Methodology Deep Dive
How Predictium Projects Player Props
A 5-layer, data-driven projection model that accounts for game context, opponent matchups, pace, and teammate absences — built on the same infrastructure as our game predictions.
← Back to Props OverviewThe Problem with Most Prop Models
Why simple averages fail
Most prop tools start with a player's season average and maybe adjust for recent form. Some aggregate projections from multiple sources into a consensus number. Both approaches have the same blind spots:
- No game context. A game with a 15-point spread plays differently than a 1-point spread. Starters sit earlier in blowouts, bench players get extended run. Season averages don't know this.
- No stat-level opponent adjustment. A team might be average defensively overall but elite at limiting three-point shooting. Generic "defense rating" adjustments miss these matchup edges.
- No role change modeling. When a team's primary ball handler sits, every teammate's role shifts — but not every stat shifts the same way. A guard doesn't suddenly grab a center's rebounds just because the center is out.
The result: stale projections that miss situational edges. The exact situations where sharp bettors find value.
Our 5-Layer Approach
Each layer refines the projection
Our prop model is a pipeline. Each layer takes the output of the previous layer and adds context. Nothing is a black box — every adjustment has a clear rationale.
Per-Minute Rates
We compute per-minute production rates for every stat category — points, rebounds, assists, 3PM, steals, blocks. These rates use exponentially weighted moving averages (EWM) with a half-life of roughly 10 games, so recent performance carries more weight while maintaining enough sample to avoid noise. A player who's been scoring efficiently over the last two weeks matters more than their October averages.
Projected Minutes
Minutes are sourced from RotoWire and adjusted for game context. The key adjustment is blowout minutes regression: if the predicted spread is large, starters' minutes come down and bench players' minutes go up. A starter on a -15 favorite might play ~28 minutes instead of their usual 34. This is a direct consequence of having an integrated game prediction model — we know the spread before we project the props.
Pace Adjustment
Both teams' pace ratings affect volume. A fast-paced game means more possessions, more shots, more counting stats across the board. We use our game prediction model's pace estimates — not some independent lookup — so the pace adjustment is consistent with the projected total and game flow. A game projected at 230+ total plays differently for prop purposes than one projected at 205.
Opponent Adjustment
Stat-specific defensive profiles make a real difference. We don't use a single "defensive rating" number — we look at how each team defends each stat category independently. If Cleveland allows the 2nd-fewest 3PM per game, your shooter's three-point projection goes down. If Indiana allows the most assists in the league, the opposing point guard's assist projection goes up. These opponent factors are computed from our team feature matrix and updated daily.
Smart Redistribution
When a key player is injured or resting, their production doesn't just disappear — but it doesn't all stay on the same team either. How we handle redistribution depends on the stat:
Points, Assists, 3PM → Redistribute by usage rate (USG%)
Ball-dominant players absorb more of the scoring and playmaking load. A player with 28% usage picks up a larger share than one with 15%.
Rebounds, Blocks, Steals → No usage redistribution
These stats don't scale with ball-handling usage. A guard doesn't grab a center's rebounds just because the center is out. Some of those rebounds go to the other team entirely. We use baseline per-minute rates instead.
Why this distinction matters
A naive model would inflate Jalen Brunson's rebound projection when Karl-Anthony Towns sits. We don't. Rebounds, blocks, and steals are physical-presence stats that don't follow the ball. When a center is out, some of his rebounds go to other bigs on the roster, and some go to the opposing team. Treating all stats like scoring stats is one of the most common mistakes in prop modeling.
What We Don't Do
Intentional design choices
Integration with Game Predictions
One consistent system
Most prop tools operate in isolation — they have their own pace numbers, their own opponent ratings, their own sense of how a game will play out. Our prop model is integrated with our B6 game prediction engine. Everything flows from the same data:
- Pace estimates from the game model feed directly into prop projections — so a game we project as fast-paced produces higher-volume prop numbers
- Predicted spread drives blowout minutes regression — if we think the game is a blowout, starters' minutes come down automatically
- Opponent defensive profiles are shared across both systems — the same team strength data that powers game predictions also adjusts individual stat lines
- RAPM data informs both player impact on game outcomes and individual stat projections
The result: if we project a high-scoring, fast-paced game, the prop projections for players in that game reflect it. No contradictions, no inconsistencies.
Coming Soon
On our roadmap
With/Without Teammate Splits
Actual performance data for when key players sit — not just modeled redistribution, but real historical with/without numbers to validate and sharpen projections.
Stat-Specific Redistribution Weighting
Per-minute rebound rate, block rate, and other positional stat weights to further improve how we redistribute production when players are absent.
Historical Prop Accuracy Tracking
Transparent tracking of how our projections compare to actual outcomes — by stat, by player tier, by game context. Full accountability.
Access Player Props & Fantasy Projections
7-day free trial. Full access to prop projections, downloadable fantasy data, and everything else. No credit card required.
Start Free Trial