Down and distance are genuine, documented inputs to a win probability model — nflfastR’s published feature list includes both down and yards to go alongside score, time, field position and timeouts. And yet a viewer watching a live win-probability chart during most of a game could be forgiven for concluding they barely matter at all, because their effect on the displayed number is frequently too small to see. That gap between “included in the model” and “visible in the output” is a real usability question, distinct from any question about whether the model itself is correctly built.
What Down and Distance Actually Contribute
Down and yards to go describe how likely the current possession is to continue — first-and-10 and third-and-15 imply very different odds of the drive producing points, even holding score, clock and field position fixed. That is a real, defensible reason to include both as separate model inputs, covered in Part 1 of the building series alongside the other four core features. Nothing about how down and distance get used inside a fitted model is in question here; the question is how much they move the final number relative to the game’s other inputs.
The two features are also asymmetric in how directly they translate to outcomes. Down maps fairly cleanly onto urgency — fourth down carries obvious finality that first down doesn’t — while distance interacts more continuously with field position and the specific offense’s tendencies, which is part of why a tree-based model, covered in Part 2 of the building series, can represent the relationship between the two more flexibly than a model that treats “yards to go” as a single linear effect regardless of down.
A related, easily confused question is whether down and distance are worth including at all, given how small their typical marginal contribution looks. That is a different question from usability, and the answer to it is straightforward: a feature earns its place in a model by improving calibration and reducing log loss when tested against held-out data, not by producing a visually dramatic swing on a chart. A feature that quietly shifts the estimate by half a percentage point in most situations but by several points in the specific high-leverage situations where it matters most is exactly the kind of input worth keeping — removing it would degrade the model precisely where accuracy matters most, even though its absence would be hardest to notice everywhere else.
Why the Effect Is Often Invisible
Score differential and time remaining dominate a win probability estimate in most game states, particularly in blowouts or in the early stages of a close game, because together they bound how many scoring possessions realistically remain. Against that backdrop, changing the down from first to third, or the distance from 2 yards to 12, typically shifts the estimate by a small fraction of a percentage point — mathematically real, but visually indistinguishable on a chart whose scale runs from 0% to 100% across an entire game. A viewer tracking the win-probability line through three quarters of a comfortable lead has no practical way to notice that down and distance are doing anything at all, because in that context they aren’t doing much.
The effect is not uniform, though. It grows sharply in exactly the situations where the game’s outcome is still genuinely contested and the clock is short: a third-and-2 inside the two-minute warning of a one-score game carries a meaningfully different win probability than a third-and-15 in the identical spot, because the gap in conversion odds directly changes how likely a team is to keep possession, run clock, or extend a scoring chance with limited time left. The same feature that looked inert in the third quarter becomes one of the more consequential inputs in the closing minutes — an interaction with time remaining, not a fixed importance level.
A Usability Fix That Doesn’t Touch the Model
The usability gap described above has a fix that never requires changing the underlying win probability model at all: presenting a supplementary number alongside the main chart, specific to the current down-and-distance situation, rather than expecting a single running probability to visibly carry every input at once. A “conversion probability” or “drive-success estimate” shown next to a win-probability chart on third-and-2 versus third-and-15 would make the exact contribution of down and distance legible in the moment it matters most, without asking the headline win-probability number to do double duty as both a game-outcome estimate and a play-level explainer. That kind of layered presentation — one number for the game, a second number for the specific decision in front of it — is closer to what the NYT bot already does by pairing an expected-points view with a win-probability view, just generalized beyond fourth downs specifically.
Where the Usability Question Actually Gets Solved
The New York Times 4th Down Bot sidesteps the visibility problem by making down and distance the explicit subject of its output rather than one buried input among several. A fourth-down decision only exists because of the down and distance directly in front of it, so the bot’s recommendation — go for it, punt, or kick — is built around exactly the feature that a running win-probability chart tends to understate. It does this using an expected-points model earlier in a game, switching to a win-probability model in the final minutes of the fourth quarter, on the premise that once winning (not scoring) becomes the priority, the right measure of a decision changes too (Advanced Football Analytics, cited below). David Romer’s earlier academic work made a related point in a different framing: fourth-down decisions, evaluated through the lens of expected points built from down-and-distance-driven historical outcomes, showed teams departing systematically and consistently from what the numbers implied was the better choice (Romer, 2002).
The pattern across both examples is the same: down and distance become usable, visible drivers of a football analysis the moment the tool is built around the specific decision they govern, rather than folded into a single running probability where their marginal contribution gets crowded out by score and time. That is a display and framing choice, not evidence the feature was ever missing from the underlying model — down and distance were never absent from a win probability calculation; they were simply hard to see inside one.
The broader lesson generalizes past down and distance specifically. Any feature engineering choice covered earlier in the building series — timeouts remaining, field position, the spread-adjusted pregame prior — can end up in the same position: correctly included, correctly weighted, and still functionally invisible to a reader if the only way it gets presented is as one contributor buried inside a single running number. Usability and correctness are separate design problems, and solving the first one is a presentation exercise, not a reason to revisit the feature list.
Frequently asked questions
Do down and distance actually affect win probability?
Yes — down and yards to go are both standard inputs to a football win probability model, including nflfastR's published implementation, because they describe how likely the current drive is to continue. Their inclusion is not in question; what varies is how visible their effect is on the final displayed number next to bigger inputs like score and time.
Why might down and distance seem like they 'don't matter' when watching a win probability chart?
Because score differential and time remaining dominate the estimate in most game states, especially in blowouts or early in a game, and the marginal effect of a down-and-distance change often moves the number by a fraction of a percentage point in those situations — small enough that a viewer watching the chart may not notice it changed at all, even though the model used it.
When does down and distance visibly move a win probability estimate?
Most visibly in close, late-game situations — a two-minute-drill third-and-2 carries a meaningfully different win probability than a third-and-15 in the same spot, because the difference in conversion odds directly changes how likely the trailing or leading team is to control the clock or extend a scoring chance. The same input matters far more when the score and clock have already narrowed the range of likely outcomes.
How does the NYT 4th Down Bot use down and distance differently from a win probability chart?
It uses down and distance as the direct subject of its recommendation — a fourth-down decision only exists because of the down and distance in front of it — evaluated through an expected-points model earlier in the game and a win-probability model in the final minutes, rather than as one input among several inside a single running number.
Is a usability gap the same thing as a modeling flaw?
No. A model can correctly weight down and distance in its internal calculation while still presenting a chart or number where that weighting isn't visually obvious to a reader — that's a display and communication problem, not evidence the underlying math is wrong. The fix is in how the number gets explained or broken down, not in removing the feature from the model.