Visor

Pick predictor stats, pick a response stat, and run a regression against actual box scores. Mix player stats and team stats freely, predict a team's win total from one player's receiving yards, or a player's fantasy output from the team's pass rate.

...

Player games

...

Team games

...

Seasons Covered

Model Types

Filters

Dataset

Single Dataset

Compare Two

Team

Narrows every dataset to one franchise's games.

Player Lookup

Pick a name to model that player's own games only.

Position

Season Type

Seasons

Weeks

Sample

...

Linear Regression

What is Linear Regression?

Starter Examples

Not sure where to start? Pick one and it'll set everything up and run it for you.

Predictor variable (choose 1)

Response variable (choose 1)

Advanced Options (optional extra diagnostics; see the Glossary tab for what these mean)

Show prediction range around the line

Adds a shaded zone around the line showing where about 95% of individual games are expected to land, not just the average. A wide zone means games vary a lot at that point, a narrow zone means they tend to cluster close to the line.

Show which predictor matters most

Ranks each predictor by how much it actually swings the result, so you can see at a glance which stat is doing the most work.

Check if it holds up on new games (train/test)

Tests the model on games it has never seen, to check whether it really learned something useful or just memorized this particular set of games.

Test if the whole model is meaningful (F test)

Checks whether your predictors, taken together, actually explain something real, rather than just judging each one by itself.

Adjustable yes/no cutoff + confusion matrix

Lets you slide the yes/no cutoff up or down and watch how many right and wrong guesses the model makes at each setting.

Show ROC curve & AUC

Draws a curve and a single score showing how well the model can tell the two outcomes apart, across every possible cutoff, not just 50%.

Select at least one predictor and a response variable.

Model Output

In Plain English

Fit

R²:

...

R squared

...

Adj. R squared

...

Observations

...

RMSE

Coefficients

Variable	Coef.	Std. Err.	t stat	p value

Model output will appear here once you run it.

Run History

A reference, written in plain language, for every stat in this tool and every number the models produce. You don't need to know football or statistics to use this. Search or scroll to the term you're stuck on.

Dataset Modes

What each option in the "Dataset" filter actually compares.

Players

One player's own stats against their own other stats in the same game, e.g. does a QB's Completions predict their own Passing Yards.

Teams

A team's stats against its own other stats in the same game, e.g. does a team's Total Yards predict its own Points Scored.

Player + Team

One player's stats against their team's stats in the same game, e.g. does a receiver's Fantasy Points predict whether the team won.

Player + Player

Two different players' stats against each other, but only for games where they were actually teammates that week. Useful for questions like "does Player A's turnovers affect Player B's targets."

Pair Comparison

Two separate player pairs, compared side by side, e.g. is a QB/WR relationship stronger for one duo than another entirely different duo on a different team.

Predictions

A separate tab from the model builder above. Answers "how will this player do" by fitting a real model on their history, then projecting it forward.

Why two predictors, not one

A breakout game usually needs the QB playing well AND a soft matchup at the same time. A line using only one predictor can't produce a real spike, since every point sits on the same slope. You can pick as many predictors as you want (a real Multiple Regression), mixing QB stats with Matchup Context fields, so a good matchup can genuinely lift the projection above what QB play alone would suggest.

Picking your own predictors

You can choose a Regression (for continuous stats), Logistic Regression (binary), or Poisson Regression (counts), and pick as many predictor variables as you want for any of them. Any Matchup Context field you pick is automatically pulled from the projected player's own opponent that game; everything else is automatically pulled from their real teammate. Multicollinearity is checked live the same way it is in the main model builder.

Swap

Flips which of the two players is "being projected" and which is the "real teammate," with one click. Useful if you set them up backwards, or if you want to see the QB's own projection instead of the receiver's, using the exact same real relationship.

Do I need a third player?

No. Just picking the two real teammates already gets you a full season projection, using the real teammate's own history against next season's schedule as the driver. The third "try a different QB/teammate" field is only for testing a hypothetical ("what if a different player performed at this level instead"), not required for a normal projection.

