Beyond the Box Score
How we turn raw game logs into calibrated probability estimates and find value where the market has mispriced player performance.
Think of how a stock analyst values a company. They do not just look at yesterday's price and guess tomorrow's. They ingest earnings reports, sector trends, competitive dynamics, and historical patterns, then build a model to produce a probability-weighted valuation: “65% chance the stock is underpriced.” Our NBA props model works the same way, except the asset is a player's stat line and the question is whether it goes over or under the book's number.
This page explains how the model actually thinks, from raw data to the picks that surface in your terminal. We will keep it honest: you will understand exactly what powers the system without us handing out the recipe.
The Challenge
Why NBA Props Are Uniquely Tricky
Imagine trying to predict how many emails your coworker will send tomorrow. You know their daily average, but that average hides enormous variation: some days they are heads-down coding, other days back-to-back meetings. Now multiply that uncertainty across every NBA player, every stat, every night. That is the player props market.
Game-to-Game Noise
A star can score 40 one night and 18 the next. Blowouts, foul trouble, and hot streaks create massive variance even for elite players.
Context Dependence
The same player facing the league's best defense on a back-to-back is a completely different proposition than that player rested against the worst defense.
Efficient Market
Sportsbooks have their own models and sharp bettors moving their lines. The obvious edges get priced out quickly. Subtlety wins.
The opportunity: Most bettors and even many models treat all players the same: generic assumptions, surface-level averages. When our model accounts for individual volatility, real matchup dynamics, and calibrated probabilities, the mispriced lines reveal themselves.
The Projection Engine
Beyond the Box Score
Imagine you are diagnosing why a car is slow. You would not just look at one gauge. You would check the engine RPM against the speed, the fuel pressure against the throttle, and the temperature against the load. Each comparison tells you something the raw numbers alone cannot. Our model works the same way: it engineers specific comparisons that expose where the market has mispriced a player.
The model computes purpose-built features for each player prop, each designed to capture a distinct edge signal. These features feed into a trained machine learning model that has learned which combinations predict OVER hits.
| Signal Category | What It Captures |
|---|---|
| Workload Analysis | Detects shifts in player usage that the book's line has not yet absorbed. |
| Matchup Context | Evaluates how tonight's specific opponent environment affects the player's expected output. |
| Schedule & Fatigue | Quantifies rest and fatigue dynamics as direct model inputs rather than manual adjustments. |
| Game Environment | Captures broader game-level conditions that influence overall tempo and opportunity volume. |
| Market Baseline | Incorporates the market's own pricing as an anchor, ensuring the model finds value relative to efficient odds. |
Key insight: Raw stats are noisy. Engineered features isolate the specific comparisons that predict outcomes. The model looks beyond season averages to find the dimensions where the market consistently misprices players.
The Feature Pipeline
Each feature captures a different dimension of the prediction problem. They act as independent scouts, each watching the game from a different angle. The model learns how to weigh their reports together to produce a single probability estimate.
Workload Analysis
The model detects when a player's recent usage has shifted in ways the book's line has not yet reflected. These workload signals are among the most predictive inputs in the system.
Usage changes often precede line adjustments by hours or days, creating a window of opportunity.
Matchup Context
Not all opponents are created equal. The model evaluates how tonight's specific defensive environment affects the player's expected output, isolating true opponent impact from raw averages.
This captures what flat season averages consistently miss.
Schedule & Fatigue
The model quantifies rest and fatigue dynamics as direct inputs, letting it learn the precise impact from training data rather than applying manual adjustments. This removes guesswork from a factor that meaningfully affects player output.
Game Environment
Broader game-level conditions influence total opportunity volume. The model captures these dynamics as context features, learning how they interact with other signals rather than applying flat multipliers.
Styles Make Stats
The oldest truth in sports analytics: not all numbers are created equal. Two players can both average 22 points per game and be completely different bets. The raw average hides the context that actually matters: playing time trends, matchup quality, schedule demands, and market efficiency.
