How DMP Works
Under the hood: projections, EV, and AI insights
DMP doesn't just show you odds. It builds projections from real game data, calculates expected value against the market, and uses AI to assess whether edges are durable or fragile. Here's how each piece works — and how it adapts by sport.
Projections
Every prop you see on DMP has a projection behind it — a statistical estimate of what a player is likely to produce in their upcoming game. Projections power two things: the PROJ column on the odds screen and the P(over) probability used for model-based EV.
Stats We Project
Combo stats (PRA, PR, PA, RA) use covariance modeling — they account for how stats correlate within the same game, not just simple addition.
1Weighted Baseline
DMP starts with a player's real game logs and computes per-100-possession rates for each stat across three time windows. This normalizes for pace — a player in a fast game and a slow game are compared apples-to-apples.
NBA Weights
- Last 5 games: 50%
- Last 10 games: 30%
- Season: 20%
Recent form weighted heavily — NBA roles shift fast.
NCAAB Weights
- Last 5 games: 30%
- Last 10 games: 30%
- Season: 40%
Season weighted more — college samples are smaller, so we lean on the larger window.
2Matchup Adjustment (DVP)
Not all defenses are created equal. DMP adjusts projections based on how the opponent defends players like this one — what we call Defense vs. Position (DVP).
NBA: 10-Archetype System
Players are classified into archetypes based on usage, assist rate, and play style — like primary ball handler, stretch big, or rim runner.
DVP is calculated per-archetype, per-stat. A team might be tough on scoring wings but soft against stretch bigs. Each stat gets its own softness score (0-100 percentile).
NCAAB: 3-Bucket Positions
College basketball uses a simpler model: Guard, Forward, and Center.
With less data per team, fine-grained archetypes would overfit. The 3-bucket approach is more reliable with smaller sample sizes, and uses per-possession rates to avoid double-counting pace.
3Context Adjustments
After the matchup adjustment, several contextual factors are applied:
Pace
Faster-paced games create more possessions and more statistical opportunities. DMP uses actual possession data when available, or derives pace from Vegas totals.
Home Court
NBA: +2% home boost. NCAAB: +3.5% home boost (college crowds have more impact). NCAAB also handles neutral sites (tournaments) where no boost is applied.
Injury Redistribution NBA only
When a teammate is OUT, their statistical opportunity flows to remaining players. DMP models this using archetype-based redistribution rules — a primary ball handler going out boosts combo guards differently than it boosts rim runners.
Each injured player has an impact tier (A through D) based on their minutes and usage. A Tier A star (like a 30+ MPG, high-usage player) missing redistributes far more than a Tier D bench player.
NCAAB-Specific Adjustments
- Blowout risk: If the Vegas spread is 10+, starters rest earlier in a 40-minute game. Projections are compressed accordingly.
- Team shooting scheme: If a team runs a perimeter-heavy offense, 3-pointer projections are adjusted up (and vice versa).
- Defensive pressure: High-pressure defenses (lots of steals and forced turnovers) suppress assist production for ball handlers.
- Foul trouble variance: College has a 5-foul limit (vs NBA's 6). Players averaging 2.5+ fouls per game get wider variance on their projection.
4Variance and P(over)
A projection alone isn't enough — you need to know how confident that projection is. DMP computes variance (how spread out a player's outcomes are) from their recent game-to-game consistency.
Using the projection and variance together, DMP calculates P(over) — the probability of going over the betting line — using a statistical model (normal distribution).
Example: A player projected for 22.5 points with low variance against a line of 20.5 might have P(over) = 72%. The same projection with high variance might only be P(over) = 60%.
5Shrinkage (Regression to Mean)
After all adjustments, DMP blends the adjusted projection (70%) with the player's season average (30%). This prevents the model from overreacting to a hot or cold streak and keeps projections grounded when context adjustments stack aggressively.
Expected Value (EV)
Expected value tells you whether a bet is priced in your favor. DMP calculates EV differently for NBA and NCAAB because the data available to price each market is fundamentally different.
Sharp Book Consensus
For NBA props, sportsbooks like Pinnacle, Circa, and Bookmaker post lines with tight margins. These "sharp" books are considered the closest reflection of true probability.
How it works:
- DMP collects odds from 5 sharp books (FanDuel, Caesars, Pinnacle, Circa, Bookmaker)
- Each book's odds are devigged — the house edge is removed to extract the implied true probability
- A consensus probability curve is built from all sharp book lines
- For each prop at each sportsbook, EV is calculated: does the offered price exceed the consensus fair price?
The +EV% you see for NBA props represents how much the offered odds exceed the sharp book consensus. This is a pure market-based approach — it trusts that sharp books have priced the true probability correctly.