The "real" card

The left card is a normal, real regression, fit only on the projected player's actual career games, their actual teammate, their actual opponents. Nothing hypothetical about it, it's the same kind of output you'd get from the main model builder.

The "projected" card

Each driver stat prefers the stand in's real history specifically against that week's actual opponent (their real games vs that exact team, from any season). A QB's own tendency against a specific defense is better evidence than one blended career number. If they've genuinely never played that opponent, it falls back to what that defense typically allows for the same stat (a real benchmark, tagged "N/A, opp avg" in the breakdown table), and only falls back further to their career average if neither real signal exists for that stat. Each matchup field uses the projected player's own real schedule for next season, already announced, and each opponent's most recently completed full season on defense, since there's no in season data yet for a season that hasn't been played.

Why career average, not a single game

A single game is noisy. Averaging over a player's full history gives a more stable, defensible read on their typical level to project forward with, which is why this uses as much of their real career as is available rather than just their most recent game or season.

Matchup Fields for Specific Positions

A defense's weakness often varies a lot by position, not just overall. "Receiving Yards Allowed to TEs" tells you something a blended "total pass yards allowed" number can miss, since a defense can be strong against WRs but weak against TEs (or vice versa). When both players resolve, this is automatically picked as the default matchup field whenever a field specific to that player's position exists.

Season Total

The sum of every projected game for the coming season, not just an average per game, so you can see the full season picture in one number.

Why the "real games" count might look low

Every real-game count in Predictions only covers the 2022 through 2025 seasons, plus 2026 games played so far, since that's the full range of Visor's underlying data. If two players have actually been teammates longer than that (a common situation for veteran QB and receiver pairs who've played together for four, five, even six real seasons), the earlier seasons simply aren't part of this tool, so the count you see here will be lower than their full real history together. This isn't a counting error, it's a real, honest reflection of what data Visor currently includes.

Confidence Interval (CI)

A range around each projected number, shown at 90%, 95%, or 99% (pick which one at the top of the page). It does not mean "there is a 95% chance the real result lands in this range." It means: if this same projection method were used many times on new data, about 95% of the ranges it produced would contain the true value. A wider range usually means less data or a noisier relationship to work with; it is honest uncertainty, not a mistake.

Confidence Score

A percentage shown per week (and averaged for the season) that reflects how much to trust that specific projection. It combines two things: how well the underlying relationship has explained real results in the past, and how typical this particular week's matchup is compared to what the model was built on. It is not a probability that the projection will come true, and it is a different measurement than the Confidence Interval above; a tight range and a high score usually go together, but they answer different questions. In the QB + Receiving Corps table, this score describes the QB's projected number in that row, not the individual receivers listed earlier in the same row; each receiver's own projection has its own separate Confidence Interval further down the page, under "Each Receiver's Own Projection."

Keeping up with trades and free agency

The projected player's team comes from real, current roster data whenever available, not from guessing based on old game logs, so a real offseason move (a trade, a free agent signing) is reflected correctly. If a player isn't found on a current roster yet, a note says so and falls back to their most recently logged team instead of silently guessing.

Why some fields aren't offered

The QB stat list and the list of stats you can project are both restricted to fields that make sense for each player's own position, the same restriction used in Player + Player mode.

QB + Receiving Corps

A separate tab under Predictions. Projects a QB's own season (yards, TDs, completions, whatever you pick) using multiple real pass catchers at once, instead of just one teammate.

How the real model gets fit

Each receiver you add becomes its own predictor (using the same stat, applied to everyone), fit as a real Multiple Regression on the QB's actual career games where that whole group of receivers was genuinely on the field with him. A Matchup Context field is always included too, for real variation from week to week. It's picked to actually match what you're projecting (Passing TDs Allowed for a Passing TDs projection, Turnovers Forced for Interceptions, and so on), not a single field used for everything.

Why one stat for every receiver

You pick one stat (like Targets) that gets applied to every receiver in the corps, rather than a different stat per person. Keeps the model interpretable: you're asking "how does the QB's output respond to each of these receivers getting more of the same thing," not mixing incompatible measures.

