Analysis
Forecasting Accuracy in NBA Game Outcomes
A Comparative Analysis of Bookmakers and Public Prediction Markets
Koleman StrumpfProfessor
Open main menu
The following is a paper created by a group of students in Professor Strumpf's ECN 345 class at Wake Forest University.
The rise of data-driven decision-making has sparked growing interest in the predictive power of betting markets, where professionals and the public attempt to forecast the outcomes of competitive events. In professional basketball, bookmakers and public prediction markets offer two distinct approaches to forecasting NBA game results. While professional bookmakers have long held reputations as reliable and accurate predictors due to their access to sophisticated models and large datasets, the emergence of decentralized public prediction markets, where everyday participants bet using real money, offers a compelling alternative grounded in the "wisdom of crowds."
This paper seeks to evaluate the predictive accuracy of these two systems during the 2024–25 NBA regular season. Specifically, we compare the forecasts generated by professional bookmakers, aggregated via OddsPortal, with those from the public prediction platform Polymarket. By converting betting odds into implied probabilities and comparing those to actual game outcomes, we assess which system more accurately anticipates winners across a wide sample of games.
Understanding these differences is more than a niche inquiry into sports analytics—it also holds broader implications for economic theories of market efficiency, behavioral finance, and forecasting reliability. If prediction markets can match or even outperform professional bookmakers, this could reshape how we value decentralized information aggregation in high-stakes environments.
We begin with the following hypotheses:
Through statistical analysis, we aim to determine whether one method offers a consistently superior forecast or if both reflect a similarly efficient processing of available information.
The question of how accurately different market structures predict real-world outcomes has been explored in several economic and behavioral finance studies. Central to this line of research is the comparison between expert-based systems, such as professional bookmakers, and decentralized systems, such as prediction markets. These two forecasting mechanisms rely on fundamentally different assumptions about information aggregation, participant incentives, and market behavior.
A seminal paper by Franck, Verbeek, and Nüesch (2010) offers one of the most comprehensive comparisons between professional bookmakers and betting exchanges. Using data from 5,478 football matches across the top five European leagues (England, Germany, Spain, Italy, and France) over three seasons, the authors compared the predictive performance of traditional bookmakers with that of Betfair, a peer-to-peer betting exchange. By employing statistical tools like the probit model, Brier scores, and goodness-of-fit metrics, they found that Betfair markets, driven by individual bettors wagering their own money, produced more accurate forecasts than bookmakers. This finding supports the hypothesis that markets driven by real financial risk and participant diversity can outperform professionally set odds.
In a related study, Servan-Schreiber et al. (2004) investigated whether the presence of real monetary stakes improves prediction accuracy in sports markets. The study compared TradeSports, a real-money betting site, with NewsFutures, a play-money market offering small prizes. Using a sample of 208 NFL games, they found that both markets produced similarly accurate forecasts, with 65.9% of TradeSports' favorite teams winning their games and 66.8% of NewsFutures' favorites also winning. This suggests that while financial stakes might slightly sharpen predictions, even play-money markets can yield meaningful forecasts, likely due to competitive pressure, frequent, accessible, and accurate NBA sports news information, and high participant engagement. A key element in this study was that in the process of analyzing both sites the authors sorted each trading price into bins at 10% intervals (from 0-100); the binning method allowed for a closer comparison of tradesports and new futures accuracy in observed victory of an NFL compared to the market price for that outcome. Across all bins, to total correlation coefficient between trading prices and the outcome of an NFL game was 0.96 for TradeSports and 0.94 for NewsFutures.
In another degree of analysis, the authors extended their investigation to compare the overall accuracy of prediction markets and a formal betting market (with real money) to that of the accuracy of individual betting in a market. To do this, the authors took data from a site called ProbabilityFootball; in this prediction contest, participants are asked to provide a percentage likelihood of a game's outcome. Overall, the authors found that both Tradesports and Newsfutures had a higher forecast accuracy than the average percentages for each game on Probability football, and when the average prices for each market were entered against individual betters in the contest, Newsfutures ranked 11th and Tradesports ranked 12th out of 1,947. In essence, the work "Does Money Matter?" demonstrates the accuracy of both a prediction market and a betting site (with the cumulative wisdom of the entire market) over one individual bettor as seen by their outperformance in the probability Football competition.
