How to Research an NBA Player Prop: Build an Evidence Note First

Research an NBA player prop by defining the event, checking the games behind the statistics, recording changes in role, and separating facts from assumptions. The result should be an evidence note you can reconstruct after the game. A high historical hit rate is only one input. The player, rebound totals, and odds in this guide are fictional teaching examples.

Define the exact market

Start with a question that stays the same throughout your research. Sandve and colleagues recommend preserving the inputs and settings behind computational results so someone can reconstruct them. Sandve et al. (2013) Applied to a player prop, that means recording the player, game, statistic, threshold, price, time checked, and settlement rules. In our example, over 8.5 rebounds means nine or more. If the line moves to 9.5, the earlier count still describes nine or more and must be recalculated for ten. A precise market definition gives every later number a stable target.

Save what was knowable before tipoff

Keep a dated boundary around your evidence. Kapoor and Narayanan explain how information leakage can produce misleading evaluations when a test includes information unavailable in the intended real-world setting. Kapoor and Narayanan (2023) That finding concerns scientific machine learning; the practical lesson here is to preserve the data and reports available when you made your note. A rumored minutes restriction belongs under unresolved information until a credible dated report supports it. If new information arrives, timestamp the revision and state which assumption changed. This prevents a postgame explanation from quietly becoming a pregame reason.

Show the games behind the percentage

A hit rate needs a visible denominator and eligibility rule. Brown, Cai, and DasGupta show why uncertainty in small binomial samples needs careful treatment. Brown et al. (2001) Suppose a fictional center has 12 recorded appearances with rebound totals of 6, 10, 8, 12, 7, 9, 11, 5, 10, 8, 13, and 9. Seven reached nine rebounds, giving 58.3%. Two separate DNP fixtures should remain recorded as absences rather than zero-rebound appearances. Under independent, constant-probability teaching assumptions, a nominal 95% Wilson interval for 7/12 is about 32.0% to 80.7%. NBA roles and opponents change, so this interval is neither a forecast for tonight nor a cure for an unsuitable sample. Carry the count and its limitations together.

Choose filters for a basketball reason

Decide why a comparison matters before choosing the version that looks strongest. Cawley and Talbot show that model selection can overfit a noisy evaluation criterion and bias reported performance. Cawley and Talbot (2010) In a prop worksheet, searching many combinations of venue, minutes, opponent, and date window creates a similar selection problem. A documented rotation change can justify comparing roles, but a filter invented after seeing a favorable percentage should be labeled exploratory. Save the full sample and all filters tried. The next useful step is a future test or better role evidence, rather than another attractive slice of the same games.

Keep the price separate from the forecast

A quoted price supplies a break-even rate; it does not supply a trustworthy player probability. In the fictional +105 market, a standard cash wager with two win-or-lose outcomes breaks even at 100/205, or about 48.8%. Comparing that with the historical 58.3% does not prove an edge because the historical count has not been justified as tonight's probability. Gneiting and Raftery's work on proper scoring rules supports evaluating probability forecasts saved before their outcomes. Gneiting and Raftery (2007) One binary check is squared error, (p - y)^2, where y is 1 or 0. Keep forecast evaluation and obtainable prices in separate fields so neither can stand in for the other.

Use this evidence-note template

The completed note should expose unresolved questions as clearly as observed facts. The reproducibility guidance above supports keeping inputs and decisions together; the table shows how that works for our fictional center. It records a valid historical calculation while leaving tonight's role, participation, and probability unresolved. That makes the next action specific: check the missing inputs or retain a documented pass. A research process can be useful without ending in a wager.

For current product information, visit the StatChecker NBA 26/27 launch page.

References

• Geir Kjetil Sandve, Anton Nekrutenko, James Taylor, and Eivind Hovig (2013). Ten Simple Rules for Reproducible Computational Research. PLOS Computational Biology, 9(10), e1003285. DOI: 10.1371/journal.pcbi.1003285.
• Sayash Kapoor and Arvind Narayanan (2023). Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4(9), 100804. DOI: 10.1016/j.patter.2023.100804.
• Lawrence D. Brown, T. Tony Cai, and Anirban DasGupta (2001). Interval Estimation for a Binomial Proportion. Statistical Science, 16(2), 101-133, including discussion. DOI: 10.1214/ss/1009213286.
• Gavin C. Cawley and Nicola L. C. Talbot (2010). On Over-fitting in Model Selection and Subsequent Selection Bias in Performance Evaluation. Journal of Machine Learning Research, 11, 2079-2107.
• Tilmann Gneiting and Adrian E. Raftery (2007). Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association, 102(477), 359-378. DOI: 10.1198/016214506000001437.

Prepared with AI assistance and source checks. Published by StatChecker.