Projecting forward

Same real, opponent by opponent approach as the rest of Predictions: each receiver's input for a given week prefers their own real history against that specific opponent (from any team they've played for), falling back to what that defense typically allows when they've never faced them.

Accounting for a Missing Corps Mate

A receiver's numbers can look inflated in games where a fellow pass catcher was actually out (more targets to go around). So each receiver's own history excludes games where another named receiver in this corps was confirmed absent, keeping the average closer to what a normal, fully healthy week looks like rather than blending in performances boosted by someone else's absence.

A receiver new to the team

Someone with zero real shared games with this QB (a trade, a free agent signing, a rookie) can't get a real fitted coefficient, but is still included in the season projection using the average value per unit the other real receivers show with this QB, applied to their own career history. This is clearly labeled "estimated role" everywhere it appears, so it's never confused with a real, fitted relationship.

A QB with an entirely new receiving corps

If NONE of the named receivers have ever played with this QB (he just switched teams entirely), there's nothing specific to this QB to fit at all. The projection falls back to the QB's own real career average as a baseline, adjusted using a real relationship pooled across every real QB and teammate pair league wide, labeled "league wide role." This is honestly a rougher, less personalized estimate than the other cases here (the league wide fit itself tends to be weak, since one receiver's volume only explains a small slice of a QB's total output). It's the best available before any real games are played together, and should be revisited once they are.

Each Receiver's Own Projection

Beyond the QB's combined number, every receiver gets their own season total and average per game for the stat you picked (like Targets for the full season), shown separately, whether they're a real fitted part of the model or an estimated new addition.

Teammate Presence

The "Teammate Presence" filter, available once you've picked a player under Player Lookup.

Only games with them

Narrows your selected player's games to ones where the named teammate had 1 or more offensive snaps that week, i.e. actually played.

Only games without them

Narrows your selected player's games to ones where the named teammate had 0 recorded offensive snaps, a genuine explicit absence, not just a quiet stat line. A player can have zero catches while still playing every snap; this filter looks at snap counts specifically so it doesn't confuse the two. It can't tell you why they were out, though: injury, bye week, and healthy scratch all look identical in this data (zero snaps), so treat "without" as "confirmed didn't play," not as a specific reason.

Bye weeks and Teammate Presence

A team's bye week isn't specially handled: since neither player has a game that week at all, it simply never shows up as a row for either "with" or "without," the same as any other week where the main player themselves didn't play.

Teammate Presence in Player + Player

Works the same way when comparing two players against each other, but checks a third player's presence instead. For example, comparing Joe Burrow's Pass Attempts to Ja'Marr Chase's Targets, filtered by whether Tee Higgins played that week.

Excluded games

A game only gets excluded when the reference player has never, in any season, played for the team in question at all. If they've been on that team at any point, every game there where they personally didn't record snaps counts as a real "without," whether that's before they joined, after they left, hurt while on the roster, or any other reason. The note tells you how many games were excluded and why.

Head to Head panel

In Player + Player mode, shows both players' average stats side by side for whichever variables you've selected. If a Teammate Presence reference is set, it splits each average into "with" vs "without" so you can see the shift directly, regardless of which one the toggle below is currently set to.

With vs. Without Comparison

Also in Player + Player mode when a Teammate Presence reference is set: two complete, independent models, one fit only on games with the reference player, one fit only on games without them, each with its own equation, chart, R squared, RMSE, and p value. Lets you see whether the relationship itself changes, not just the average.

Pair Comparison

A separate dataset mode for comparing two entirely different player pairs, e.g. does a QB's Fantasy Points predict their WR's Fantasy Points, and is that relationship stronger for one QB/WR duo than another. Runs two full, independent models side by side. If both pairs share the same first player, an extra filter appears letting you isolate each pairing to games where the other pair's second player did or didn't also play.

Understanding Your Results

What the numbers under "Model Output" actually mean.

Coefficient

How much the response variable changes for every +1 unit of the predictor, holding everything else constant. A coefficient of 0.5 on Passing Yards means each extra passing yard adds 0.5 to the response, on average.