Why Most Prop Models Get This Wrong
Picture two archers. Both shoot at a target 3 inches to the right of center. Archer A groups every shot within a 1-inch circle. Archer B sprays arrows across a 3-foot spread. For Archer A, being 3 inches off-center is a meaningful, correctable bias. For Archer B, it is noise lost in the chaos. The same principle applies to player props.
Most models treat every player the same way: they compute an average, compare it to the line, and call it a day. They miss the comparisons that actually predict outcomes: Is the player getting more minutes than the line implies? Is the opponent's defense creating a favorable matchup? Is fatigue a factor tonight?
Our model does not just project a number. It engineers proprietary features that capture the dimensions where the market tends to misprice players, then lets a trained machine learning model learn the optimal weighting.
Learned Weights, Not Manual Rules
Traditional models use hand-tuned multipliers and rules of thumb. Our model learns the optimal relationship between each feature and the outcome from thousands of historical props. The weights are not guesses; they are the result of training on real data with proper cross-validation.
The Fifth Feature: Market Itself
The market's de-vigged implied probability is itself a feature in the model. This is deliberate: the market is smart, and ignoring it would be arrogant. By including the market's own estimate as an input, the model learns to find value relative to what the market already knows, not in a vacuum.
| Feature Role | What It Does |
|---|---|
| Primary Signals | Where the model finds mispricing — the features most likely to diverge from what the book has priced |
| Context Signals | Game conditions and situational factors that modify the strength of the primary edge |
| Market Anchor | Prevents the model from ignoring efficient pricing — value is found relative to the market, not in a vacuum |
Finding Value
Where Edges Come From
Sportsbooks are not setting NBA prop lines in a vacuum. They have their own models, sharp bettors moving their lines, and oceans of data. The market is efficient, but not perfectly efficient. We look for specific situations where the market tends to misprice players:
- Workload mispricing: A player's recent usage has shifted but the book's line still reflects the old patterns. Our model catches these changes before the market adjusts.
- Matchup-driven mispricing: The book line reflects the player's average, but tonight's opponent creates a materially different environment. Our model captures what flat averages miss.
- Situational mispricing: Schedule dynamics and fatigue create measurable drags on output that the model quantifies directly, not as a guess.
- Model vs. market divergence: When the trained model's predicted probability diverges meaningfully from the market's de-vigged probability, that gap is the edge.
Model vs. Market
The model outputs a probability for each prop. We strip the bookmaker's margin from both sides of the line to get a fair implied probability. The gap between the model's P(OVER) and the fair probability is the edge. If a book prices a line at -110 each side, the de-vigged fair probability is 50.0%. If our model says the true probability is 62.8%, the edge is +12.8%.
Important: Not every edge is worth taking. The model enforces a minimum edge threshold configured from production experience. Only props where the model sees a meaningful probability advantage above the market baseline are surfaced.
The Confidence System
Multi-Gate Quality Control
Every potential pick must clear multiple independent checkpoints before it reaches your terminal. If it fails any single gate, it is rejected. Failures are final.
Gate 1: Minimum Edge
The model's predicted probability must exceed the market's fair probability by a configurable threshold. Marginal edges get rejected outright.
Gate 2: Odds Range
Only props within a specific odds band are considered. Extreme favorites and longshots are excluded because the risk-reward ratio becomes unfavorable.
Gate 3: Injury Screen
Players flagged with injury concerns are excluded. Uncertainty around playing time or effectiveness makes the prediction unreliable.
Gate 4: Deduplication
Only one pick per player per day. If multiple lines qualify, the highest-edge option wins. This prevents overexposure to a single player outcome.
Edge-Ranked, Flat-Sized
Picks that survive all gates are ranked purely by edge: the gap between the model's predicted probability and the market's fair probability. The top candidates by edge are surfaced each day, capped at a maximum to maintain quality over quantity.
| Metric | How It Works |
|---|---|
| Edge | Model probability minus fair (de-vigged) probability. Higher edge = stronger signal. |
| EV | Expected value calculated from the model's probability and the actual odds offered. |
| Units | Flat 1.0 unit per pick. Consistent sizing removes emotion and overconfidence from the equation. |
Important: Individual picks still lose. NBA props are inherently volatile. The edge-ranking system is about expected value over many bets, not guarantees on individual picks. Flat sizing ensures no single loss can overshadow the portfolio.
