Accuracy

Season results, historical backtests and clear metric definitions. The published weekly record is on the Scoreboard.

Season:

2026 · half-PPR · completed weeks since the Week 4 repair

Weeks 1–3 are excluded from these accuracy figures. Those forecasts were affected by the range-floor defect fixed before Week 4. Accuracy, evaluated projections and average error below all use the same completed weeks after that repair. The original full-season results are retained separately below.

Loading the completed-week breakdown…

—

In-Range Accuracy · Since the Week 4 Repair

Actual score landed within projected floor–ceiling range

—

Evaluated Predictions

Projections with confirmed actual scores

—

Mean Absolute Error

Average miss distance in fantasy points

Current-season context, original archive and starter breakdown

2026 published record and context

Half-PPR coverage of published graded projections, with the original season record preserved separately.

The weekly breakdown could not be matched to this season's published total. Accuracy for the repair period is unavailable.

Recorded range repair. The main figure uses completed weeks from Week 4 onward; the completed-week cutoff is currently unavailable. The original full-season record retains all published results.

The forward-only verified record →Weekly autopsy availability could not be verified right now.

What "In Range" Means

Every projection produces three numbers: a floor, a median (our best single-point estimate), and a ceiling. These define the projected range. The floor and ceiling are coverage targets measured across all players — see "How Floors and Ceilings Are Set" below for exactly what they aim at.

A prediction is "in range" when the player's actual fantasy score lands anywhere between the floor and ceiling — inclusive. We do not cherry-pick a single point estimate and claim a hit when the actual is one point off. The range must contain the actual.

Why the width matters: a range wide enough to contain almost anything would score well and tell you nothing. Ours is deliberately not that: we target roughly 12% of results finishing below the floor and roughly 15% above the ceiling, so the band is meant to be beaten about a quarter of the time. An in-range rate far above that target would mean the bands had drifted too wide to be useful, which is a calibration failure we track rather than a result we would advertise.

How Floors and Ceilings Are Set

The weekly model estimates workload and production, calibrates point targets and constructs a constrained range. Its inputs can include:

Historical position-level scoring distributions (2019–2025)
Opponent defensive strength (adjusted points allowed by position)
Injury/practice status at projection time
Game context: home/away, Vegas total, weather for outdoor stadiums
CAF width adjustments; separate availability, allocation and environment paths can also affect the central estimate

The floor and ceiling are coverage targets, not percentiles of a per-player distribution. Across all players we aim for roughly 12% of actual results to finish below the floor and roughly 15% to finish above the ceiling. That is a statement about the population over a season, not a claim about any single row. Optional MC estimates may use a band-based approximation; they do not make these endpoints calibrated per-player quantiles, so we do not describe them as p10/p90. "All players" means every player we actually projected for that week — a player the model expects not to play carries no range, so there is nothing to be right or wrong about. Everyone else counts, including players whose floor is zero. The median is our central point estimate. Final stat scenarios can be reconciled to those point targets; arithmetic agreement is not evidence of independent stat accuracy. The weekly confidence score is a separate 0–100 composite, not the MC in-range chance or this page's observed coverage. Full methodology →

Walk-forward calibration adjusts these distributions each week based on recent prediction errors, so the model self-corrects over the course of a season.

Our Commitment

No retroactive edits to a published week. Once a week is published, the floor, median and ceiling you saw are the ones evaluated against actuals — we do not rescore a week against a later model.

Historical calibration is not a forward forecast. The 2019–2025 record is a walkforward backtest: the model never saw a game's result before projecting it, but some inputs — weather, injury and betting context — were reconstructed after those games were played, so they are not pre-kickoff captures. Verified forward-only forecasting, captured before kickoff, begins in 2026 and is tracked separately on the Scoreboard.

Evaluation periods are explicit. The 2026 primary record starts with completed Week 4, after the range-floor defect was repaired. The affected early weeks are excluded from that figure and retained in the separately labelled original archive. Historical backtests remain available through the season selector.

Source counts are disclosed. The "Evaluated Predictions" number in Season summary (—) counts published graded player-weeks in the selected half-PPR population — not total rows in our database (—).

← Forward-only verified recordAutopsy availability unverified

Accuracy

Published weekly results

2026 published record and context

What "In Range" Means

How Floors and Ceilings Are Set

Our Commitment