Intercept

The model's predicted value when every predictor equals zero. Usually not meaningful on its own, it's just the starting point the line is drawn from.

R squared (R²)

The percentage of the response's variation that this model explains, from 0% (explains nothing) to 100% (explains everything). Real world sports stats rarely get above 30 to 40%. That's normal, not a failure.

Adjusted R squared

Same idea as R², but penalized for adding predictors that don't really help. If Adjusted R² is a lot lower than R², some of your predictors are probably just noise.

RMSE

"Root Mean Squared Error," roughly how far off a typical prediction is, in the same units as your response. An RMSE of 20 on Fantasy Points means predictions are usually off by about 20 points.

Std. Error

How uncertain the coefficient estimate is. A small standard error relative to the coefficient means the estimate is precise; a large one means it's shaky.

95% CI

A range that, roughly speaking, is likely to contain the true coefficient. If it doesn't cross zero, that's consistent with the p value being significant. A wide range means the estimate isn't very precise, even if it's statistically significant, so treat the exact number with some caution. On the same scale as the coefficient itself (log odds for logistic, log rate for Poisson), not the Odds Ratio or Rate Ratio.

Methods Summary

A paragraph ready to paste into a write up, summarizing the current model in standard reporting style (coefficient, standard error, test statistic, p value, 95% CI, and odds/rate ratio where relevant), for pasting straight into a report or paper.

t stat / z stat

The coefficient divided by its standard error. Bigger (positive or negative) means the effect is more likely real rather than random noise. As a rough rule of thumb, anything past ±2 is usually significant.

P Value

The probability you'd see a relationship this strong purely by chance, if the predictor actually had no real effect. Lower is more convincing. Under 0.05 is the common "statistically significant" threshold.

VIF (Variance Inflation Factor)

Flags when two predictors are basically measuring the same thing (multicollinearity). Under 5 is fine, 5 to 10 is worth a second look, over 10 means the coefficients above may be unreliable.

Odds Ratio (logistic only)

How much the odds of the outcome multiply for each +1 unit of the predictor. An odds ratio of 1.5× means the odds of "yes" go up 50% per unit; under 1× means the odds go down.

Rate Ratio (Poisson only)

How much the expected count multiplies for each +1 unit of the predictor. A rate ratio of 1.2× means the expected count goes up 20% per unit; under 1× means it goes down. The Poisson equivalent of an odds ratio.

Pseudo R² (Poisson, deviance based)

A different formula than logistic's McFadden pseudo R², but the same idea: roughly, how much better this model explains the counts than just guessing the average every time. Not directly comparable to linear R² or logistic's pseudo R².

Dispersion (Poisson only)

Poisson regression assumes a count's variance equals its mean. This checks whether that's actually true in your data. Close to 1.0 is expected; well above 1.5 means the data is more spread out than the model assumes (overdispersion), which makes the p values above look more confident than they really are.

Mean Abs. Error (Poisson only)

On average, how far off this model's predicted count was from the actual count, in the same units as the response. A Mean Abs. Error of 0.8 on Passing TDs means predictions are typically off by less than one touchdown.

Pseudo R² (McFadden's, logistic only)

Logistic regression's version of R². It uses a different scale. Even a "good" logistic model often only reaches 0.2 to 0.4, so don't compare it directly to a linear model's R².

Accuracy (logistic only)

The percent of games where the model's yes/no guess (using a 50% cutoff) matched what actually happened.

Advanced Options (opt in)

Extra diagnostics you can turn on under "Advanced Options" when building a model.

Prediction Range

The shaded zone around the fitted line shows where about 95% of individual games are expected to land, not just where the average line sits. A wide zone means games vary a lot at that point; a narrow zone means they tend to cluster close to the line.

Passing Stats

Completions

Passes thrown that were caught (in bounds) by the QB's own teammate.

Pass Attempts

Every pass thrown, whether it was caught, dropped, or incomplete.

Completion %

Completions divided by attempts. How often the QB's throws were caught.

