CoChat

A statistical theory of optimal decision-making in sports betting

Open access
Dmochowski, Jacek P. · 2023 · PLOS ONE

The DOI matches its Crossref record; 42 of the 42 references it cites were found in a registry or an index; none of the 15 sources its claims cite is checked for support yet.

Abstract

The recent legalization of sports wagering in many regions of North America has renewed attention on the practice of sports betting. Although considerable effort has been previously devoted to the analysis of sportsbook odds setting and public betting trends, the principles governing optimal wagering have received less focus. Here the key decisions facing the sports bettor are cast in terms of the probability distribution of the outcome variable and the sportsbook's proposition. Knowledge of the median outcome is shown to be a sufficient condition for optimal prediction in a given match, but additional quantiles are necessary to optimally select the subset of matches to wager on (i.e., those in which one of the outcomes yields a positive expected profit). Upper and lower bounds on wagering accuracy are derived, and the conditions required for statistical estimators to attain the upper bound are provided. To relate the theory to a real-world betting market, an empirical analysis of over 5000 matches from the National Football League is conducted. It is found that the point spreads and totals proposed by sportsbooks capture 86% and 79% of the variability in the median outcome, respectively. The data suggests that, in most cases, a sportsbook bias of only a single point from the true median is sufficient to permit a positive expected profit. Collectively, these findings provide a statistical framework that may be utilized by the betting public to guide decision-making.

Overview

CoChat Summary

TL;DR

The paper develops a statistical framework for optimal sports-betting decisions, showing that the median outcome and nearby quantiles determine when to bet, which side to choose, and when to avoid wagering. Using 5,412 NFL regular-season games from 2002–2022, it finds sportsbooks’ point spreads and totals explain 86% and 79% of the variability in the median outcome, and that in many cases a 1-point bias from the true median is enough to permit positive expected profit.

Key findings

  • For point-spread betting, the bettor should wager on the home team if and only if the spread is less than the relevant quantile of the margin of victory; with equal payouts (typical 0.91), this becomes a rule based on the median margin of victory.
  • A positive expected profit is only possible if the sportsbook’s spread lies outside the interval between the 0.476 and 0.524 quantiles; for standard 4.5% commission, if the spread is within 2.4 percentiles of the true median, wagering yields negative expected profit.
  • In NFL data, sportsbook point spreads explained 86% of the variability in the true median margin of victory (r² = 0.86, n = 21), with OLS slope 0.93 and intercept -0.41; sportsbook point totals explained 79% of the variability in the median point total (r² = 0.79, n = 24).
  • For the largest stratified NFL samples, a deviation of only 1 point from the true median was enough to permit positive expected profit in point-spread betting and in point-total betting; across all stratifications, expected profit increased with 1-, 2-, and 3-point deviations, e.g. spread overestimation by 1/2/3 points yielded 0.022 ± 0.011, 0.090 ± 0.021, and 0.15 ± 0.030 expected profit.

Methods

  • The paper develops a theoretical model of point-spread, moneyline, and over-under betting using the outcome distribution and sportsbook proposition, deriving theorems/corollaries for optimal side selection, when to wager, and error bounds.
  • Empirical analysis uses historical NFL regular-season games from 2002–2022 (n = 5,412), with sportsbook point spreads, point totals, payouts, and consensus values; for spread and total analyses, matches were stratified by sportsbook line into groups with at least 100 matches each.
  • Quantiles and medians were estimated with kernel density estimation (Gaussian kernel, bandwidth 2), confidence intervals used 1,000 bootstrap resamples, and relationships between sportsbook lines and medians were assessed with ordinary least squares regression.

Limitations

  • The paper states it does not propose or analyze specific predictors/features or models; it focuses on the wagering theory rather than forecasting models.
  • The empirical validation is limited to NFL regular-season games; the text notes that outcomes and quantiles may differ across sports such as the NBA.
  • The stratification process did not account for varying payouts (for example, -110 vs -105), which the paper says likely introduces negligible error but is still a simplifying assumption.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1371/journal.pone.0287601

