TD Leaders · Ron's Tud Tool

Loading…

How Picks Work ▶

Every skill player on the slate, priced to score a touchdown — and every starting QB, priced to throw one.

What the price is. The modelled chance the player scores at least one offensive touchdown (rushing or receiving), shown as American odds with the raw probability beneath. These are fair odds with no vig — a real sportsbook will always price worse, and the gap is their margin.
It is calibrated, not a rating. Out-of-sample across 2017–2025 the number means what it says: players priced around 30% scored about 30% of the time. That is the whole point of showing a probability rather than a 0–100 grade.
What drives it. Overwhelmingly recent snap share — the share of his offense's plays he was on the field for over his last three games. That single input is about 87% of the signal. Then the game's implied team total (from the spread and total), his recent red-zone touches, his share of the touchdowns his team scores, and whether a same-position teammate has just been ruled out.
Opportunity, and conversion. Almost everything above measures how much a player is on the field. None of it measures whether that turns into points, and for a long time this board didn't either — which is why it topped out near −147 while sportsbooks price lead backs at −300. The tell was in the history: the player-seasons that actually scored in 70%+ of their games carried lower snap share and fewer red-zone touches than this model's own top-priced players. Jahmyr Gibbs scored in 78% of his 2024 games on 54% of the snaps, and the model had him at +103. It now also reads each player's share of his team's touchdowns, which is worth roughly double every other candidate tested at that stage combined — goal-line carries inside the five, team red-zone share and touch share were all nulls. Fair warning: this is information sportsbooks price too, so it mostly makes our numbers agree with theirs rather than beat them.
What is deliberately absent. Opponent defense, red-zone defense, pace, weather, home/away and blowout risk were each tested over ten seasons and none improved the forecast. Defensive touchdowns-allowed is close to pure noise year to year, and the implied total already prices the scoring environment.
Five markets. 1+ TD is the anytime price. 2+ TD is that multiplied by the chance a scorer scores again — so it can never come out shorter than the anytime price, which two separately fitted models would get wrong sooner or later. That second step is its own model and is nowhere near a constant: it runs 8% to 28% across red-zone workload, and 23% for backs against 11% for tight ends.
1st TD is a race, not a coin flip. Exactly one player scores the opening touchdown, so a price here is a share of his game rather than a standalone chance. It is the anytime price divided by the total scoring threat on the field, which is why two backs at the same anytime number can sit far apart: one may be boxed in with three other likely scorers while the other has the game to himself. Pricing that competition is worth real accuracy, and each game's prices sum to the 97% of games whose opening score goes to someone we carry.
It is recalibrated. The raw model runs hot on favourites — out of sample it said 75% where 60% actually happened, and 82% where 61% did, because a logistic extrapolates in a straight line while scoring saturates. A monotone recalibration, fitted on walk-forward out-of-sample results, trims the top end: the 70%-plus band's error drops from 15.5 points to under 3. It is why a back who looks like a lock is priced near -150 rather than -350, and each pick's breakdown shows the trim as its own step rather than hiding it.
Pass TD is a separate model. The anytime model scores a quarterback as a rusher — his arm never enters it — so passing touchdowns needed their own. It is a Poisson on three inputs: the implied team total, his pass volume, and his own passing-TD rate over his last two seasons of starts. The line is Over 1.5, the number books hang, and the base rate is 44% — a coin flip rather than a longshot like the other four. Starting quarterbacks only, chosen by who actually throws rather than by depth chart.