These case studies underscore an important distinction in forecasting systems: bookmakers operate within a framework that balances risk and profit, and may adjust lines to account for public sentiment, not just outcome likelihood. Prediction markets, in contrast, are more direct aggregators of participant beliefs and often reflect a broader base of information, particularly when participation is both incentivized and decentralized.
Our study builds on this literature by applying similar comparative techniques to a different context: the NBA, during the 2024–25 season. While prior work has focused heavily on European football and NFL games, relatively few studies have examined the predictive performance of public prediction markets versus bookmakers in professional basketball, and fewer still have done so using data from modern, blockchain-based platforms like Polymarket. By revisiting this comparison in a new sport and market environment, our research contributes to a growing body of literature assessing the reliability of decentralized forecasting systems and the conditions under which they may match or exceed the performance of professional institutions.
At the heart of this study lies the economic concept of information aggregation—the idea that markets can efficiently combine dispersed information held by individuals to produce accurate forecasts. In financial and betting markets, this aggregation is often evaluated through the lens of market efficiency, which posits that prices (or odds, in this case) fully reflect all available information.
Prediction markets, such as Polymarket, operate as platforms where participants buy and sell shares in the outcome of future events. The market price of a contract for a particular outcome (e.g., a team winning a game) reflects the collective belief about the probability of that outcome. These platforms rely on the "wisdom of crowds" hypothesis: the belief that a diverse group of independent individuals can, on average, make better predictions than experts. When participants risk real money, they are incentivized to seek accurate information, evaluate probabilities critically, and avoid biases.
In contrast, professional bookmakers use internal models, historical data, and expert intuition to set odds, which they then adjust based on betting volume to balance their risk exposure. Their incentives are not solely to predict outcomes accurately but also to ensure profitability by attracting balanced bets on both sides of an event. This means their odds may incorporate elements of behavioral economics, reflecting public sentiment or psychological pricing strategies that deviate from pure probability.
To evaluate predictive accuracy, we use statistical tools like the Brier score, which measures the mean squared difference between predicted probabilities and actual outcomes. Lower Brier scores indicate greater predictive accuracy. We also examine calibration (how well predicted probabilities match observed frequencies) and discrimination (how well the model distinguishes between different outcomes).
Finally, this study touches on elements of behavioral bias and reputation economics. Bookmakers have long held social legitimacy and trust due to their established presence and perceived expertise, whereas public prediction markets may be unfamiliar to many and perceived as less credible, even if they perform equally well. By comparing these two systems, we gain insight into not only forecasting ability but also how reputation, risk, and structure influence market outcomes.
To compare the predictive accuracy of professional bookmakers and public prediction markets, we compiled a dataset spanning the entire 2024–25 NBA regular season. This dataset includes 1,000 games and reflects two distinct forecasting systems: traditional sportsbooks, represented by OddsPortal, and decentralized public prediction markets, represented by Polymarket.
OddsPortal aggregates moneyline odds from a wide range of major sportsbooks, including MGM, Bet365, and BetInAsia. For each game, we recorded the final pre-tip-off odds, capturing the most updated expectations of professional bookmakers. Because moneyline odds are expressed in American format, we converted them into implied probabilities to allow for meaningful comparison and quantitative analysis. The equations we used are:
For positive American odds: Implied Probability = 100 / (Odds + 100)
For negative American odds: Implied Probability = |Odds| / (|Odds| + 100)
For example, a moneyline of +150 becomes: Implied Probability = 100 / (150 + 100) = 0.4 A moneyline of -200 becomes: Implied Probability = 200 / (200 + 100) = 0.67
Excel Formula: =IF(Odds>0, 100/(Odds+100), IF(Odds<0, ABS(Odds)/(ABS(Odds)+100), ""))
This formula enabled consistent transformation of all odds into probabilities on a 0–1 scale, which was crucial for applying statistical accuracy measures like Brier scores. On the other hand, Polymarket functions as a blockchain-based prediction market in which individuals wager real money on the outcomes of future events, including sports games. For each NBA matchup, we recorded the closing market-implied probabilities just before tip-off to ensure alignment with bookmaker data and to capture real-time crowd sentiment. Since Polymarket prices already reflect implied probabilities, no further conversion was necessary.
