PropJuice Terms & Glossary
A comprehensive glossary of terms used throughout the PropJuice platform, from edge calculations to model accuracy metrics.
Edge
The mathematical advantage a bet offers based on the difference between model probability and implied odds. A positive edge indicates expected profitability if the bet were made repeatedly. For example, if a model estimates a 60% win probability but the line implies 50% (even odds), the edge is approximately 10%.
Example: Model probability: 60%. Bet 100 times at -110 odds. Win 60 times (+$100 each) = $6,000. Lose 40 times (-$110 each) = -$4,400. Net profit: $1,600 over 100 bets, representing a 14.5% return on total risk.
+EV (Positive Expected Value)
A bet with positive expected value has a higher probability of winning than the odds imply. Over many bets, +EV wagers should produce profit even if any individual bet may lose. A +EV pick at +200 odds (33% implied probability) might have a model probability of 40%—meaning you lose more often than you win, but the payouts when you win more than compensate.
Example: A +200 underdog with 40% model probability: Bet $100 a hundred times. Win 40 times at +$200 = $8,000. Lose 60 times at -$100 = -$6,000. Net: +$2,000 profit despite 60% loss rate.
Ensemble Model
A prediction system that combines multiple independent models to produce a consensus output. By aggregating different algorithms, training approaches, and feature sets, ensembles typically outperform any single model. PropJuice uses 30+ models in its ensemble.
Example: Five models predict a player's points: 22, 24, 23, 25, 24. The ensemble consensus might be 23.6 points, smoothing out individual model variations.
Component Model
An individual model that contributes to an ensemble. Each component brings different strengths—some excel at certain game types, bet types, or conditions. Component performance is tracked individually to weight their contributions appropriately.
Example: A neural network component might have 58% accuracy on NBA totals while a gradient boosting component hits 61% on the same bets. The ensemble weights the gradient boosting model more heavily for this bet type.
Projection
A model's predicted value for a game outcome or player statistic. Projections may be point estimates (Player X will score 24 points) or probability distributions (60% chance of covering the spread). Most useful when compared against betting lines to identify edge.
Example: Model projection: Team A wins by 6 points. Current spread: Team A -3.5. The 2.5-point difference between projection and line suggests potential value on Team A.
RMSE (Root Mean Squared Error)
A measure of prediction accuracy that calculates the average magnitude of errors. Lower RMSE indicates better predictions. RMSE penalizes large errors more heavily than small ones, making it useful for evaluating point projections.
Example: A model predicting player points with RMSE of 4.2 means predictions are, on average, about 4 points off from actual results. An RMSE of 3.1 would indicate meaningfully better accuracy.
MSE (Mean Squared Error)
The average of squared differences between predicted and actual values. MSE is related to RMSE (RMSE = square root of MSE) and is used in model training to optimize predictions. Like RMSE, lower values indicate better accuracy.
Example: If predictions miss by 3, 4, 2, and 5 points, squared errors are 9, 16, 4, 25. MSE = (9+16+4+25)/4 = 13.5. RMSE = √13.5 = 3.67.
Backtest
Testing a model on historical data it wasn't trained on to evaluate real-world performance. Proper backtesting simulates actual betting conditions—training only on data available before each prediction. Models failing to achieve minimum accuracy thresholds (typically ~55-60% for spread predictions) are not deployed.
Example: A model trained on 2019-2022 NFL data is tested on 2023 games. If it achieves 58% accuracy against the spread in this backtest, it passes validation.
Accuracy
The percentage of predictions that are correct. For binary outcomes (over/under, spread cover), accuracy measures how often the model picks the right side. Higher accuracy is better, but even 55% accuracy can be highly profitable given standard betting odds.
Example: A model correctly predicts 57 out of 100 spread outcomes. Accuracy = 57%. At -110 odds, this produces profit: (57 × $100) - (43 × $110) = $5,700 - $4,730 = +$970.
Calibration
How well a model's probability estimates match actual outcomes. A well-calibrated model's 60% confidence picks should win about 60% of the time. Poor calibration means the model is overconfident or underconfident in its predictions.
Example: If a model assigns 70% probability to 100 different picks and 85 of them win, the model is underconfident—it should have assigned ~85% probability. Calibration adjustments correct for this.
Specificity
In over/under predictions, specificity measures how often the model correctly identifies unders when the actual result is under. High specificity means the model rarely predicts over when the game goes under.
Example: Of 50 games that went under, the model correctly predicted under in 35. Specificity = 35/50 = 70%.
Sensitivity
In over/under predictions, sensitivity measures how often the model correctly identifies overs when the actual result is over. High sensitivity means the model rarely misses overs by predicting under.
Example: Of 60 games that went over, the model correctly predicted over in 42. Sensitivity = 42/60 = 70%.
Confidence Level
An indicator of how strongly models agree and how reliable a prediction is likely to be. High confidence typically requires model consensus, sufficient sample size, and favorable historical accuracy for similar predictions. Not all picks deserve equal weight.
Example: A high-confidence pick might have 5 out of 6 component models agreeing, a historical accuracy of 62% for similar situations, and strong recent model performance. A low-confidence pick might show split models and limited comparable data.
Line Movement
Changes in betting odds between opening and game time. Line movement reflects new information (injuries, weather) and betting action (sharp money, public betting patterns). Tracking how and when lines move can reveal market sentiment and potential value.
Example: A spread opens at -3 and moves to -4.5 by game time. This 1.5-point move suggests significant action on the favorite, possibly from sharp bettors who see value at the original number.
Sharp Money
Bets placed by professional or highly informed bettors. Sportsbooks track sharp action and often adjust lines accordingly. Line movements caused by sharp money tend to be predictive—the 'smart money' is frequently correct.
Example: A total opens at 220 with balanced public betting. A large sharp bet on the under causes the line to drop to 217. This 3-point move on limited public action suggests professional opinion that the total is set too high.
Ready to see these concepts in action?
Get access to our AI-powered picks and see edge calculations, +EV picks, and ensemble projections in your dashboard.
View Plans