What was tested and dropped there. Red-zone efficiency, weapon quality and opposing pass defense are all genuinely stable traits — team receiving production split-halves at r=0.89 — and all three added nothing once the implied total was in. That is not a null result about football; it is the betting market having already priced them. Out of sample the passing model separates the top fifth of the board from the bottom fifth by 2.6× on Over 1.5, against 0.500 on a shuffled control.
Only players on the active roster are priced. The weekly roster feed carries every transaction, and reading it whole put 180 of 480 players on this board who could not take the field — 66 released, 31 on injured reserve, 80 on the practice squad, plus retired and exempt. They held 30% of the board's probability. That is bad on its own, but it is worse once snaps are normalized below: a cut player doesn't sit there harmlessly, he consumes part of his old team's snap budget and takes it off the players who are actually playing. The real players were left holding 1.51 expected scorers per team against the 2.04 that actually happens, which made every honest price about 35% too long. The board is now filtered to active status, and it prices 2.06 per team.
Snap shares are normalized to the team. A player's recent snap share belongs to the team he last played for, and in week 1 that is always last season's — so a roster that added talent carries several players all wearing a lead role they can't all have. Tennessee projected Ridley, Robinson, Ayomanor and Dike at their old shares; Pittsburgh carried Dowdle and Warren both as the lead back. Because a team fields exactly five skill players and one quarterback on every snap, those shares must sum to five and one — measured over ten seasons they do, 4.99 and 1.00 — so each team's projections are resolved back to that budget. Not by cutting everyone equally, which would take the lead man down as hard as the imports. Rank a team's players within their position and the real distribution is steeply top-heavy — the top back plays 59% of snaps, the third 11% — so a crowded room resolves by the depth collapsing while the starter holds. Each player is blended halfway toward the share his depth rank historically earns, and only the small residual is scaled. Depth comes from the published depth chart, which is reloaded every build: inferring it from snap history works for established players and fails exactly where it matters most, because every player without NFL history carries the same position prior and a rookie backfield is a pile of ties. It corrects both ways: thin rooms get scaled up. On weeks 1-3 of 2016-2025 that beats both the raw feature and an equal cut on log loss, AUC and calibration (ECE 0.0138 raw, 0.0099 equal-cut, 0.0054 here). By week 4 the underlying feature is all current-season anyway and the adjustment fades on its own.
Return men are priced whole. A kick or punt return touchdown pays as an anytime touchdown, so for anyone with a return role his price is his offense and his return chance combined — flagged RET, with the split shown in his card. Pricing offense alone understated every returner on the board.
Injury status is in the price. Anyone ruled Out or Doubtful is off the board entirely. Questionable players stay — they usually play — but their price is shaded 12% and badged Q. That number is measured, not assumed: across 2016-2025 the board priced Questionable players at 0.213 and they scored at 0.187, while unlisted players came in at 0.997 of their price. The model is calibrated; the tag is information its features can't see, since snap history describes a healthy player.
Cut lines. Each minimum-odds filter shows the hit rate that cut actually produced in the backtest for that market — so tightening the board is a measured trade, not a guess.
Honest limits. Fitted on 2016–2025 and never yet graded against a live week. It is also not measured against sportsbook prices, so a fair number is not the same thing as an edge.

