Research
Experiments we actually ran, written up with the methodology, the numbers, and the limitations. Free to cite with attribution. No picks, no affiliate links — just what the data shows.
Market margin ≈ 4.5–6% · line shopping removed about half of it (57% in UFC, only 29% in boxing)
What the vig actually costs, by sport: 315 games of measured bookmaker margin
We measured the real bookmaker margin across six sports from our own captured odds — and how much of it line shopping removes. Boxing was the most expensive market; shopping cut the margin roughly in half almost everywhere.
Read the study →65% winner accuracy → −33% bankroll betting every pick, −50%+ betting only its “value” picks
A 65%-accurate NBA model still lost money: a full-season market-efficiency test
We built a walk-forward Elo model, predicted 1,200+ NBA games at 65% accuracy, and bet it against real closing lines. It lost — and its most confident disagreements with the market lost fastest.
Read the study →Same model, same method: NBA 65% winner accuracy vs NHL ~53% — a near coin-flip
How predictable is each sport? The same model called 65% of NBA winners — and 53% of NHL winners
An identical rating model run across full NBA and NHL seasons. Basketball was far more predictable than hockey — with direct consequences for variance, sample sizes, and staking by sport.
Read the study →How these are run
We publish a study when we have run something ourselves and the answer is useful whether or not it flatters us. Two of the three below found that an approach people commonly recommend does not work.
Our own captured data
Odds come from prices we recorded ourselves at the time, not from a vendor's summary. The vig study measures 315 games of real bookmaker margin; the archive behind it keeps one permanent closing record per game, market and side.
Walk-forward, never retro-fitted
Any model in a study makes each prediction using only what was known at that point in the season. It is the difference between a result and a backtest that has quietly seen the answer.
Limitations stated, not buried
Every study ends with what would change the result and what it does not cover. Where a comparison is missing — a naive baseline, a second season — we say so on the page rather than leaving you to notice.
No commercial interest in the answer
We have no sportsbook affiliate, referral or revenue-share relationships of any kind. Nothing here is written to move you toward a book, because there is nothing in it for us if you go.
Citing this research
Free to cite, quote and chart with attribution to Bankroll Guardian and a link to the study page — no permission needed, for commercial outlets as well as personal ones. If you are writing something up and want a methodology detail, a figure checked, or a cut of the data we have not published, get in touch and we will answer properly rather than send a press release.
How we compute everything
Fair no-vig odds, market consensus, EV, CLV, and out-of-sample validation — documented in full on the Methodology page.