How Accurate Are These Rankings?

Every completed week is graded player by player. Each player is scored against the last board published before his own game kicked off, and his actual fantasy points are computed under that format's real scoring rules. Below the live results is a backtest across two seasons the model was not built on.

Every figure in this section is a backtest, not a live result. For each week of 2024 and 2025 the rankings were rebuilt from scratch using only what was known beforehand, then compared with the week's actual results. The comparisons below are against what a reasonable person would do instead — not against nothing.

Of the players the model listed as startable each week, this is the share who actually finished startable — top twelve at quarterback and tight end, top twenty-four at running back, top thirty-six at receiver. Alongside it, the same figure for ranking players by their season average instead.

Tight end shows the largest gap, which makes sense — it is the shallowest position and the one where a change in role moves a player furthest. Running back shows the smallest, because the position is easier to predict by any method.

Ranking players by their season average is a reasonable approach and hard to beat on the players everyone already agrees about. Where a model can add something is on the players whose situation just changed.

So the sharper test is what happens when the two methods disagree. When the model and season-average ranking differ by ten or more places on a player, the model lands closer to what actually happened 61% of the time.

That held in all twenty-four combinations of position, scoring format and season tested, ranging from 56% to 68%.

Ordering is one question; the numbers themselves are another. A board can be in the right order and still project everyone too high.

Measured as the average gap between projected and actual points, the projections come out close to correct at every position — under a point at quarterback and under half a point everywhere else.

The typical error looks large next to a normal weekly score, and it is — because weekly fantasy football is genuinely noisy. A receiver projected for twelve points might score four or twenty-two, and no model changes that. What the first column shows is that the projections are not systematically high or low.

Defense is where the model's edge over the obvious alternative is largest.

Ranking defenses by their own recent scoring tracks actual results at a correlation of 0.11. Ranking them the way this model does — driven mainly by how many points the opposing offense is expected to score — reaches 0.22, roughly double.

That is the case for treating defense as a matchup decision rather than a judgment about which unit is better.

Kicker rankings track actual results at a correlation of 0.15. That is weak, and it is worth saying why rather than hiding it: ranking kickers by their season average reaches 0.14, and by last week's points 0.09.

Weekly kicker scoring is close to unpredictable by any method available. The model is marginally better than the alternatives, and nobody should expect much more than that from any kicker ranking anywhere.

Correlation measures how closely the order of a board matches the order the week actually finished in. A correlation of 1.0 would mean the board matched the results exactly; 0 would mean no relationship at all. Weekly fantasy football is noisy enough that nothing approaches 1.0.

These are correlation coefficients on a −1 to 1 scale, not percentages. The model beats both alternatives at every position and in all twenty-four cells tested. The margin over last week's points is large. The margin over season average is small, which is why the disagreement figure above is the more useful measure.

Quarterback is the weakest position, at 0.47 against 0.71 for running back. The model measures opportunity — dropbacks and designed carries — and does not measure how efficiently a quarterback converts them.

The model measures volume — snaps, carries, targets, air yards, red-zone work — and does not measure how efficiently a player converts it. Several candidate additions have been tested against the same two seasons: red-zone usage, an efficiency adjustment, and blending with a season-average baseline. Each gained under a hundredth of a point of correlation, which suggests the remaining headroom is not in these inputs.

Red-zone usage is the clearest illustration. It predicts touchdowns, and touchdowns are the least predictable part of a weekly score — a running back with three carries inside the ten might score twice or not at all. The signal is real and it is included in the model, but it is attached to the noisiest outcome, and most of what is stable about a player's role already shows up in his snaps and targets.