Two projection sources looking at the same player in the same week frequently publish numbers that are far apart. The disagreement comes from modelling choices rather than from one side having better information.
Projections are built from separate components
A projection is rarely a single estimate. It is usually a chain: how much playing time the player gets, how often he is involved when on the field, and how efficiently he converts that involvement.
Each component is estimated separately and then multiplied, which means small differences compound rather than average out.
A modest disagreement about snap share and another about conversion rate can produce a substantial gap in the final number without either model being obviously wrong.
Sample weighting is a genuine judgement call
Recent performance carries the most current information but the smallest sample, while a full season carries the largest sample and the most outdated context.
Every system chooses how far back to look and how steeply to discount older data, and there is no correct answer because the right weighting depends on what changed.
Models that lean on recent form react quickly to real change and also to noise; models that lean on longer histories are steadier and slower to notice a genuine shift.
Regression assumptions pull in opposite directions
Efficiency statistics fluctuate heavily in small samples, so most models pull extreme values back towards a baseline before projecting forward.
How hard to pull, and towards which baseline, differs between systems. One may regress towards a player's career norm, another towards a positional average.
The choice matters most for exactly the players managers are unsure about, which is why disagreement is largest where it is least helpful.
Opportunity forecasts are the biggest divergence
Predicting how a coach will distribute work is closer to reporting than to modelling, and models handle it with wildly different assumptions.
Some infer usage purely from historical patterns, while others incorporate depth-chart information and stated intentions that cannot be validated statistically.
Because usage is the largest single driver of fantasy output, this is where two credible sources most often part company.
How to use the disagreement
The spread between projections is itself information: a wide spread signals genuine uncertainty about a player rather than a flaw in one source.
Narrow agreement suggests the outcome is relatively predictable, which usually means the player is priced accordingly and offers little edge.
Treating projections as a distribution rather than a point estimate is the practical adjustment, because lineup decisions depend on range as much as on expected value.