Commas match any of them — narrows whatever the filters already show

Loading…

How Leaders Work ▶

Season touchdown leaders — sort by total, rushing, receiving, defensive, special-teams, first-TD, or opportunity.

Loading…

How Recap Works ▶

Every touchdown of the season, by week — scorer, type, distance, and quarter.

Loading…

How the Schedule Works ▶

The week game by game: every game's projected box score, its lines and how they've moved, graded once it's played.

Loading…

How Birthdays Work ▶

Who is playing on or near his birthday — and whether birthday boys actually find the end zone.

Loading…

How Due Works ▶

Heavy red-zone workload, no touchdown to show for it — the guys whose usage has outrun their scoring.

Loading…

How Milestones Works ▶

Who is closing in on a career touchdown milestone — 50, 100, 150, 200.

Loading…

How Returners Works ▶

Who takes the punts and kickoffs, and what that adds to his anytime touchdown price.

There is no return-TD market. A kick or punt return touchdown pays as an anytime touchdown, so that is what this board prices: his offensive chance and his return chance combined. The +ret column isolates what the return role adds — the piece a pure offense model like our Picks board can't see, and the reason a returner's anytime price should sit shorter than his target share alone suggests.
Volume comes from the opponent. They punt when their offense stalls and they kick off when they score, so a returner's workload is mostly a fact about the other team. Both mechanisms check out: punt returns track the opponent's actual punts at r=+0.80, kick returns track their points at r=+0.38.
But the two cancel. This is the part worth knowing before betting a "high total means more returns" hunch. A weak opponent punts more and scores less; a strong one does the reverse. Across opponent-implied-total terciles a team sees 6.2 / 6.4 / 6.3 returns — flat. The matchup moves the mix, not the volume.
The mix is what matters. Only because the two types don't pay the same: a punt return reaches the end zone about 1.5× as often as a kick return. So a punt specialist facing a stalling offense is the one real angle here — and even then it's small, moving expected return-TD from roughly 2.5% to 2.6% a game.
Share is the bigger lever. Once the opponent effects wash out, what separates returners is how much of his own team's work he takes. Usage is a genuinely stable role — split-half r=0.56 on punts, 0.50 on kicks — so the share column is doing more than the matchup column.
Kick numbers are 2025 only. The dynamic-kickoff change took the league return rate from 33% to 74.5% in one off-season. Earlier seasons describe a game that no longer exists, so they're excluded rather than averaged in. Punt rates are stable and use five seasons.
Honest scale. A return touchdown happens in about 6% of games league-wide, so for most returners the return role adds a point or two of anytime chance, not more. The board is sorted by that add rather than by the combined price, because the combined price is mostly his offense — which the Picks board already tells you.

Loading…

How Completions Works ▶

Projected completions for every starting quarterback, priced against the standard lines.

Loading…

How Interceptions Works ▶

Every starting quarterback, priced to throw an interception.

Loading…

How Kickers Works ▶

Field goals made, PATs made and kicker points for every team on the slate, priced from the game — not the kicker.

Loading…

How Yards Works ▶

Projected passing, rushing and receiving yards for everyone on the slate, priced at any line.

Loading…

How Receptions Works ▶

Projected catches for every pass-catcher on the slate, priced against the standard lines.

It is a volume market, not a scoring one. Touchdowns turn on who gets the ball at the three; receptions turn on how often a player is on the field and thrown at. That makes the role far more predictable — a receiver's game-to-game catch volume splits half-against-half at 0.947, which is why this is the market we built next rather than kickers (0.365) or interceptions thrown (0.145), both of which are close to noise.
Recent snaps carry it, again. The largest term in the model is his snap share over the last three games, the same input that dominates the touchdown board. His longer-run catch rate, his share of the team's targets and his team's pass volume fill in the rest.
Prices are negative binomial, not Poisson. This matters and it is not a technicality. Catch counts are more spread out than a Poisson allows — variance runs 1.1–1.3× the mean — so pricing them Poisson pushes too much probability past the line. It would have over-priced the over by 3 points at 2.5 catches and 9 points at 6.5. The wider distribution fixes it.
Wind is in, and nothing else about the weather is. Every outdoor game carries a forecast at kickoff and the projection is scaled by a measured multiplier — about −2% at 10 mph, −6% at 15, −15% at 25. Two mechanisms, both measured: teams throw less (37.0 attempts in calm air, 34.2 at 20+) and complete less of what they throw (66.3% down to 63.0%). Domes and closed roofs are unaffected by definition. Temperature, indoors and game script were each tested against this market and are nulls.
Why a multiplier and not a coefficient. Wind started life as a term inside the fit and it did not work, which the Results tab is what caught: 20+ mph games were projected at 2.25 catches and actually caught 1.71. Two reasons. The term used a threshold at 15 mph, but the real decay is continuous from calm and accelerating, so everything below 15 was thrown away. And a coefficient fitted against a column that is zero in 89% of rows comes out far too weak. It is now measured directly from out-of-sample error by wind bin.
Opposing defense is a null here too. Receptions allowed to a position carries no predictive weight once usage and team pass volume are in — the fifth time defense has failed a test on this site.
No wind forecast is wired up yet, so every game on this board is currently priced calm. When it is added, windy games will move and nothing else will.

Loading…

How Pairings Work ▶

Which players find the end zone together, and which quarterbacks throw it to whom — both priced, with the correlation naive parlay math misses.