Model + Market Blend
Sharp books rarely post college basketball player props. Without sharp lines to anchor a consensus, DMP uses a dual approach that combines its own model with the retail market.
How it works:
- A consensus is built from 14 retail books (FanDuel, DraftKings, BetMGM, Caesars, Bet365, and others)
- Each book is evaluated using leave-one-out — its own line is excluded from the consensus when calculating its EV, preventing self-reference bias
- DMP's projection model independently calculates P(hit) from the player's projected stats and variance
- The final EV blends: 60% model probability + 40% market consensus
When market data exists (2+ books)
Blended EV gives you both the model's view and the market's view, weighted toward the model. If they disagree by more than 8%, the EV chip shows an amber warning.
When only 1 book posts a line
No consensus is possible. EV falls back to pure model EV — 100% from DMP's projection. These are higher-conviction but unconfirmed by the market.
Why blend? Retail books still carry signal (real money is wagered), but they're less precise than sharp books. The model fills the gap with statistical projections.
The EV Formula
Regardless of how the probability is determined (sharp consensus or model blend), the EV formula is the same:
A positive EV means the offered odds imply a probability lower than the true probability. Over enough bets, positive EV plays are expected to profit.
AI Insights
Projections and EV tell you whether a bet has mathematical value. AI Insights answers a different question: is today an outlier game or a typical game? It evaluates whether the statistical edge is durable (likely to persist) or fragile (likely to evaporate due to context).
Six Factors Evaluated
Role & Minutes
Are minutes locked in or volatile? Is usage trending up or down? High stability = durable edge.
Game Script
Vegas total and spread predict pace and blowout risk. High totals favor OVER; large spreads threaten early benchings.
Matchup (DVP)
Stat-specific defensive data. A team can be tough overall but soft on a specific stat — the AI evaluates each separately.
Injuries
Tier-weighted injury impact. Not just "who's out" — the AI considers whether the absence is new (opportunity) or established (already priced in).
Line Movement
Is the line moving with or against you? Movement direction signals whether sharp money or soft action is driving the price.
Model Output
The projection, P(over), fair odds, and market classification provide the statistical baseline the AI evaluates against.
League-Specific Context
What the AI considers
- Archetype DVP with per-stat softness: Separate percentiles for points, rebounds, assists, and 3PM against the specific archetype
- Injury tiers (A/B/C/D): Star absences (Tier A) trigger major redistribution analysis. Bench players (Tier D) are flagged as negligible.
- New vs established absences: A player out 10+ games is "already priced in" — the AI won't treat it as a fresh edge
- Opponent defender injuries: If an elite perimeter defender is out, the AI boosts the matchup softness assessment
- Game theory scenarios: If a star is QUESTIONABLE, the AI models both outcomes (plays vs sits) and their impact
What the AI considers
- 40-minute game context: Stats are ~17% lower than NBA. Per-40 normalization is used for comparison.
- Team shooting scheme: Whether the team is perimeter-heavy (40%+ of shots from 3) or inside-oriented — directly impacts 3PM ceilings
- Opponent defensive pressure: High-pressure defenses (70th+ percentile in steals/forced TOs) suppress assists for ball handlers
- Foul risk: 5-foul limit means a player averaging 2.5+ fouls/game has real minutes variance
- Conference context: Conference games are well-scouted (DVP more reliable). Non-conference games carry more uncertainty.
- Neutral sites: Tournament games have no home boost — affects all player projections
- Player role classification: Primary Ball Handler, Key Creator, Key Scorer, Rotation, or Role Player — derived from position, tier, and usage
What the AI Prioritizes by Prop Type
Different stats are driven by different factors. The AI evaluates each prop type through its own lens:
Points
NBA: Pace, DVP archetype, usage trend, blowout risk. NCAAB: Tempo is the #1 driver (70+ possessions = high ceiling, <64 = compressed), then per-40 rate, DVP, foul risk.
Rebounds
Position stability, opponent FG% (more misses = more boards), game script (competitive games keep starters in), pace. Blowouts compress rebound opportunities as benches come in.
Assists
Ball-handler role and usage rate are primary. NCAAB: Opponent defensive pressure is key for handlers — disruptive defenses force turnovers instead of assists. Teammate shooting % also matters (better shooters = more assist opportunities).
3-Pointers Made
Higher variance than other stats. The AI flags when L10 average is above the line but hit rate is below 50% — a sign that one or two hot games are inflating the mean. NCAAB: Team shooting scheme is the primary driver (attempt ceiling matters more than accuracy).