Cites

42 of 42 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig 1 How accurately do sportsbooks predict the median outcome? (a) The distribution of margin of victory for National Football League matches with a consensus sportsbook point spread of s = 6. The median outcome of 4.26 (dashed orange line, computed with kernel density estimation) fell below the sportsbook point spread (dashed blue line). However, the 95% confidence interval of the sample median (2.27-6.38) contained the sportsbook proposition of 6. (b) Same as (a), but now showing the distribution of point total for all matches with a sportsbook point total of 46. Although the sportsbook total exceeded the median outcome by approximately 1.5 points, the confidence interval of the sample median (42.25-46.81) contained the sportsbook’s proposition. (c) Combining all stratified samples, the sportsbook’s point spread explained 86% of the variability in the median margin of victory. The confidence intervals of the regression line’s slope and intercept included their respective null hypothesis values of 1 and 0, respectively. (d) The sportsbook point total explained 79% of the variability in the median total. Although the data hints at an overestimation of high totals and underestimation of low totals, the confidence intervals of the slope and intercept contained the null hypothesis values.
Fig 1 How accurately do sportsbooks predict the median outcome? (a) The distribution of margin of victory for National Football League matches with a consensus sportsbook point spread of s = 6. The median outcome of 4.26 (dashed orange line, computed with kernel density estimation) fell below the sportsbook point spread (dashed blue line). However, the 95% confidence interval of the sample median (2.27-6.38) contained the sportsbook proposition of 6. (b) Same as (a), but now showing the distribution of point total for all matches with a sportsbook point total of 46. Although the sportsbook total exceeded the median outcome by approximately 1.5 points, the confidence interval of the sample median (42.25-46.81) contained the sportsbook’s proposition. (c) Combining all stratified samples, the sportsbook’s point spread explained 86% of the variability in the median margin of victory. The confidence intervals of the regression line’s slope and intercept included their respective null hypothesis values of 1 and 0, respectively. (d) The sportsbook point total explained 79% of the variability in the median total. Although the data hints at an overestimation of high totals and underestimation of low totals, the confidence intervals of the slope and intercept contained the null hypothesis values.CC BY · PubMed Central
Fig 2 Do sportsbook point spreads deviate from the 0.476-0.524 quantiles? With a standard payout of ϕ = 0.91, achieving a positive expected profit is only feasible if the sportsbook point spread falls outside of the 0.476-0.524 quantiles of the margin of victory. The 0.476 and 0.524 quantiles were thus estimated for each stratified sample of NFL matches. Light (dark) black bars indicate the 95% confidence intervals of the 0.476 (0.524) quantiles. Orange markers indicate the sportsbook point spread, which fell within the quantile confidence intervals for the large majority of stratifications. An exception was s = 5, where the sportsbook appeared to overestimate the margin of victory. For two other stratifications (s = 3 and s = 10), the 0.524 quantile tended to underestimate the sportsbook spread, with the 95% confidence intervals extending to just above the spread.
Fig 2 Do sportsbook point spreads deviate from the 0.476-0.524 quantiles? With a standard payout of ϕ = 0.91, achieving a positive expected profit is only feasible if the sportsbook point spread falls outside of the 0.476-0.524 quantiles of the margin of victory. The 0.476 and 0.524 quantiles were thus estimated for each stratified sample of NFL matches. Light (dark) black bars indicate the 95% confidence intervals of the 0.476 (0.524) quantiles. Orange markers indicate the sportsbook point spread, which fell within the quantile confidence intervals for the large majority of stratifications. An exception was s = 5, where the sportsbook appeared to overestimate the margin of victory. For two other stratifications (s = 3 and s = 10), the 0.524 quantile tended to underestimate the sportsbook spread, with the 95% confidence intervals extending to just above the spread.CC BY · PubMed Central
Fig 3 Do sportsbook point totals deviate from the 0.476-0.524 interval? The 0.476 and 0.524 quantiles of the true point total were estimated for each stratified sample of NFL matches. For all but one stratification (t = 47, 95% confidence interval [41.59-45.42], sportsbook overestimates the total), the confidence intervals of the sample quantiles contained the sportsbook proposition. Visual inspection of the data suggests that, in the NFL betting market at least, sportsbooks are very adept at proposing totals that fall within the critical 0.476-0.524 quantiles.
Fig 3 Do sportsbook point totals deviate from the 0.476-0.524 interval? The 0.476 and 0.524 quantiles of the true point total were estimated for each stratified sample of NFL matches. For all but one stratification (t = 47, 95% confidence interval [41.59-45.42], sportsbook overestimates the total), the confidence intervals of the sample quantiles contained the sportsbook proposition. Visual inspection of the data suggests that, in the NFL betting market at least, sportsbooks are very adept at proposing totals that fall within the critical 0.476-0.524 quantiles.CC BY · PubMed Central
Fig 4 How large of a bias in the point spread is required for positive expected profit? In order to estimate the magnitude of the deviation between sportsbook point spread and median margin of victory that is required to permit a positive profit to the bettor, the hypothetical expected profit was computed for point spreads that differ from the true median by 1, 2, and 3 points in each direction. The analysis was performed separately within each stratified sample, and the figure shows the results of the four largest samples. For 3 of the 4 stratifications, a sportsbook bias of only a single point is required to permit a positive expected return (height of the bar indicates the expected profit of a unit bet assuming that the bettor wagers on the side with the higher probability of winning; error bars indicate the 95% confidence intervals as computed with the bootstrap). For a sportsbook spread of s = 3 (dark black bars), the expected profit on a unit bet is 0.021 [0.008-0.035], 0.094 [0.067-0.119], and 0.166 [0.13-0.2] when the sportsbook’s bias is +1, +2, and +3 points, respectively (mean and confidence interval over 500 bootstrap resamples).
Fig 4 How large of a bias in the point spread is required for positive expected profit? In order to estimate the magnitude of the deviation between sportsbook point spread and median margin of victory that is required to permit a positive profit to the bettor, the hypothetical expected profit was computed for point spreads that differ from the true median by 1, 2, and 3 points in each direction. The analysis was performed separately within each stratified sample, and the figure shows the results of the four largest samples. For 3 of the 4 stratifications, a sportsbook bias of only a single point is required to permit a positive expected return (height of the bar indicates the expected profit of a unit bet assuming that the bettor wagers on the side with the higher probability of winning; error bars indicate the 95% confidence intervals as computed with the bootstrap). For a sportsbook spread of s = 3 (dark black bars), the expected profit on a unit bet is 0.021 [0.008-0.035], 0.094 [0.067-0.119], and 0.166 [0.13-0.2] when the sportsbook’s bias is +1, +2, and +3 points, respectively (mean and confidence interval over 500 bootstrap resamples).CC BY · PubMed Central
Fig 5 How large of a bias in the point total is required for positive expected profit? Vertical axis depicts the expected profit of an over-under wager, conditioned on the sportsbook’s posted total deviating from the true margin by a value of 1, 2, or 3 points (horizontal axis). The analysis was performed separately for each unique sportsbook total, and the figure displays the results for the four largest samples. A deviation from the true median of a single point permits a positive expected profit in all four of the depicted groups. For a sportsbook total of t = 44 (green bars), the expected profit on a unit bet is 0.015 [0.004-0.028], 0.075 [0.053-0.10], and 0.13 [0.10-0.17] when the sportsbook’s bias is +1, +2, and +3 points, respectively (mean and confidence interval over 500 bootstrap resamples).
Fig 5 How large of a bias in the point total is required for positive expected profit? Vertical axis depicts the expected profit of an over-under wager, conditioned on the sportsbook’s posted total deviating from the true margin by a value of 1, 2, or 3 points (horizontal axis). The analysis was performed separately for each unique sportsbook total, and the figure displays the results for the four largest samples. A deviation from the true median of a single point permits a positive expected profit in all four of the depicted groups. For a sportsbook total of t = 44 (green bars), the expected profit on a unit bet is 0.015 [0.004-0.028], 0.075 [0.053-0.10], and 0.13 [0.10-0.17] when the sportsbook’s bias is +1, +2, and +3 points, respectively (mean and confidence interval over 500 bootstrap resamples).CC BY · PubMed Central

