Receipts
Every projection we publish gets graded.
This is the ledger. Every week from Week 1, the projections we published get scored against what players actually did. The row posts, it settles once the NFL's late stat corrections land, and then it is final. Nothing gets backfilled, and no losing week quietly disappears. It is empty right now because the season has not kicked off. You are looking at the honest starting state, not a page that failed to load.
Quick answer
How accurate are these fantasy football projections?
Nothing is graded yet. Week 1 2026 is the first live test. In backtesting, the model beat the prior-season baseline in three of four held-out seasons, by 0.0024 to 0.0045 Spearman. It lost the fourth by 0.0063, wider than any of its wins. It cleared the gate; it did not clear it easily. The rows below are the real answer.
The 2026 ledger
| Week | Our MAE | Consensus MAE | Beat rate | Range hit |
|---|
No rows yet. The first grade posts once Week 1 is complete, and from there you get one new row every week.
MAE is mean absolute error: the average gap, in fantasy points, between a projection and what the player actually scored. Lower wins. Our MAE is computed on the players we projected that week who played. Bench zeros do not pad it, and a player we never ranked cannot flatter it. Beat rate is the share of that week's graded players where our projection landed closer to the actual than consensus did. 50% is a coin flip. It needs the consensus column, and that license is not signed. If it lands, both MAE columns narrow to the players we and the consensus source both project. Neither side gets credit for a player the other never ranked. Range hit is the share of graded players whose actual score landed inside the P10–P90 range we published for them: the test of whether our ranges are the right width, run in the open every week. A range that is too narrow will show it here, and we would rather you see that than not.
The model behind these grades is documented in the open, backtest and loss included: how our projections work
The projections already on the record, free and public: the 2026 draft board
FAQ
Why is the accuracy ledger empty?
Because the 2026 season hasn't been played. Live grading starts at Week 1. Until a week is in the books, there's nothing honest to put in the table. An empty ledger on launch day is the cost of refusing to backfill it.
Why not backfill the ledger with past seasons?
Because we'd be marking our own homework after seeing the answers. A model tuned on 2025 will always look good on 2025. The backtest is real work, published in full on the methodology page, loss included. It stays there, labeled. It never becomes a row in this table. Only weeks projected before kickoff get graded here.
When does the ledger update?
The first grade posts once the Sunday slate is final. If a game is still to come, that grade is partial, and it completes the day after the last one is played. Week 18 has no Monday night game, so its Sunday grade is already the full one. Thursday's pass folds in the stat corrections the NFL files late, and that is when the row is final. Until then it can still move. After that, it stands whether it flatters us or not.
What is the consensus column?
A licensed industry consensus projection, bought from a third party. We grade against their number so the comparison is not us against a strawman we drew ourselves. That license is not signed yet. If it is not in place by Week 1, the consensus column stays empty and the ledger still runs. Our error against actual results is the promise; the benchmark annotates it.
Do you grade the P10–P90 ranges too?
Yes — from Week 1 the Range hit column shows, every week, the share of graded players whose actual score landed inside the band we published. This is where we are weakest. Every projection ships as a P10–P50–P90 band, and that band is supposed to contain about 80% of real outcomes. In backtesting it held 66.7% to 83.2%. Two of those four seasons now reach the target; two, the worst included, still fall under it, and there the bands stay too tight. v2 narrowed this — it did not close it — and the whole backtest, loss and all, lives on the methodology page. Whatever the weekly column reads, it reads in public.