How to Project NFL Player Props Without Relying on Hunches

Before modeling, write the wager as a settlement specification: player, statistic, threshold, side, odds, game window, and operator. “Over 64.5 rushing yards” at -110 is not interchangeable with 65+ yards, an alternate line, or a rushing-plus-receiving market.

Confirm the book’s NFL player prop definitions and rules, including:

• Whether overtime counts
• Whether the player must take a snap
• How pushes, voids, stat corrections, and postponements are handled
• Which official data provider governs settlement

Then model that exact event. A 65-yard median does not imply a 50% chance of clearing 64.5 when outcomes are skewed or discrete; the full distribution must map to the posted line and price.

Reconcile house rules with the sportsbook’s data provider before modeling. NFL passing yards and attempts exclude sacks; sack yardage reduces team net passing, not the quarterback’s individual passing total. Official rushing attempts generally include kneels, so accounting for quarterback kneel-downs can materially shift low rushing-yard or attempt lines. Confirm how the feed handles scrambles, aborted plays, and penalties.

Raw season averages are fragile: they mix changing roles, uneven opponents, injuries, and small samples. Start with a weighted baseline that emphasizes games reflecting the player’s current job rather than blindly favoring the most recent box scores.

Blend several signals:

• Role: snaps, routes, backfield participation, alignment, and depth-chart status
• Usage: targets per route, carries per snap, air-yard share, and goal-line work
• Performance: yards, receptions, touchdowns, or other market-specific production
• Opponent adjustment: production relative to what each defense normally allows
• Credible prior: multi-year performance, position-level aging curves, draft capital, or comparable-player rates

Shrink noisy statistics toward the prior. Touchdown rate and yards per target generally require more regression than route participation or target share. Weight games by role similarity and health, then adjust efficiency separately; a backup’s 35% route rate should not be averaged directly with a new 90% role.

Estimate team plays first, using situation-neutral pace, expected game script, opponent pace, weather, and offensive continuity. Next divide plays into dropbacks and designed rushes. Pass rate over expectation is more informative than raw pass rate because it controls for score, down, distance, and win probability.

Allocate that volume through a connected roster model:

1. Project active players and expected limitations.
2. Assign snaps, routes, carries, targets, and red-zone roles.
3. Account for teammate competition and formation constraints.
4. Verify that every share sums to a realistic team total.

Internal consistency is non-negotiable. If an injured receiver loses 12 routes, those routes must transfer elsewhere or disappear through personnel and play-calling changes. Raising a running back’s carry share while also increasing the backup’s workload creates phantom volume unless team rushing attempts rise.

Finish with reconciliation checks: player routes cannot exceed team dropbacks, target shares should sum near 100% after throwaways and spikes, and projected carries must include quarterback scrambles where the sportsbook’s grading requires them.

A projection should move from team environment to player opportunity, then conversion and efficiency. For a receiver, the chain is team dropbacks × route participation × targets per route × catch probability. This structure clarifies why receptions and receiving yards require different assumptions: yards add depth of target, yards after catch, and game-state effects after the catch is secured.

Use a prop-specific chain:

• Receptions: routes, target rate, catch probability, and defensive coverage tendencies.
• Receiving yards: targets, catch rate, air-yards distribution, and yards after catch—not merely historical yards per reception.
• Rushing yards: team rush volume, backfield share, quarterback scramble effects, box counts, and yards gained conditional on carry type.
• Touchdowns: projected team scoring opportunities, red-zone usage, goal-line role, and conversion rate.
• Longest play: opportunity count plus the full gain distribution, especially its right tail.

Regression should strengthen as outcomes become less repeatable. Route participation and carry share generally deserve modest shrinkage; touchdown conversion and breakaway rates deserve much more. Longest-play projections should use a distribution or simulation rather than average efficiency, because one rare gain determines settlement. The final mean is less informative than the probability of crossing the sportsbook’s exact line.

A matchup modifier needs a traceable path to the prop. Stat-based NFL matchup analysis should identify how an opponent changes opportunity or efficiency—not merely label the defense strong or weak.

• Pressure: Adjust sack, scramble, time-to-throw, and completion assumptions according to protection injuries and pressure tendencies.
• Coverage: Use man-zone rates, safety structure, and likely shadow assignments to alter target distribution or depth—not every receiver equally.
• Box counts: Connect personnel and defensive fronts to rushing efficiency and play selection.
• Weather and venue: Meaningful wind can suppress deep passing and kicking; rain alone usually warrants little movement. Dome effects are generally small.
• Injuries and coaching: Translate absences and tactical changes into routes, snaps, touches, or team pace.

Totals and spreads can inform projected plays and pass rate, but the adjustment should remain modest because those markets already reflect many underlying inputs. Apply the change once: if a low total has reduced team touchdowns, avoid separately penalizing every scorer for the same environment.

A central projection is only the distribution’s balance point. Two players projected for 64.5 receiving yards can have very different chances of beating 63.5: a short-area receiver may cluster near the mean, while a deep threat produces more low totals and a longer right tail.

Choose a distribution that respects the statistic:

• Counts such as receptions, attempts, and touchdowns often require a negative-binomial or beta-binomial model rather than Poisson, because role and game-to-game variance create overdispersion.
• Yardage is better represented as a compound process: opportunity count multiplied by variable gain per opportunity. Lognormal, gamma, or empirical residual distributions can preserve asymmetric upside.
• Playing time should be modeled as a mixture. A normal-role state, limited-role state, and early-exit state capture benching, aggravation, and conditioning uncertainty more honestly than a single injury haircut.

Volume, efficiency, and script should not be sampled independently. A shared latent game environment can jointly move team plays, pass rate, target share, and yards per target. This preserves outcomes such as trailing teams generating more attempts but lower efficiency, or positive rushing scripts increasing carries while shortening the game.

Finally, calculate probabilities from the discrete settlement line itself. For a line of 65 receiving yards:

• Over: P(Y ≥ 66)
• Push: P(Y = 65)
• Under: P(Y ≤ 64)

The mean does not identify any of these probabilities. Tail shape, variance, integer mass, and mixture weights determine the fair price—and push probability must be separated before comparing it with sportsbook odds.

A projected hit rate becomes actionable only after conversion to a fair price. With no push probability, fair decimal odds are 1 / p. Fair American odds are -100 × p / (1 − p) when p > 50%, or +100 × (1 − p) / p otherwise.

For integer lines, remove pushes before pricing: use the conditional win rate p(win) / [p(win) + p(loss)]. A 52% win probability with an 8% push rate is effectively 56.5%, implying fair odds near -130, not -108.

Evaluate the line and juice as one instrument. Over 74.5 at -115 cannot be compared directly with over 75.5 at +105; each threshold requires its own win, loss, and push probabilities. Expected value per unit risked is the cleanest comparison.

Require a margin beyond fair value—often several percentage points or a meaningful EV threshold—to absorb parameter and distribution error. Market prices remain useful diagnostics: a large disagreement should trigger checks for injuries, grading rules, and stale inputs, but should not force the projection toward consensus.