1. Start with compatible sources.
FF combines available FantasyPros, ESPN, Sleeper, CBS and FFToday inputs. Availability varies by week, position and provider. Raw categories from additional sources are rescored where the selected scoring rules can be supported; incompatible or incomplete categories are excluded rather than guessed.
Source identity includes the player, team and position. Missing projections stay missing, and a genuine zero is distinct from a missing value. The remaining compatible source weights are normalized.
2. Weight by position, then adjust carefully.
v4 uses different provider weights for QB, RB, WR, TE, K and D/ST. These are research-informed starting weights, not a claim of fitted or proven accuracy at each position.
The existing bounded adjustments remain: recent form can influence large deviations by up to 20%; extreme positional defensive matchups adjust by up to 10%; qualified advanced metrics adjust by up to 5%. Inputs and sample limitations affect whether an adjustment is usable.
Compatible saved Boosted Trees and Workload estimates can each contribute up to 10% of the final public expected score, with bounded deviations from the provider-based estimate. Monte Carlo is not added as a duplicate mean vote on top of overlapping workload inputs. When compatible saved inputs are unavailable, the provider-based forecast remains usable.
3. Keep the mean, median and range distinct.
- FF expected points / mean: the ordering statistic used for every player in Fortified Rankings.
- Median: the midpoint of the modeled probability distribution, shown as additional context when available. It does not replace the mean for some players.
- Experimental ranges: middle 50% and middle 80% intervals from a compatible saved simulation. They are modeled intervals, not guaranteed floors or ceilings.
Normal-availability simulation residuals can be recentered and scale-adjusted to the v4 expected score. Availability-conditional mixtures are preserved and labeled separately. Missing or unordered distributions stay unavailable.
4. Apply the right football context.
A weekly points preference uses the current FF forecast. Saved experimental evidence does not create a second close-call weighting. Availability, kickoff locks and eligible lineup slots remain separate constraints.
Pickups, Drop/Add and GamePlan can consider several weeks, roster capacity, legal transactions and your league’s actual playoff calendar. A different objective or custom scoring can legitimately produce a different roster decision from a public single-week ranking. Old saved reports are snapshots of their original inputs; they are not silently rewritten as new-model forecasts.
5. Evaluate prospectively.
We preserve original forecast, capture and training timestamps and distinguish pregame forecasts from retrospective runs. Future actuals are not training inputs for their own evaluation. Forecast error, bias and interval coverage require an appropriate sample of original predictions followed by actual outcomes.
v4 coverage is collecting. Model agreement is not independent confirmation when the models share inputs. We do not publish an accuracy advantage or win probability merely because the model version increased.
Release notes.
- v4: expanded CBS/FFToday coverage where compatible, position-specific provider blending, bounded saved-model influence, medians and experimental distribution windows.
- October 31, 2026: aligned weekly Start/Sit and screenshot pickup recommendations with the current FF expected-points order; removed the extra close-call blend and hidden pickup priority adjustments. Updated shared preset-board enrichment and explanation consistency.
- Historical forecasts: remain associated with the model and inputs that produced them. Their scores and grades are not relabeled as v4.
Public previews use common scoring presets. Confirm your league’s scoring, lineup and playoff settings before making a move.