Combos (PRA, PR, PA, RA)
Highly minutes-dependent — blowout risk is the #1 threat. Pace matters most since more possessions lift all component stats. The AI also considers stat correlation (playmakers have correlated PTS+AST).
The Verdict System
After evaluating all six factors, the AI delivers one of five verdicts:
All factors aligned. Edge is structural, not situational.
Strong structural support with one minor uncertainty.
The thesis is sound but you need better odds. Look for a better line elsewhere.
Fundamentals are strong but a key variable (injury status, starting lineup) is unresolved.
The thesis itself is weak — the edge exists on paper but context undermines it.
Recommended Bets
DMP scans thousands of player props every few minutes and surfaces a short list of the most promising bets using 14 deterministic signals. These are designed to approximate the same contextual judgment that AI Insights provides, but computed at scale without any LLM calls.
How Bets Are Selected
Signal Scoring
Each prop is scored across 14 signals — market-derived signals (EV edge, sharp book divergence, price advantage, line movement), projection signals (projection edge, DVP softness, pace, stability), and context signals (hit rate consistency, game script sensitivity, role trend, injury freshness, and factor alignment). Each signal is percentile-ranked against all other props in the same slate.
Composite Score
Signals are combined into a weighted composite score (0-100). Market edge signals (EV, sharp divergence, projection edge) carry the highest weights. Context signals act as tiebreakers and quality filters — they help distinguish between a bet that looks good on paper vs one where the context actually supports it.
Qualification
Props must pass minimum thresholds: composite score of 65+, positive EV (with variance-adjusted minimums per market type), coverage from 3+ sportsbooks including FanDuel, and confirmed game lineups (GREEN status). High-variance markets like 3-pointers require a higher EV floor than core stats like points.
Scoring vs AI Insights
The deterministic scoring pipeline and AI Insights use the same underlying data, but they reason about it differently:
Scoring Pipeline (14 signals)
- Runs every few minutes across thousands of props
- Deterministic, reproducible computation
- Quantifies each factor as a 0-100 percentile
- Cannot assess whether today breaks from historical patterns
- Cannot prioritize conflicting signals (weights are fixed)
AI Insights (Gemini)
- Runs per-prop when you open a research card
- LLM-based qualitative reasoning
- Can judge "this game may not follow historic patterns"
- Dynamically weighs which factors matter most today
- Produces a verdict (VALUE/LEAN/SHOP/WAIT/PASS) with reasoning
Think of recommended bets as a curated shortlist — props where the numbers line up across many dimensions. The best workflow is to start from recommended bets and click through to read what AI Insights says about each one before placing a bet.
Performance Tracking
Every recommended bet is tracked after the game completes. We measure:
Hit Rate
Did the bet win? Tracked by composite score tier, market type, and EV bucket.
Closing Line Value
Did we get a better price than the market's closing number? The strongest indicator of long-term edge.
Signal Effectiveness
Which of the 14 signals actually correlate with positive CLV? This feedback loop tunes the model.
Results are published on the Results page so you can see exactly how recommended bets perform over time — including what happens if you simply take every recommendation without doing additional research.
NBA vs NCAAB at a Glance
| Aspect | NBA | NCAAB |
|---|---|---|
| Game Length | 48 minutes | 40 minutes |
| Matchup System | 10 archetypes, per-stat softness | 3 position buckets (G/F/C) |
| Season Weighting | 20% (more recent-heavy) | 40% (trusts larger sample) |
| Home Boost | +2% | +3.5% (neutral sites: 0%) |
| Injuries | Full tier system, redistribution | Not tracked (no reliable data) |
| EV Source | 5 sharp books consensus | 60% model + 40% retail consensus |
| Leave-One-Out | Not needed (sharp books reliable) | Yes (prevents self-reference bias) |
| Foul Risk | 6 fouls (less impactful) | 5 fouls (inflates variance) |
| Blowout Adjustment | Standard Q4 bench risk | Aggressive (spread 10+ compresses) |
| AI Insights Focus | Injuries, redistribution, defender matchups | Tempo, pressure, scheme, foul risk |
Key Takeaway
NBA and NCAAB use the same projection framework (weighted baselines, DVP adjustments, variance modeling), but each is tuned to the realities of its league. NBA has richer data (archetypes, injuries, defense profiles), so the system is more granular. NCAAB compensates for thinner data with heavier season weighting, model-based EV, and context-specific adjustments like foul risk and defensive pressure.
The numbers on the odds screen — PROJ, +EV%, and AI Insights — each answer a different question. The projection says "what should happen."EV says "is this bet priced in your favor."AI Insights says "will today's context let it happen."