Methodology

No single model solves the NBA. We run five independent methodologies in parallel and surface where they disagree.

When all five models align, the market has usually priced it in. The value exists in the divergence — when one model sees something the others don't. We surface the conflict, not the consensus.

Normalization, confidence scoring, regime classification, and accountability.

The engine scans every game for situations where context diverges from what the market expects. A key rotation player ruled out two hours before tip. A team on the back end of a road trip facing a rested opponent. A spread bucket where the covering team has historically dominated.

We surface the patterns and the data behind them. You decide if there's edge.

A player averaging 18 points might be doing it in 32 minutes one night and 22 the next. We normalize performance per-36 minutes so a starter coasting in a blowout doesn't skew the data.

When a player is absent, usage redistributes. We measure the actual historical shift in each teammate's per-36 production across every relevant game — not estimates, measurements.

Small samples mislead. When evidence is thin, projections anchor to the player's established baseline. As games accumulate, the observed data takes over. This is Bayesian smoothing — it prevents volatile early-season signals from distorting projections.

Every projection includes a floor and ceiling: the range of probable outcomes given the volatility in both groups (with the player and without). A tight range is a strong pattern. A wide range means uncertainty remains.

Not all matchups have the same statistical character. The engine classifies each game into regimes based on recent performance variance:

Stable — tight performance clusters on both sides. Projections carry weight.

Volatile — wide variance on one or both sides. Error bars expand.

Regime classification feeds directly into confidence scoring. Volatile matchups widen floor-ceiling ranges and temper confidence badges. A HIGH signal in a Stable game carries more weight than the same signal in a Volatile one.

We also cross-reference against line movement from 39 sportsbooks. When the market diverges from what the stats suggest, that's worth investigating.

The engine generates assertions — timestamped, machine-generated claims locked before tip-off — and grades every one against actual box scores the next morning.

The pipeline: Gating filters signals through minimum hit rate and sample size thresholds. Assertion records the exact claim, threshold, and market odds. Grading compares against the official box score. CLV measures whether we captured value versus the closing line.

Hit rates use Wilson score confidence intervals, not raw percentages. Going 7/10 feels like 70%, but the 95% CI is [38%, 90%]. Wilson intervals account for sample size. As assertions accumulate, the intervals tighten and the true signal becomes visible.

NBA teams submit injury reports by 5:30 PM ET. Our engine runs at 6:00 PM ET after reports finalize. We filter for meaningful transitions: a player ruled OUT, upgraded from a worse status, or returning after missed time. Routine PROBABLE designations are excluded.

Each alert includes beneficiary impact: which teammates have historically absorbed the usage, with per-36 normalized differentials, Bayesian-smoothed confidence, and current prop lines for context.

Impact baselines use temporal decay weighting with a 60-day half-life. Recent form is weighted over season averages. Both “with” and “without” game groups use identical decay, isolating the absence effect from mid-season trends.

The Lab lets you build custom prop queries with historical hit rates. Query Receipts close the loop: did the pattern hold going forward?

When you save a query, the system snapshots your parameters, the targeted players, their current prop lines, and the applicable games. After those games complete, the resolver grades every player against their actual box score.

Over means the stat exceeded the line. Under means it fell below. Exact ties push (excluded from hit rate). Players who didn't play are voided. Resolution runs at 4 AM ET after box scores finalize.