Passing Yards

Total distance the ball traveled downfield on completed passes (from the QB, credited on the throw).

Yards / Attempt

Passing yards divided by attempts. How much ground each throw gains on average, whether or not it's caught.

Passing TDs

Touchdown passes thrown.

Interceptions

Passes caught by the other team instead of the QB's own team. A turnover.

Sacks Taken

Times the QB was tackled behind the line of scrimmage before getting a pass off.

Passing First Downs

Completed passes that gained enough yardage to earn a new set of downs (a fresh chance to keep the drive alive).

Passing EPA

"Expected Points Added," a play by play value for how much a pass improved (or hurt) the team's scoring chances, summed over the game. Positive is good.

300+ Passing Yards

Yes/No: did the QB throw for 300 or more yards in this game? A commonly used "big game" benchmark.

Rushing Stats

Carries

Times a player ran with the ball (a "rushing attempt").

Rushing Yards

Total distance gained while running with the ball.

Yards / Carry

Rushing yards divided by carries. How much ground the runner gains per attempt, on average.

Rushing TDs

Touchdowns scored by running the ball into the end zone.

Rush Fumbles Lost

Times the runner dropped the ball and the other team recovered it. A turnover.

Rushing First Downs

Runs that gained enough yardage to earn a new set of downs.

Rushing EPA

Expected Points Added from running plays, summed over the game. Positive is good.

Receiving Stats

Targets

Times a pass was thrown to this player, whether or not they caught it.

Receptions

Passes this player successfully caught.

Catch Rate

Receptions divided by targets. How often a pass thrown their way was actually caught.

Receiving Yards

Total distance gained after catching the ball.

Yards / Reception

Receiving yards divided by receptions. How much ground gained per catch, on average.

Receiving TDs

Touchdowns scored by catching the ball in the end zone.

Rec. Fumbles Lost

Times this player fumbled after catching the ball and the other team recovered it.

Receiving First Downs

Catches that gained enough yardage to earn a new set of downs.

Receiving EPA

Expected Points Added from this player's catches, summed over the game.

Target Share

The percent of the team's total pass attempts that were thrown to this player. A measure of how central they are to the passing game.

100+ Scrimmage Yards

Yes/No: did rushing + receiving yards combined reach 100 or more? Another common "big game" benchmark, used for both runners and receivers.

Team / Game Stats

Points For

Points this team scored in the game.

Points Against

Points the opponent scored.

Point Differential

Points For minus Points Against. Positive means they outscored the opponent (a win); negative means they didn't.

Win

1 if the team won, 0 if they lost, 0.5 for a tie. This is the variable used for logistic regression's win/loss modeling.

Total Yards (Team)

The team's passing yards plus rushing yards for the game. Total offensive output.

Total TDs (Team)

All touchdowns the team scored on offense (passing + rushing) in the game.

First Downs (Team)

The team's total first downs earned, from both passing and rushing plays.

Turnovers (Giveaways)

Interceptions thrown plus fumbles lost. Every time the team handed the ball to the opponent.

Sacks (Team)

Times the team's QB was tackled behind the line before throwing.

Fantasy & "Big Game" Stats

Fantasy Points (Standard)

A single score combining a player's yards, touchdowns, and turnovers using common fantasy football scoring rules (does not count catches on their own).

Fantasy Points (PPR)

Same as standard fantasy points, but each reception (catch) is worth an extra point. "Points Per Reception."

Boom Game

Yes/No: did the player score 20 or more PPR fantasy points in this game? A "great fantasy day" benchmark.

Bust Game

Yes/No: did the player score under 5 PPR fantasy points? A "disappointing fantasy day" benchmark.

Scored a TD

Yes/No: did this player score any touchdown (passing, rushing, or receiving) in this game?

Multi TD Game

Yes/No: did this player score 2 or more touchdowns in this game?

Turnover Committed

Yes/No: did this player throw an interception or lose a fumble in this game?

Offensive Line Stats

Centers, Guards, and Tackles don't record passing, rushing, or receiving stats, so they use a different set of numbers based on playing time instead.

