kickoff 27 August 2026, last kickoff 31 October 2026 (US time).
  • Window: pregame only. Every row's price was posted before kickoff (ts < start_ms). The earliest quotes date from 7 August.
  • Markets: full-game moneyline, spread, and total, at every line each book posted (alternate spreads and totals included, as in the study).
  • Books: Pinnacle, Bookmaker, BetOnline, Kalshi, Polymarket, Novig, ProphetX, 4Caster, FanDuel, DraftKings, Fanatics, and Caesars.
  • Rows: 172,700,424 (about 1 GB of Parquet, one file per kickoff day).

One row is a run of one price

Books re-send an unchanged price over and over. Each row here is one run: a book's price for one selection staying the same from ts (first seen) to ts_last (last re-sent). A new row starts when the price, or whether it is bettable, changes. A price holds until the next row for the same selection, so to read a book's price at time T, take the latest row with ts <= T.

Sportsbooks (Pinnacle, Bookmaker, BetOnline, Novig, ProphetX, FanDuel, DraftKings, Fanatics, Caesars) are at the feed's millisecond resolution.

Kalshi, Polymarket, and 4Caster show many resting offers at once, so their rows are the best price you could actually bet, minute by minute: ts and ts_last fall on whole minutes, and a row with eff = null marks a minute with nothing bettable. A 4Caster offer counts as gone if it is not re-posted within 90 seconds, and a Kalshi offer under $100 is ignored rather than treated as a pulled price.

A selection is (matchKey, market, line, side, book, contract).

Schema

column type description
matchKey string Per-game identity. Group on this.
market string Moneyline, Spread, or Total Points.
book string Sportsbook, exchange, or prediction market.
contract string Kalshi's contract id (Kalshi lists each team as its own contract). Empty for every other book.
side string A = home (moneyline, spread) or over (total); B = away or under.
line double 0 for moneylines; the home team's spread for spreads (so both sides of a spread share one line); the total for totals.
ts int64 First time this price was seen, epoch milliseconds (UTC).
ts_last int64 Last time the book re-sent this same price, epoch milliseconds (UTC).
start_ms int64 Scheduled kickoff, epoch milliseconds (UTC).
home string Home team.
away string Away team.
eff double The bettable decimal price. Null when the price was pulled, or when the book reported a max bet under $100.
del bool The book flagged the selection as removed while still showing a price (a feed refresh; it usually comes back).
mx double The book's reported max bet (or size at the price) when the run started, where the feed carries one.
kick_day date Partition key: the kickoff date, taken as kickoff time minus 8 hours in UTC (so a late-night Pacific kickoff stays on its Saturday).

How it was built

  1. scripts/extract_day.sh pulls each day of SmartStake's raw odds archive down to pregame NCAAF moneylines, spreads, and totals at the twelve books.
  2. scripts/stage1_compact.sh turns the sportsbooks' ticks into the runs above. A pulled price (null odds, or odds of 1 or less) and a price with a reported max bet under $100 both become eff = null, so neither counts as a price anyone could bet.
  3. scripts/xraw_day.sh and scripts/xtop_day.sql do the same for the three exchanges, keeping each resting offer separate and then taking the best live one each minute.

The study's other rules are applied in the analysis, not here: a new exchange price has to hold for five minutes before it counts as a move, and a cross showing an edge above 5% is thrown out as a bad print.

Notes and caveats

  • Odds are raw. Nothing is de-vigged. The study de-vigs each book's two sides before comparing books, except for crossed markets, which use posted prices.
  • Exchange bots are left in on purpose. Refill-bot prints are part of what the study measures; the five-minute hold above is how the study keeps them from counting as moves.
  • Sparse feeds. BetOnline and Fanatics confirm their prices rarely, so a run can last hours without the book re-sending it. Check ts_last before treating an old price as live.
  • No results. Game outcomes are not included.

Loading

import pandas as pd
df = pd.read_parquet("hf://datasets/SmartStake/cfb-mainlines/kick_day=2026-10-03")
-- DuckDB, straight from the hub
SELECT book, count(*) FROM 'hf://datasets/SmartStake/cfb-mainlines/kick_day=*/*.parquet' GROUP BY 1;

Provenance, license, and responsible use

Odds were collected from public sportsbook and exchange feeds. Released under CC BY 4.0 for research and educational use. This dataset is informational and historical: past prices do not predict future outcomes, and betting carries risk of loss. Not affiliated with any sportsbook. 21+.

Citation

SmartStake (2026). SmartStake College Football Mainline Odds (2026). Hugging Face.