Our analysis focuses on games involving twenty NBA teams, representing a diverse mix of high-profile and smaller-market franchises. These include the Milwaukee Bucks, Los Angeles Lakers, Miami Heat, Boston Celtics, Denver Nuggets, Chicago Bulls, New York Knicks, Minnesota Timberwolves, Los Angeles Clippers, Memphis Grizzlies, Toronto Raptors, Brooklyn Nets, New Orleans Pelicans, Indiana Pacers, Sacramento Kings, Phoenix Suns, Philadelphia 76ers, Orlando Magic, Charlotte Hornets, Cleveland Cavaliers, and Oklahoma City Thunder. This variety ensures coverage of different fan bases, performance levels, and betting dynamics.
After compiling the data, we matched the implied probabilities from both sources with actual game outcomes (win/loss), which were also sourced from OddsPortal. We then conducted statistical analyses using metrics such as Brier scores, correctness rates, and binning techniques. These tools allowed us to assess the degree to which each prediction source aligned with real world results, ultimately enabling a head-to-head comparison of forecasting performance between bookmakers and the public.
After assembling a dataset of 1,000 NBA games with matched bookmaker and public market predictions, we employed a series of statistical techniques to evaluate the forecasting accuracy of each source. Our primary objective was to determine how well each system's predicted probabilities aligned with actual game outcomes, both overall and within specific probability ranges.
To measure predictive performance, we primarily used the Brier score, a standard metric for probabilistic forecasts. The Brier score quantifies the mean squared difference between predicted probabilities and actual outcomes (coded as 1 for a win, 0 for a loss). Lower Brier scores indicate more accurate and better-calibrated predictions. This metric was calculated for each prediction source across all games, providing a robust average measure of forecast quality. In addition to Brier scores, we computed simple correctness rates—the proportion of games in which each source's forecast correctly identified the winning team based on the higher implied probability. This offered a straightforward comparison of directional accuracy.
We also performed a binning analysis, grouping predictions into intervals based on their confidence level (e.g., 0.30–0.39, 0.40–0.49, etc.). Within each bin, we calculated the proportion of correct predictions for both sources. This allowed us to evaluate how forecast accuracy varied by level of certainty and to assess whether either source systematically overestimated or underestimated win probabilities, especially in coin-flip scenarios versus high-certainty matchups.
To further examine the behavior of each system, we calculated the standard deviation of predicted probabilities and explored their correlation with actual outcomes. These measures helped identify whether one source consistently produced more conservative or aggressive probability estimates and whether their forecasts tracked with real-world results.
All calculations were conducted using spreadsheet software, and results were visualized through tables and charts to identify patterns and differences between the two systems. By applying the same statistical tools across a large and consistent sample, we ensured a fair and methodologically sound comparison of bookmaker and prediction market performance.
Overall, the 9 major sportsbooks that are recorded through Oddsportal maintained a prediction accuracy of NBA game outcomes of 66%. Polymarket maintained an overall predictive accuracy of 67%. By increasing our scope of data analysis to include 20 bins for all predictions within 5% of each other both polymarket and Oddsportal maintained a predictive correctness rate of over 90% for the for the four bins predicting above a 95% and or below 5% of a teams ability to win in any given game. Conversely, games that were in the center bins or considered a coin flip of a victory decreased both Oddsportal and Polymarkets' ability to predict a game's outcome very closely. Specifically, if games were within the bins with a prediction of victory between .30 and .70, then the probability of either Oddsportal or Polymarket being correct was below 70%.