What it counts. Across the season, how often each group of players scored a TD in the same window. Ranked by how many times they've gone together.
Window. Same week (anyone on the slate) or same day (Thu / Sun / Mon) — day is the tighter correlation.
Group size. Pairs, Trios, Quads, or Quints — the bigger the group, the rarer (and longer-odds) the combo.
SGP. When 2+ members share a team they play the same game — a true same-game-parlay stack, flagged SGP·TEAM.
Basis. Count every touchdown, or narrow to one of the markets the Picks board prices: each game's opening score, its closing score, or only multi-TD games. The narrower bases are far rarer, so the co-occurrence floor drops to 2.
The price, and the correlation. When every member is on this week's Picks board, the card carries a fair price — and next to it, struck through, what you'd get by simply multiplying the legs together. The gap is the whole point.
Teammates compete. The gap runs the opposite way to what "correlation tax" implies. A team scores about 2.5 touchdowns a game; when the back takes one, that's one fewer for the tight end. A same-team double lands 11% below the naive product, a trio 29% below, a four-man stack 53% below. Legs split across the two teams are effectively independent (1.02×).
Where the numbers come from. Every within-game combination from 2016-2021 was priced with our own model and compared to what actually happened, then checked on 2022-2025. At four legs the naive product overstates the parlay by 126%; the correction leaves 7%. Pure same-team groups use the multiplier measured for that exact size, because cannibalisation compounds faster than a per-pair rule implies; mixed groups apply it once per pair.
"But books charge me MORE for same-game legs." They do — and that isn't a contradiction, because a worse price and a higher probability are different things. Three reasons a book shortens these that have nothing to do with TD scorers being positively correlated: same-game parlays carry the highest hold on the board; a book's correlation engine is generic across every market in the game, and most same-game legs (passing yards + receiving yards, team total + props) genuinely are positively correlated; and it has to defend against the one combination that really is — a quarterback and his own receiver. So a same-team TD double is bad twice over: less likely than multiplying suggests, and priced shorter than multiplying suggests. That gap is the reason to show it.
Why we trust the number. The comparison that matters doesn't depend on our model being perfectly calibrated: at the same naive price, same-team pairs hit about 13% less often than opposing-team pairs. And the opposing-team control lands at 1.02× — when legs genuinely are independent, our product is right, which is what says the same-team gap is real rather than a pricing error. It holds for every combination: two pass-catchers 0.89×, back plus pass-catcher 0.92×, two backs 0.83× — cannibalisation wins even where two receivers share a quarterback.
What this is not. A fair, no-vig number is not a claim you can beat the book. It tells you what the bet is worth, so you can see how far the offered price sits from it.
QB → Receiver. That chip switches to a different relation entirely: every quarterback-to-receiver touchdown connection, ranked by how many he actually threw him. Receiver means whoever caught it — WRs are 38% of these, tight ends 14% and running backs 11%, so calling it "QB→WR" would leave out a quarter of them. It isn't a filter on the pairs above, and it can't be — only 11 quarterbacks scored at all last season, because a QB's own touchdown is a rushing one, while 67 threw 854 scoring passes. Filtering the co-occurrence groups for QB+receiver would have surfaced almost nothing, and what it did surface would be "the QB ran one in and, separately, someone caught one" — not a connection at all.
Trios and up. The size tabs work here too: a trio is one quarterback and two different receivers he threw touchdowns to in the same week, which happens more than you'd think — 229 quarterback-weeks last season had two different targets score, 67 had three. Prescott to Pickens and Ferguson hit four separate weeks.
A trio isn't the pair squared. It mixes a positive link with a negative one: each receiver runs above independence with the passer, but the two of them compete with each other for his throws. Multiplying the pair multiplier by itself would say 1.58×; measured, it's 1.44×. Each group size carries its own number, fitted on 2016-2021 and checked on 2022-2025, where it lands within 5% against 15% for the naive composition. Quads are listed but not priced — that multiplier didn't hold out of sample, and an unreliable price is worse than none.
They price the opposite way. The two legs are the same event — the receiving touchdown IS the passing touchdown — so where two teammates who both score come in 11% below the naive product, a quarterback and his receiver come in 33% above it. Same tool, opposite sign, and the reason is structural rather than statistical: teammates compete for a fixed pool of touchdowns, a passer and his target share one.
Search. Type names (comma-separated) to require specific players in every group.

Comma-separate names to require ALL of them in a group

Loading…