Offensive Snaps Played

The number of offensive plays this player was on the field for during the game.

Offensive Snap %

Offensive Snaps Played divided by the team's total offensive plays that game. 100% means they played every offensive snap.

Started (90%+ Snaps)

Yes/No: did this player take part in at least 90% of the team's offensive snaps? A stand in for "started and played the whole game" without relying on the official start list.

Defensive Player Stats

Defensive Line (DL), Linebacker (LB), Cornerback (CB), and Safety (S). These positions don't have offensive stats at all (their passing/rushing/receiving numbers are all 0, not missing), and their Matchup Context is left blank rather than showing the opponent's own defense, which wouldn't mean anything for them.

Sacks

Times this player brought down the quarterback behind the line of scrimmage. Can be a fraction like 0.5, since a sack is sometimes split between two players who both got there. Because of that, Sacks isn't offered as a Poisson response (Poisson needs true whole number counts); "Recorded a Sack" is a clean 0/1 substitute if you want that kind of model instead.

QB Hits

Times this player hit the quarterback on a pass play without necessarily bringing them down for a sack. A broader measure of pressure on the passer than sacks alone.

Tackles For Loss

Tackles made behind the line of scrimmage, stopping the play before it gained any yards.

Total / Solo / Assisted Tackles

Total Tackles is every tackle this player was involved in. Solo means they made it alone; an Assisted Tackle means they shared credit with a teammate.

Passes Defended

Passes this player batted down or broke up, preventing a completion, not counting interceptions (those are tracked separately).

Forced Fumbles

Times this player knocked the ball loose from the ball carrier. Different from a fumble recovery, which is picking the loose ball up, possibly by a different player entirely.

Fumble Recovery (Own vs. Opponent)

Own Team's Fumble means recovering a loose ball your own offense fumbled. Opponent's Fumble means recovering one the other team fumbled, the more meaningful defensive play of the two.

Defensive Touchdowns / Safeties

Touchdowns scored directly by the defense (returning an interception or fumble to the end zone) and safeties (tackling the ball carrier in their own end zone), the rarest defensive plays with the biggest impact.

Why some fields are 0/1 instead of a count

"Recorded a Sack," "Recorded an Interception," "Forced a Turnover," and similar fields exist so you can ask yes/no questions with Logistic Regression, like "does pressure predict whether a player gets at least one sack," separate from asking how many.

Matchup Context

How tough was the opponent's defense, based only on what happened before this game (never the game itself, so there's no lookahead). Available on every player and team, since it describes who they're facing, not who they are.

Opponent Points Allowed Per Game (Entering This Week)

The average points the opponent had allowed per game so far that season, counting only games before this one. A lower number means a tougher defense coming in.

Opponent Pass Yards Allowed Per Game (Entering This Week)

Same idea, for passing yards. Useful for asking whether a weaker pass defense predicted a bigger receiving or passing game.

Opponent Rush Yards Allowed Per Game (Entering This Week)

Same idea, for rushing yards.

Opponent Total Yards Allowed Per Game (Entering This Week)

Passing and rushing yards allowed combined.

Opponent Sacks Forced Per Game (Entering This Week)

How often the opponent's defense had been getting to the quarterback so far that season. A higher number means a tougher, more disruptive pass rush.

Opponent Turnovers Forced Per Game (Entering This Week)

How often the opponent's defense had been forcing fumbles or interceptions so far that season.

Why "Entering This Week"?

These numbers only use games before the one you're looking at, on purpose, so a defense's stats from this exact game never leak into a prediction about this exact game. Week 1 of a season shows no value for any team, since there's no earlier data yet to average.

Field Goals Made / Attempted / Missed / Blocked

A kicker's real field goal volume and accuracy that game. FG% is makes divided by attempts. FG Long is the longest make; the 50+ fields track attempts and makes specifically from 50 yards or beyond, the hardest, most doubtful range.

PATs Made / Attempted / Missed

Extra-point kicks after a touchdown. Nearly automatic in the modern NFL, but still a real, trackable stat.