Verification

Claims the paper makes

0 Valid · 4 Not valid · 8 Not checked · 96 Own results and interpretations

73 more claims

Create a free account to see all 108 claims and to verify the 2 references still open.

See all claims — free account
  • It may state two separate claims as one. 20
  • Probably not a substantial claim. 13
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Citations

Cites

Research works printed but not linked: 1

  1. Neal B, Mittal S, Baratin A, Tantia V, Scicluna M, Lacoste-Julien S, et al. A modern take on the bias-variance tradeoff in neural networks. arXiv preprint arXiv:181008591. 2018;.

Cited by

See all 5

Metadata

Authors
Dmochowski, J. P.
Affiliations
City College of New York (US)
Corresponding author
Dmochowski, Jacek P.
Venue
PLOS ONE
Published
2023-06-28
Volume
18
Issue
6
Pages
e0287601
ISSN
1932-6203 (ISSN-L)
OpenAlex
W4382468061
Cited by
5
References
42
Citations by year
2026: 1 · 2025: 2 · 2024: 2
FWCI
1.02
Type
journal article
Language
en
Open access
Locations
  • PLoS ONE (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Social Sciences › Economics, Econometrics and Finance › Economics and Econometrics › Sports Analytics and Performance
  • Physical Sciences › Mathematics › Statistics and Probability › Probability and Statistical Research
  • Social Sciences › Decision Sciences › Management Science and Operations Research › Forecasting Techniques and Applications
Keywords
sports betting, point spread, quantiles, expected profit, National Football League
MeSH terms
Football, Gambling, Humans, North America

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.