How We Measure Success
Anyone can cherry-pick a hot streak. We believe in full transparency across large sample sizes. Here is what we track and why it matters:
Hit Rate
Every pick is tracked against its predicted probability. We monitor overall hit rate to validate that the model's edge estimates translate to real-world results.
Return on Investment
Win rate alone can mislead. ROI accounts for the odds on each bet, showing actual profit relative to amount wagered.
Closing Line Value
Did we get better odds than the final line? Consistent CLV proves we are identifying real edges, not getting lucky.
Sample Size
Results over 20 picks mean little. Results over hundreds of picks reveal true model performance through the noise.
Closing Line Value (CLV)
This is the gold standard. CLV measures whether the line moved in your direction after you locked in your pick. If you bet a player Over 24.5 points and the line closes at Over 26.5, the market confirmed your read. You got a better price than the final consensus.
Sportsbooks use CLV to identify their sharpest customers. Consistent positive CLV is the single strongest predictor of long-term profitability, because it means you are consistently ahead of the market, not just getting lucky.
How our picks are timed:
Picks lock 10:00 AM ET daily. This gives you time to place bets before lines move, maximizing CLV capture.
The Bottom Line
NBA props reward those who can see past the season averages, the name recognition, and the narratives. Our model does the heavy lifting of analyzing thousands of data points across every player's season, but the philosophy is simple:
- Engineer features, not averages: purpose-built signals that expose mispricing the market has missed
- Learn from data: a trained machine learning model that weighs each signal based on historical predictive power
- Filter ruthlessly: through multiple independent gates. Only genuine edges survive
- Size consistently: flat unit sizing removes emotion and keeps the portfolio disciplined
- Track everything: so you can evaluate us with full transparency
The box score is noise. Our job is to find the signal within it.
Technical Breakdown
For the quantitative readers: a deeper look at the architecture and algorithms powering our NBA prop predictions.
Prediction Engine
The core prediction engine is a trained machine learning model. Engineered features are normalized and combined to produce P(OVER), the probability the player goes over the line. The architecture is optimized for interpretability and speed.
The model was trained on historical NBA props data with walk-forward validation. Model parameters are stored as artifacts and loaded at runtime, not retrained on the fly.
Data Pipeline
Player statistics are sourced from current-season game logs via proprietary data feeds. Every projection run fetches fresh data and computes the engineered features that power each prediction.
Real-time odds are aggregated from 30+ licensed sportsbooks via our real-time odds feed. Both sides (over/under) are paired for each line to enable proper de-vigging and fair probability calculation.
Edge Calculation
Edge is the difference between the model's predicted probability and the market's fair (de-vigged) probability. Fair probability is computed by removing the bookmaker's margin from both sides of the line. Only picks where the model sees a probability advantage above the configured threshold are surfaced.
Expected Value (EV) is calculated from the model probability and the actual American odds offered. This accounts for the real payout structure, not just the probability gap.
Model Workflow
5:00 AM ET: Settlement: previous day's picks graded against actual results
9:00 AM ET: Projection: full pipeline runs (stats, odds, feature engineering, prediction, filtering)
10:00 AM ET: Lock: top picks ranked by edge and surfaced to subscribers
Every 30 min: Odds sync and CLV tracking throughout the day
Monthly Access
- Predictions only go live when the model finds true edge
- Closing line value tracked on every prediction so you can verify it yourself
- Covers every market we model and we're always adding more
- Cheaper than your average unit size
Annual Access
- Get 4 months free on us when you go annual
- Every new model we ship is included automatically
- Full platform access for less than most services charge monthly
- Models run 365 days, your subscription should too