Made a Field Goal / Perfect Kicking Game

Yes/no versions of a kicker's game: whether they made at least one field goal, and whether they went perfect on every field goal and PAT attempt that game.

Punts / Punt Yards / Punt Average

How many times a punter kicked, the total yards on those punts, and the average per punt -- the standard measure of a punter's leg strength that game.

Punts Inside 20 / Touchbacks / Fair Catches Forced / Blocked

What actually happened to each punt: pinned inside the 20 (good for the punting team), sailed into the end zone for a touchback (a wasted opportunity to pin the offense deeper), forced a fair catch (no return allowed), or got blocked (a real special-teams breakdown).

Booming Punt Game

A yes/no marker for a game where the punter's average distance was unusually long, a quick way to flag a standout leg-strength performance.

Kick / Punt Return Attempts, Yards, Average, TDs

A returner's real production on kickoffs and punts separately: how many they returned, total yards, yards per return, and how many went all the way for a touchdown. Total Return Yards adds both return types together for one overall number.

Scored a Return TD

A yes/no marker for whether a returner took any kick or punt back for a touchdown that game.

Opponent Field Goals Made / Attempted Allowed Per Game

How often the opponent's defense has let a kicker attempt, and make, field goals so far that season -- a real signal for how often that defense's drives stall out in scoring range instead of reaching the end zone.

Opponent Completions Allowed Per Game

How many completions the opponent's defense has allowed per game so far that season, entering the week being looked at.

Opponent Passing TDs Allowed Per Game

How many passing touchdowns the opponent's defense has allowed per game so far that season -- useful matchup context specifically for a QB's touchdown-scoring stats, separate from their raw yardage.

Opponent WR / TE / RB Yards & Receptions Allowed Per Game

The same "what has this defense allowed" idea as the team-wide matchup fields, but broken out by position -- how many yards and receptions this defense has given up specifically to wide receivers, tight ends, or running backs so far that season. A defense can be tough on receivers but soft against tight ends, so the position-specific version is often more meaningful than one blended number.

Opponent Offense fields (Matchup Context for Defensive Players)

The mirror image of the "Opponent Defense" fields, used when projecting a defensive player instead: how much the opposing OFFENSE has scored, thrown for, run for, turned the ball over, or been sacked so far that season. A pass rusher's sack total, for example, is driven by how often the opposing offense's own quarterback gets sacked league-wide, not by what a defense allows.

"Team Stat" vs. "Another Player's Stat" (Special Teams & Defense)

Two different ways to drive a kicker's, punter's, returner's, or defensive player's projection. "Team Stat" uses the player's own team's real per-game number (like the team's own field goal attempts) as the predictor. "Another Player's Stat" instead uses a specific teammate's real history, the same way Two Players mode works. Neither is forced -- pick whichever real relationship makes more sense for what you're trying to see.

Week Snapshot

Runs several stats for one specific week side by side (any week, past or future), each showing the projection with all three confidence levels at once. For a week that's already been played, it also shows the real result and how close the projection actually was.

Compare Players

Runs several full-season projections side by side, one per player or stat, so you can visually compare two or more outlooks at a glance instead of running them one at a time.

Actual (Real Result)

Once a week has actually been played, its real result is shown right alongside what the model projected for it -- a direct, honest check of the projection against what really happened, not just a number with nothing to compare it to.

Accuracy (Close / A bit off / Way off)

How far a real result landed from the projection, measured in standard deviations, the same real unit of uncertainty the confidence interval itself is built from. Within 1 SD is labeled "Close," 1-2 SDs "A bit off," and beyond that "Way off." This is a genuine statistical distance, not an arbitrary percentage.

Why some weeks use a different fit than others

A week that's already been played is projected using a fit that excludes the current season entirely, so its own real result can never leak into the very model used to "predict" it -- keeping the Actual/Accuracy comparison honest. Every week that hasn't happened yet instead uses a fit that includes real games from the current season, weighted more heavily than older seasons (see below), since there's no leakage risk for a game that hasn't been played.

