Beating the Bookies with Their Own Numbers (Kaunitz et al. 2017)

Overview

A widely-cited demonstration that football bookmakers can be consistently beaten — not by
building a better forecasting model, but by exploiting the probability information already
implicit in the published odds. The result is direct evidence that the football betting market
is inefficient, and a sober lesson in why that inefficiency is hard to monetize at scale.

The strategy — bet against the consensus, on outliers

The core idea inverts the usual "build a model to out-predict the bookmaker":

  1. Across many bookmakers, the average (consensus) odds for an outcome are a good estimate of
    its true probability — the "wisdom of the crowd of bookmakers."
  2. A bet has value when an individual bookmaker's odds deviate above that consensus by more
    than a threshold — that book is offering a mispriced (too-generous) price.
  3. Stake on those outliers. You are exploiting the spread between books, using the market's own
    numbers as ground truth, with no independent forecasting model.

Validation (the credible part)

The strategy was profitable across three independent settings:
- a 10-year historical simulation using closing odds,
- a 6-month simulation using minute-to-minute odds, and
- a 5-month live experiment staking real money with bookmakers.
Profitability held across thousands of games, in simulation and in real life. Code, data, and
models were released publicly (see "Dataset" below). See the paper for the specific returns.

The "rigged market" lesson (the practical caveat)

The paper's second half is the part most relevant to anyone planning to act on an edge:
bookmakers compensate for market inefficiency with discriminatory practices against successful
clients — limiting stakes to trivial amounts and closing winning accounts. So a real,
validated edge does not translate into scalable profit: the books simply refuse winners'
action. Monetizing an edge is a harder problem than finding one. Sharp exchanges (Betfair) and
low-margin books (Pinnacle) that don't limit winners are the usual answer — at the cost of a
sharper price to beat.

Dataset

The released "Beat the Bookie" closing-odds dataset — 479,441 football matches, Jan 2005 –
Jul 2015, across many leagues. Per match it records average and maximum closing 1X2 odds
across up to ~32 bookmakers (plus which book held the max and the book count) — it does not
expose per-bookmaker columns, so it gives a multi-book consensus (avg) and the best available
(max), not an isolated sharp book like Pinnacle. Covers the 2010 and 2014 World Cups (not
2018/2022). Kaggle: austro/beat-the-bookie-worldwide-football-dataset.

Relevance to this KB / our models

  • Corroborates world-cup-market-inefficiencies and market-efficiency: markets are
    beatable, especially where pricing disperses across books.
  • Method note: the multi-book average is a strong, transparent probability benchmark — often
    more robust than a single book — useful for validating a model (value-bet-identification).
  • Caveat for our betting work: even a confirmed edge faces account limiting; treat any
    positive backtest as "interesting," not "bankable," until the monetization path (exchange /
    non-limiting book) is solved.

See Also