Our analysis of 1,000 NBA games from the 2024–25 season revealed that professional bookmakers and public prediction markets predicted game outcomes with similar accuracy, both correctly identifying winners around 66% of the time. This finding is particularly notable given the structural, informational, and institutional differences between the two systems. The results suggest that prediction markets like Polymarket, where individuals wager real money, are capable of aggregating information as effectively as professional bookmakers, despite lacking formal modeling teams or decades of industry expertise. This supports the economic theory that markets composed of informed, incentivized participants can yield efficient outcomes, even without centralized control.
One surprising element of the analysis was the lack of clear superiority in either system. While bookmakers may be expected to have an edge due to proprietary analytics and refined risk models, their odds are also influenced by the need to balance betting volume, a factor that can lead them to adjust lines based on public sentiment rather than purely on win probabilities. Meanwhile, prediction markets draw directly from the collective beliefs of participants, which can incorporate real-time reactions to injury news, team trends, and other soft information that formal models might underweight.
From a behavioral economics perspective, this parity in performance raises questions about trust, perception, and institutional legitimacy. Professional bookmakers benefit from long-standing reputations and regulatory structures that make them appear more credible to the average bettor. Prediction markets, particularly decentralized platforms like Polymarket, may still suffer from user unfamiliarity and skepticism despite demonstrating similar levels of forecasting accuracy. This suggests that perceived legitimacy may not always align with empirical performance.
These findings also hold practical significance. For bettors, analysts, and fans seeking predictive insight, both systems offer value, though prediction markets may become increasingly attractive due to their transparency and crowd-sourced foundation. For policymakers and economists, the results reinforce the potential of decentralized prediction markets in non-sporting contexts, such as elections, economic forecasting, or even climate-related risk assessments.
Finally, the study contributes to academic debates around market efficiency, the wisdom of crowds, and the role of financial incentives in prediction accuracy. If both systems perform equally well in a highly scrutinized and competitive setting like NBA betting, it suggests that decentralized systems could serve as viable complements, or even substitutes, for expert-driven forecasting in other domains.
This study set out to evaluate the predictive accuracy of professional bookmakers and public prediction markets in the context of the 2024–25 NBA season. Using data from OddsPortal and Polymarket, we analyzed 1,000 games across 20 teams, converting all odds into implied probabilities and assessing the performance of each forecasting system using metrics like Brier scores and correctness rates.
Our findings show that there is no significant difference in predictive accuracy between the two systems. Both bookmakers and prediction markets accurately predicted game outcomes approximately 66% of the time. These results support our null hypothesis and suggest that both systems, despite their differing structures, motivations, and reputational standing, are similarly effective at aggregating information and forecasting sports outcomes.
This outcome aligns with previous literature suggesting that decentralized prediction markets, especially those involving real money, can perform on par with or even outperform expert-driven systems. While professional bookmakers benefit from sophisticated modeling and risk management strategies, public markets offer a broader, crowd-sourced approach that appears equally reliable in practice.
However, our study is not without limitations. While we focused on final pre-tip-off odds, we did not have access to information on trading volume, liquidity, or participant composition, factors that could influence the precision and robustness of prediction market probabilities. Moreover, Polymarket's structure, while innovative, is relatively new, and future data could reflect maturation in its forecasting behavior.
Looking ahead, several avenues for future research emerge:
In conclusion, our results highlight the viability of prediction markets as forecasting tools and challenge the notion that expert-driven bookmakers are inherently superior. As public markets continue to evolve and gain traction, they may offer an increasingly powerful and democratized alternative to traditional forecasting systems.
Franck, Egon P., Erwin Verbeek, and Stephan Nüesch. "Prediction Accuracy of Different Market Structures: Bookmakers Versus a Betting Exchange." International Journal of Forecasting, vol. 26, no. 3, 2010, pp. 448–459. ISU Working Paper No. 96. SSRN, https://ssrn.com/abstract=1503350.
OpenAI. ChatGPT Model Information. OpenAI, https://openai.com/.
Wolfers, Justin. Are Voters Rational? Evidence from Gubernatorial Elections. National Bureau of Economic Research, 2002, http://users.nber.org/~jwolfers/Papers/DoesMoneyMatter.pdf.
"Wikipedia: Sportsbook." Wikipedia, https://en.wikipedia.org/wiki/Sportsbook.