Recency-weighted fitting

For any week that hasn't been played yet, real games from the current season count for more than older seasons, and the current season's overall share of the model's attention grows automatically as more of it gets played -- a handful of early-season games count for a modest amount, not an outsized one, while a fuller current season carries considerably more weight. Older seasons are never thrown out; they still make up real, substantial part of the fit even once the current season is well underway.

Defensive Penalties / Penalty Yards

Real penalties called on a defensive player that game, and the yardage given up because of them -- a discipline signal separate from their actual production (tackles, sacks, coverage).

Forced a Turnover (Yes/No)

A yes/no marker for whether a defensive player was directly responsible for a turnover that game (an interception or a forced fumble), regardless of who actually recovered it.

10+ Tackle Game (Yes/No)

A yes/no marker for a standout tackling performance: whether a defensive player reached double-digit tackles (combined solo and assisted) in a single game.

How this prediction actually works

This isn't a straight line "if the teammate plays well, so does the response" model. That misses the thing that actually creates a breakout game. A real spike usually comes from multiple things lining up at once: a teammate playing well, and a soft matchup that week, and sometimes more. So you can build the model here just like the main model builder: pick a Regression, Logistic, or Poisson model, and choose as many predictor variables as you want, fit on the projected player's own real career history.

Player's Stat = Teammate's Stat(s) + Opponent's Defense Strength That Week

Any Matchup Context field you pick is automatically pulled from the projected player's own opponent (computed only from games before the one in question, never leaking in information from the game itself); everything else is pulled from their real teammate.

Once that relationship is fit on real history, you can project it forward using a different player's real career average stats as a stand in for "what if this teammate performed at that level" applied against the projected player's own real, already-announced schedule for next season, and each opponent's most recently completed full season on defense (since there's no in-season data yet for a season that hasn't been played). Every projected point is a real calculation from real numbers, not a guess, but it's still a model output: treat it as "what the pattern says," not a guarantee.

Confidence interval

The interval above and the "Confidence" score in the tables are two different things. The interval is a range derived from real statistical formulas, with the repeated sampling meaning explained above. The confidence score is a single percentage per week — it blends how well this relationship has explained the data historically (R²) with how tight that week's own interval is relative to the response's typical spread. It is a rough, built-in "how much to trust this specific projection" gauge, not a probability that the outcome happens and not the same statistical guarantee the interval carries.

Set up the comparison

Player you're projecting

Their real teammate

Must be someone they've actually played with. Any position works, not just a QB. This is where the real relationship gets fit from.

Enter both names to see their shared game history.

Try a different QB/teammate instead (optional)

Leave this blank and you'll still get a real season projection. It'll just use the real teammate above's own history against next season's schedule. Only fill this in if you want to test "what if a different player performed at this level instead."

Starter Examples

Predictor variables (choose 1)

Stat to project (choose 1)

Pick a player, their real teammate, and a stat, then click Run Prediction for a real season projection. No third player needed.

What is Linear Regression?

Starter Examples

Predictor variable (choose 1)

Response variable (choose 1)

Model Output

Head to Head

With vs. Without Comparison

Fit

Coefficients

Which Predictor Matters Most

Does It Hold Up On New Games?

Is The Whole Model Meaningful?

Adjustable Yes/No Cutoff

ROC Curve

Pair Comparison

Run History

Dataset Modes

Predictions

QB + Receiving Corps

Teammate Presence

Understanding Your Results

Advanced Options (opt in)

Passing Stats

Rushing Stats

Receiving Stats

Team / Game Stats

Fantasy & "Big Game" Stats

Offensive Line Stats

Defensive Player Stats

Matchup Context

How this prediction actually works

Starter Examples

Predictor variables (choose 1)

Stat to project (choose 1)

Prediction

What Goes Into Each Projected Week

Starter Examples

Prediction

What Goes Into Each Projected Week

Starter Examples

Prediction

What Goes Into Each Projected Week

Starter Examples

Prediction

What Goes Into Each Projected Week

Week Snapshot

Compare Players

Export